Authors: Gyubok Lee, Woosog Chay, Edward Choi
Abstract: Recent advances in Large Language Models (LLMs) have enabled the development of text-to-SQL models that allow clinicians to query structured data stored in Electronic Health Records (EHRs) using natural language. However, deploying these models for EHR question answering (QA) systems in safety-critical clinical environments remains challenging: incorrect SQL queries-whether caused by model errors or problematic user inputs-can undermine clinical decision-making and jeopardize patient care. While prior work has mainly focused on improving SQL generation accuracy or filtering questions before execution, there is a lack of a unified benchmark for evaluating independent post-hoc verification mechanisms (i.e., a component that inspects and validates the generated SQL before execution), which is crucial for safe deployment. To fill this gap, we introduce SCARE, a benchmark for evaluating methods that function as a post-hoc safety layer in EHR QA systems. SCARE evaluates the joint task of (1) classifying question answerability (i.e., determining whether a question is answerable, ambiguous, or unanswerable) and (2) verifying or correcting candidate SQL queries. The benchmark comprises 4,200 triples of questions, candidate SQL queries, and expected model outputs, grounded in the MIMIC-III, MIMIC-IV, and eICU databases. It covers a diverse set of questions and corresponding candidate SQL queries generated by seven different text-to-SQL models, ensuring a realistic and challenging evaluation. Using SCARE, we benchmark a range of approaches-from two-stage methods to agentic frameworks. Our experiments reveal a critical trade-off between question classification and SQL error correction, highlighting key challenges and outlining directions for future research.
Authors: Yuechi Zhou, Yi Su, Jianxin Zhang, Juntao Li, Qingrong Xia, Zhefeng Wang, Xinyu Duan, Baoxing Huai
Abstract: Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the $\textbf{A}$ttention-$\textbf{A}$ware $\textbf{A}$ccurate KV Cache Fusion algorithm ($A^3$), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that $A^3$ achieves the best task performance compared to four baselines while reducing the time-to-first-token (TTFT) by 2$\times$.
Authors: Huimin Ren, Yan Liang, Baiqiao Su, Chaobo Sun, Hengtong Lu, Kaike Zhang, Chen Wei
Abstract: The ability of Large Language Models (LLMs) to precisely follow complex and fine-grained lexical instructions is a cornerstone of their utility and controllability. However, evaluating this capability remains a significant challenge. Current methods either rely on subjective and costly human evaluation or on automated LLM-as-a-judge systems, which suffer from inherent biases and unreliability. Existing programmatic benchmarks, while objective, often lack the expressiveness to test intricate, compositional constraints at a granular level. To address these limitations, we introduce LexInstructEval, a new benchmark and evaluation framework for fine-grained lexical instruction following. Our framework is built upon a formal, rule-based grammar that deconstructs complex instructions into a canonical
Authors: Wei Tian, YuhaoZhou
Abstract: This paper introduces ChineseErrorCorrector3-4B, a unified model for Chinese spelling and grammatical error correction based on Qwen3-4B. The model demonstrates outstanding performance in general text correction tasks and achieves state-of-the-art results in both spelling correction (CSC) and grammatical correction (CGC). On several authoritative benchmark datasets -- including SIGHAN-2015, EC-LAW, MCSC, and NaCGEC -- the model's F1 and F0.5 scores significantly surpass existing publicly available models, ranking first in both spelling and grammatical error correction tasks.
Authors: Sarthak Chakraborty, Suman Nath, Xuchao Zhang, Chetan Bansal, Indranil Gupta
Abstract: Large Language Models (LLMs) are increasingly being used to plan, reason, and execute tasks across diverse scenarios. In use cases like repeatable workflows and agentic settings, prompts are often reused with minor variations while having a similar structure for recurring tasks. This opens up opportunities for caching. However, exact prompt matching fails on such structurally similar prompts, while semantic caching may produce incorrect responses by ignoring critical differences. To address this, we introduce \ourmethod{}, a generative cache that produces variation-aware responses for structurally similar prompts. \ourmethod{} identifies reusable response patterns across similar prompt structures and synthesizes customized outputs for new requests. We show that \ourmethod{} achieves 83\% cache hit rate, while having minimal incorrect hits on datasets without prompt repetition. In agentic workflows, it improves cache hit rate by $\sim$20\% and reduces end-to-end execution latency by $\sim$34\% compared to standard prompt matching.
Authors: Patrick Gerard, Aiden Chang, Svitlana Volkova
Abstract: When large language models (LLMs) are aligned to a specific online community, do they exhibit generalizable behavioral patterns that mirror that community's attitudes and responses to new uncertainty, or are they simply recalling patterns from training data? We introduce a framework to test epistemic stance transfer: targeted deletion of event knowledge, validated with multiple probes, followed by evaluation of whether models still reproduce the community's organic response patterns under ignorance. Using Russian--Ukrainian military discourse and U.S. partisan Twitter data, we find that even after aggressive fact removal, aligned LLMs maintain stable, community-specific behavioral patterns for handling uncertainty. These results provide evidence that alignment encodes structured, generalizable behaviors beyond surface mimicry. Our framework offers a systematic way to detect behavioral biases that persist under ignorance, advancing efforts toward safer and more transparent LLM deployments.
Authors: Vladimir Berman
Abstract: We study a deliberately simple, fully non-linguistic model of text: a sequence of independent draws from a finite alphabet of letters plus a single space symbol. A word is defined as a maximal block of non-space symbols. Within this symbol-level framework, which assumes no morphology, syntax, or semantics, we derive several structural results. First, word lengths follow a geometric distribution governed solely by the probability of the space symbol. Second, the expected number of words of a given length, and the expected number of distinct words of that length, admit closed-form expressions based on a coupon-collector argument. This yields a critical word length k* at which word types transition from appearing many times on average to appearing at most once. Third, combining the exponential growth of the number of possible strings of length k with the exponential decay of the probability of each string, we obtain a Zipf-type rank-frequency law p(r) proportional to r^{-alpha}, with an exponent determined explicitly by the alphabet size and the space probability. Our contribution is twofold. Mathematically, we give a unified derivation linking word lengths, vocabulary growth, critical length, and rank-frequency structure in a single explicit model. Conceptually, we argue that this provides a structurally grounded null model for both natural-language word statistics and token statistics in large language models. The results show that Zipf-like patterns can arise purely from combinatorics and segmentation, without optimization principles or linguistic organization, and help clarify which phenomena require deeper explanation beyond random-text structure.
Authors: Sharaj Kunjar, Alyssa Hasegawa Smith, Tyler R Mckenzie, Rushali Mohbe, Samuel V Scarpino, Brooke Foucault Welles
Abstract: Computational approaches have previously shown various promises and pitfalls when it comes to the reliable identification of media frames. Generative LLMs like GPT and Claude are increasingly being used as content analytical tools, but how effective are they for frame analysis? We address this question by systematically evaluating them against their computational predecessors: bag-of-words models and encoder-only transformers; and traditional manual coding procedures. Our analysis rests on a novel gold standard dataset that we inductively and iteratively developed through the study, investigating six months of news coverage of the US Mpox epidemic of 2022. While we discover some potential applications for generative LLMs, we demonstrate that they were consistently outperformed by manual coders, and in some instances, by smaller language models. Some form of human validation was always necessary to determine appropriate model choice. Additionally, by examining how the suitability of various approaches depended on the nature of different tasks that were part of our frame analytical workflow, we provide insights as to how researchers may leverage the complementarity of these approaches to use them in tandem. We conclude by endorsing a methodologically pluralistic approach and put forth a roadmap for computational frame analysis for researchers going forward.
Authors: Thales Sales Almeida, Rodrigo Nogueira, H\'elio Pedrini
Abstract: Large Language Models (LLMs) exhibit significant variations in performance across linguistic and cultural contexts, underscoring the need for systematic evaluation in diverse languages. In this work, we present the most extensive evaluation of LLMs for the Portuguese language to date. Leveraging our newly introduced PoETa v2 benchmark -- a comprehensive suite of over 40 tasks in Portuguese -- we assess more than 20 models covering a broad spectrum of training scales and computational resources. Our study reveals how computational investment and language-specific adaptation impact performance in Portuguese, while also analyzing performance gaps in comparison to equivalent tasks in English. Through this benchmark and analysis, PoETa v2 lays the groundwork for future research on Portuguese language modeling and evaluation. The benchmark is available at https://github.com/PoETaV2/PoETaV2.
Authors: Scott Merrill, Shashank Srivastava
Abstract: Large language models offer opportunities to simulate multi-party deliberation, but realistic modeling remains limited by a lack of speaker-attributed data. Transcripts produced via automatic speech recognition (ASR) assign anonymous speaker labels (e.g., Speaker_1), preventing models from capturing consistent human behavior. This work introduces a reproducible pipeline to transform public Zoom recordings into speaker-attributed transcripts with metadata like persona profiles and pragmatic action tags (e.g., [propose_motion]). We release three local government deliberation datasets: Appellate Court hearings, School Board meetings, and Municipal Council sessions. Fine-tuning LLMs to model specific participants using this "action-aware" data produces a 67% reduction in perplexity and nearly doubles classifier-based performance metrics for speaker fidelity and realism. Turing-style human evaluations show our simulations are often indistinguishable from real deliberations, providing a practical and scalable method for complex realistic civic simulations.
Authors: Allen Roush, Devin Gonier, John Hines, Judah Goldfeder, Philippe Martin Wyder, Sanjay Basu, Ravid Shwartz Ziv
Abstract: The capacity for highly complex, evidence-based, and strategically adaptive persuasion remains a formidable great challenge for artificial intelligence. Previous work, like IBM Project Debater, focused on generating persuasive speeches in simplified and shortened debate formats intended for relatively lay audiences. We introduce DeepDebater, a novel autonomous system capable of participating in and winning a full, unmodified, two-team competitive policy debate. Our system employs a hierarchical architecture of specialized multi-agent workflows, where teams of LLM-powered agents collaborate and critique one another to perform discrete argumentative tasks. Each workflow utilizes iterative retrieval, synthesis, and self-correction using a massive corpus of policy debate evidence (OpenDebateEvidence) and produces complete speech transcripts, cross-examinations, and rebuttals. We introduce a live, interactive end-to-end presentation pipeline that renders debates with AI speech and animation: transcripts are surface-realized and synthesized to audio with OpenAI TTS, and then displayed as talking-head portrait videos with EchoMimic V1. Beyond fully autonomous matches (AI vs AI), DeepDebater supports hybrid human-AI operation: human debaters can intervene at any stage, and humans can optionally serve as opponents against AI in any speech, allowing AI-human and AI-AI rounds. In preliminary evaluations against human-authored cases, DeepDebater produces qualitatively superior argumentative components and consistently wins simulated rounds as adjudicated by an independent autonomous judge. Expert human debate coaches also prefer the arguments, evidence, and cases constructed by DeepDebater. We open source all code, generated speech transcripts, audio and talking head video here: https://github.com/Hellisotherpeople/DeepDebater/tree/main
URLs: https://github.com/Hellisotherpeople/DeepDebater/tree/main
Authors: Debashish Chakraborty, Eugene Yang, Daniel Khashabi, Dawn Lawrie, Kevin Duh
Abstract: Retrieval-Augmented Generation (RAG) enhances factual grounding in large language models (LLMs) by incorporating retrieved evidence, but LLM accuracy declines when long or noisy contexts exceed the model's effective attention span. Existing pre-generation filters rely on heuristics or uncalibrated LLM confidence scores, offering no statistical control over retained evidence. We evaluate and demonstrate context engineering through conformal prediction, a coverage-controlled filtering framework that removes irrelevant content while preserving recall of supporting evidence. Using both embedding- and LLM-based scoring functions, we test this approach on the NeuCLIR and RAGTIME collections. Conformal filtering consistently meets its target coverage, ensuring that a specified fraction of relevant snippets are retained, and reduces retained context by 2-3x relative to unfiltered retrieval. On NeuCLIR, downstream factual accuracy measured by ARGUE F1 improves under strict filtering and remains stable at moderate coverage, indicating that most discarded material is redundant or irrelevant. These results demonstrate that conformal prediction enables reliable, coverage-controlled context reduction in RAG, offering a model-agnostic and principled approach to context engineering.
Authors: Yuliang Zhan, Xinyu Tang, Han Wan, Jian Li, Ji-Rong Wen, Hao Sun
Abstract: Recently, Chain-of-Thought (CoT) reasoning has significantly enhanced the capabilities of large language models (LLMs), but Vision-Language Models (VLMs) still struggle with multi-step reasoning tasks due to limited multimodal reasoning data. To bridge this gap, researchers have explored methods to transfer CoT reasoning from LLMs to VLMs. However, existing approaches either need high training costs or require architectural alignment. In this paper, we use Linear Artificial Tomography (LAT) to empirically show that LLMs and VLMs share similar low-frequency latent representations of CoT reasoning despite architectural differences. Based on this insight, we propose L2V-CoT, a novel training-free latent intervention approach that transfers CoT reasoning from LLMs to VLMs. L2V-CoT extracts and resamples low-frequency CoT representations from LLMs in the frequency domain, enabling dimension matching and latent injection into VLMs during inference to enhance reasoning capabilities. Extensive experiments demonstrate that our approach consistently outperforms training-free baselines and even surpasses supervised methods.
Authors: Wenda Li, Tongya Zheng, Shunyu Liu, Yu Wang, Kaixuan Chen, Hanyang Yuan, Bingde Hu, Zujie Ren, Mingli Song, Gang Chen
Abstract: Heterogeneous graphs are widely present in real-world complex networks, where the diversity of node and relation types leads to complex and rich semantics. Efforts for modeling complex relation semantics in heterogeneous graphs are restricted by the limitations of predefined semantic dependencies and the scarcity of supervised signals. The advanced pre-training and fine-tuning paradigm leverages graph structure to provide rich self-supervised signals, but introduces semantic gaps between tasks. Large Language Models (LLMs) offer significant potential to address the semantic issues of relations and tasks in heterogeneous graphs through their strong reasoning capabilities in textual modality, but their incorporation into heterogeneous graphs is largely limited by computational complexity. Therefore, in this paper, we propose an Efficient LLM-Aware (ELLA) framework for heterogeneous graphs, addressing the above issues. To capture complex relation semantics, we propose an LLM-aware Relation Tokenizer that leverages LLM to encode multi-hop, multi-type relations. To reduce computational complexity, we further employ a Hop-level Relation Graph Transformer, which help reduces the complexity of LLM-aware relation reasoning from exponential to linear. To bridge semantic gaps between pre-training and fine-tuning tasks, we introduce the fine-grained task-aware textual Chain-of-Thought (CoT) prompts. Extensive experiments on four heterogeneous graphs show that our proposed ELLA outperforms state-of-the-art methods in the performance and efficiency. In particular, ELLA scales up to 13b-parameter LLMs and achieves up to a 4x speedup compared with existing LLM-based methods. Our code is publicly available at https://github.com/l-wd/ELLA.
Authors: Jianghao Wu, Yasmeen George, Jin Ye, Yicheng Wu, Daniel F. Schmidt, Jianfei Cai
Abstract: Large language models (LLMs) and multimodal LLMs (MLLMs) excel at chain-of-thought reasoning but face distribution shift at test-time and a lack of verifiable supervision. Recent test-time reinforcement learning (TTRL) methods derive label-free pseudo-rewards from self-consistency voting over sampled trajectories, yet they often collapse: the majority-vote reward prevails, responses shorten, and Pass@1 declines. We trace this to uniform sequence updates in which most tokens are low-entropy followers, while a small high-entropy subset determines the reasoning branches. Thus we propose SPINE, a token-selective test-time reinforcement learning framework that (i) updates only forking tokens, the high-entropy branch points identified from forward-pass statistics, and (ii) applies an entropy-band regularizer at those tokens to sustain exploration when entropy is too low and to suppress noisy supervision when it is too high. SPINE plugs into GRPO-style objectives, optionally with a KL anchor, and requires no labels or reward models. Across ten benchmarks spanning multimodal VQA, general and expert QA, mathematical reasoning, and medical QA, SPINE consistently improves Pass@1 over TTRL while avoiding response-length collapse and yielding more stable training dynamics on both LLM and MLLM backbones. These results indicate that aligning updates with chain-of-thought branch points is a simple and label-free mechanism for stable and effective test-time adaptation in reasoning models. Code is available at https://github.com/JianghaoWu/SPINE.
Authors: Shuo Zhang, Fabrizio Gotti, Fengran Mo, Jian-Yun Nie
Abstract: Hallucination in large language models (LLMs) is a fundamental challenge, particularly in open-domain question answering. Prior work attempts to detect hallucination with model-internal signals such as token-level entropy or generation consistency, while the connection between pretraining data exposure and hallucination is underexplored. Existing studies show that LLMs underperform on long-tail knowledge, i.e., the accuracy of the generated answer drops for the ground-truth entities that are rare in pretraining. However, examining whether data coverage itself can serve as a detection signal is overlooked. We propose a complementary question: Does lexical training-data coverage of the question and/or generated answer provide additional signal for hallucination detection? To investigate this, we construct scalable suffix arrays over RedPajama's 1.3-trillion-token pretraining corpus to retrieve $n$-gram statistics for both prompts and model generations. We evaluate their effectiveness for hallucination detection across three QA benchmarks. Our observations show that while occurrence-based features are weak predictors when used alone, they yield modest gains when combined with log-probabilities, particularly on datasets with higher intrinsic model uncertainty. These findings suggest that lexical coverage features provide a complementary signal for hallucination detection. All code and suffix-array infrastructure are provided at https://github.com/WWWonderer/ostd.
Authors: Dat Thanh Nguyen, Nguyen Hung Lam, Anh Hoang-Thi Nguyen, Trong-Hop Do
Abstract: With the rapid rise of short-form videos, TikTok has become one of the most influential platforms among children and teenagers, but also a source of harmful content that can affect their perception and behavior. Such content, often subtle or deceptive, challenges traditional moderation methods due to the massive volume and real-time nature of uploads. This paper presents MTikGuard, a real-time multimodal harmful content detection system for TikTok, with three key contributions: (1) an extended TikHarm dataset expanded to 4,723 labeled videos by adding diverse real-world samples, (2) a multimodal classification framework integrating visual, audio, and textual features to achieve state-of-the-art performance with 89.37% accuracy and 89.45% F1-score, and (3) a scalable streaming architecture built on Apache Kafka and Apache Spark for real-time deployment. The results demonstrate the effectiveness of combining dataset expansion, advanced multimodal fusion, and robust deployment for practical large-scale social media content moderation. The dataset is available at https://github.com/ntdat-8324/MTikGuard-System.git.
Authors: Gowtham, Sai Rupesh, Sanjay Kumar, Saravanan, Venkata Chaithanya
Abstract: High-quality training data is fundamental to large language model (LLM) performance, yet existing preprocessing pipelines often struggle to effectively remove noise and unstructured content from web-scale corpora. This paper presents Blu-WERP, a novel data preprocessing pipeline designed to optimize the quality of Common Crawl WARC files for LLM training. We demonstrate that Blu-WERP significantly outperforms established baselines including DCLM across multiple model scales and evaluation benchmarks. Our pipeline processes CC WARC dumps, implementing advanced filtering and quality assessment mechanisms. We conducted comprehensive evaluations using models with 150M, 400M, 530M, 750M, and 1B parameters, testing against nine standard benchmarks categorized as World Knowledge & Reasoning, Language Understanding, and Commonsense Reasoning. Results show Blu-WERP consistently achieved superior performance across all model scales. At the 1B parameter scale, Relatively Blu-WERP demonstrates a 4.0% and 9.5% aggregate improvement over DCLM and Fineweb respectively, while achieving quality-per-token efficiency gain. Categorical analysis reveals 2.4% improvement in World Knowledge & Reasoning, 6.2% improvement in Language Understanding, and 4.2% improvement in Commonsense Reasoning. These results establish Blu-WERP as a state-of-the-art preprocessing pipeline that substantially improves LLM training data quality and downstream model performance with reduced computational cost. Our findings contribute to the growing body of research on data-centric AI, demonstrating that preprocessing pipeline design significantly impacts LLM capabilities. The Blu-WERP pipeline represents a practical advancement in data quality optimization, offering researchers and practitioners an effective solution for improving LLM training efficiency and model performance.
Authors: Yomal De Mel, Nisansa de Silva
Abstract: This study introduce GeeSanBhava, a high-quality data set of Sinhala song comments extracted from YouTube manually tagged using Russells Valence-Arousal model by three independent human annotators. The human annotators achieve a substantial inter-annotator agreement (Fleiss kappa = 84.96%). The analysis revealed distinct emotional profiles for different songs, highlighting the importance of comment based emotion mapping. The study also addressed the challenges of comparing comment-based and song-based emotions, mitigating biases inherent in user-generated content. A number of Machine learning and deep learning models were pre-trained on a related large data set of Sinhala News comments in order to report the zero-shot result of our Sinhala YouTube comment data set. An optimized Multi-Layer Perceptron model, after extensive hyperparameter tuning, achieved a ROC-AUC score of 0.887. The model is a three-layer MLP with a configuration of 256, 128, and 64 neurons. This research contributes a valuable annotated dataset and provides insights for future work in Sinhala Natural Language Processing and music emotion recognition.
Authors: Sheridan Feucht, Byron Wallace, David Bau
Abstract: In order to predict the next token, LLMs must represent semantic and surface-level information about the current word. Previous work identified two types of attention heads that disentangle this information: (i) Concept induction heads, which copy word meanings, and (ii) Token induction heads, which copy literal token representations (Feucht et al., 2025). We show that these heads can be used to identify subspaces of model activations that exhibit coherent semantic structure in Llama-2-7b. Specifically, when we transform hidden states using the attention weights of concept heads, we are able to more accurately perform parallelogram arithmetic (Mikolov et al., 2013) on the resulting hidden states, e.g., showing that "Athens" - "Greece" + "China" = "Beijing". This transformation allows for much higher nearest-neighbor accuracy (80%) than direct use of raw hidden states (47%). Analogously, we show that token heads allow for transformations that reveal surface-level word information in hidden states, allowing for operations like "coding" - "code" + "dance" = "dancing".
Authors: Elias Lumer, Matt Melich, Olivia Zino, Elena Kim, Sara Dieter, Pradeep Honaganahalli Basavaraju, Vamse Kumar Subbiah, James A. Burke, Roberto Hernandez
Abstract: Recent advancements in Retrieval-Augmented Generation (RAG) have enabled Large Language Models to answer financial questions using external knowledge bases of U.S. SEC filings, earnings reports, and regulatory documents. However, existing work lacks systematic comparison of vector-based and non-vector RAG architectures for financial documents, and the empirical impact of advanced RAG techniques on retrieval accuracy, answer quality, latency, and cost remain unclear. We present the first systematic evaluation comparing vector-based agentic RAG using hybrid search and metadata filtering against hierarchical node-based systems that traverse document structure without embeddings. We evaluate two enhancement techniques applied to the vector-based architecture, i) cross-encoder reranking for retrieval precision, and ii) small-to-big chunk retrieval for context completeness. Across 1,200 SEC 10-K, 10-Q, and 8-K filings on a 150-question benchmark, we measure retrieval metrics (MRR, Recall@5), answer quality through LLM-as-a-judge pairwise comparisons, latency, and preprocessing costs. Vector-based agentic RAG achieves a 68% win rate over hierarchical node-based systems with comparable latency (5.2 compared to 5.98 seconds). Cross-encoder reranking achieves a 59% absolute improvement at optimal parameters (10, 5) for MRR@5. Small-to-big retrieval achieves a 65% win rate over baseline chunking with only 0.2 seconds additional latency. Our findings reveal that applying advanced RAG techniques to financial Q&A systems improves retrieval accuracy, answer quality, and has cost-performance tradeoffs to be considered in production.
Authors: Faheem Nizar, Elias Lumer, Anmol Gulati, Pradeep Honaganahalli Basavaraju, Vamse Kumar Subbiah
Abstract: Recent advances in Large Language Model Multi-Agent Systems enable scalable orchestration and retrieval of specialized, parallelized subagents, each equipped with hundreds or thousands of Model Context Protocol (MCP) servers and tools. However, existing agent, MCP, and retrieval methods typically match queries against a single agent description, obscuring fine-grained tool capabilities of each agent, resulting in suboptimal agent selection. We introduce Agent-as-a-Graph retrieval, a knowledge graph retrieval augmented generation approach that represents both tools and their parent agents as nodes and edges in a knowledge graph. During retrieval, i) relevant agents and tool nodes are first retrieved through vector search, ii) we apply a type-specific weighted reciprocal rank fusion (wRRF) for reranking tools and agents, and iii) parent agents are traversed in the knowledge graph for the final set of agents. We evaluate Agent-as-a-Graph on the LiveMCPBenchmark, achieving 14.9% and 14.6% improvements in Recall@5 and nDCG@5 over prior state-of-the-art retrievers, and 2.4% improvements in wRRF optimizations.
Authors: Xiaochen Zheng, Alvaro Serra, Ilya Schneider Chernov, Maddalena Marchesi, Eunice Musvasva, Tatyana Y. Doktorova
Abstract: Pharmaceutical research and development has accumulated vast, heterogeneous archives of data. Much of this knowledge stems from discontinued programs, and reusing these archives is invaluable for reverse translation. However, in practice, such reuse is often infeasible. In this work, we introduce DiscoVerse, a multi-agent co-scientist designed to support pharmaceutical research and development. The system implements semantic retrieval, cross-document linking, and auditable synthesis on a large historical corpus from Roche. To validate our approach at real-world scale, we selected a subset of 180 molecules from the Roche research repositories, covering over 0.87 billion BPE tokens and more than four decades of research. Given that automated evaluation metrics are poorly aligned with scientific utility, we evaluate the performance of DiscoVerse using blinded expert evaluation of source-linked outputs. To our knowledge, this is the first agentic framework systematically assessed on real pharmaceutical data for reverse translation, enabled by authorized access to confidential, end-to-end drug-development archives. Our contributions include role-specialized agent designs aligned with scientist workflows; human-in-the-loop support for reverse translation; expert evaluation; and a large-scale demonstration showing promising answer accuracy and decision-making insights. In brief, across seven benchmark queries covering 180 molecules, DiscoVerse achieved near-perfect recall ($\geq 0.99$) with moderate precision ($0.71-0.91$), while qualitative assessments of discontinuation rationale and organ-specific toxicity showed faithful, source-linked synthesis across preclinical and clinical evidence.
Authors: Harsh Rathva, Pruthwik Mishra, Shrikant Malviya
Abstract: The detection of hallucinations in multilingual scientific text generated by Large Language Models (LLMs) presents significant challenges for reliable AI systems. This paper describes our submission to the SHROOM-CAP 2025 shared task on scientific hallucination detection across 9 languages. Unlike most approaches that focus primarily on model architecture, we adopted a data-centric strategy that addressed the critical issue of training data scarcity and imbalance. We unify and balance five existing datasets to create a comprehensive training corpus of 124,821 samples (50% correct, 50% hallucinated), representing a 172x increase over the original SHROOM training data. Our approach fine-tuned XLM-RoBERTa-Large with 560 million parameters on this enhanced dataset, achieves competitive performance across all languages, including \textbf{2nd place in Gujarati} (zero-shot language) with Factuality F1 of 0.5107, and rankings between 4th-6th place across the remaining 8 languages. Our results demonstrate that systematic data curation can significantly outperform architectural innovations alone, particularly for low-resource languages in zero-shot settings.
Authors: Mohammad Aqib, Mohd Hamza, Ying Hei Chui, Qipei Mei
Abstract: Building codes contain critical information for ensuring safety, regulatory compliance, and informed decision-making in construction and engineering. Automated question answering systems over such codes enable quick and accurate access to specific regulatory clauses, improving efficiency and reducing errors. Retrieval-Augmented Generation (RAG) systems are essential for this task as they combine the precision of information retrieval with the generative capabilities of language models. However, tabular data are challenging to extract as they often involve complex layouts, merged cells, multi-row headers, and embedded semantic relationships that are not easily captured by traditional natural language processing techniques and Vision Language Models (VLMs). This paper explores and compares two methods for extracting information from tabular data in building codes using several pre-trained VLMs. First, a direct input method is used, where the image of the page is input directly into the VLMs, which are then tasked with answering questions based on the image. Second, an indirect input method is introduced, which involves converting an image of a page containing tables into the LaTeX code and then answering inquires based on the LaTeX-based input. The experiments find that the direct input method generally resulted in higher accuracy than the indirect input method. To further improve the performance, we fine-tuned each VLM using Low Rank Adaptation (LoRA) on a domain-specific tabular dataset. The fine-tuned models exhibited substantial improvements, with Qwen2.5-VL-3B-Instruct achieving relative accuracy gains exceeding 100%. Our results highlight the potential of parameter-efficient fine-tuning methods to adapt powerful VLMs for understanding complex structured data in specialized fields, such as building code interpretation and regulatory compliance.
Authors: Joseph Oladokun
Abstract: Large Language Model agents often retrieve context from knowledge bases that lack structural consistency with the agent's current reasoning state, leading to incoherent reasoning chains. We introduce Path-Constrained Retrieval (PCR), a retrieval method that combines structural graph constraints with semantic search to ensure retrieved information maintains logical relationships within a knowledge graph. PCR restricts the search space to nodes reachable from an anchor node, preventing retrieval of structurally disconnected information that may lead to inconsistent reasoning. We evaluate PCR on PathRAG-6, a benchmark spanning six domains with 180 nodes and 360 edges. Our results show that PCR achieves full structural consistency compared to 24-32 percent in baseline methods, while maintaining strong relevance scores. On the technology domain, PCR obtains full relevance at rank 10 with full structural consistency, significantly outperforming vector search and hybrid retrieval. PCR reduces the average graph distance of retrieved context by 78 percent compared to baselines, demonstrating retrieval of more structurally consistent information. These findings suggest that path-constrained retrieval is an effective approach for improving the reliability and coherence of LLM agent reasoning systems.
Authors: Syed Mohaiminul Hoque, Naimur Rahman, Md Sakhawat Hossain
Abstract: This paper introduces the approach of "Gradient Masters" for BLP-2025 Task 1: "Bangla Multitask Hate Speech Identification Shared Task". We present an ensemble-based fine-tuning strategy for addressing subtasks 1A (hate-type classification) and 1B (target group classification) in YouTube comments. We propose a hybrid approach on a Bangla Language Model, which outperformed the baseline models and secured the 6th position in subtask 1A with a micro F1 score of 73.23% and the third position in subtask 1B with 73.28%. We conducted extensive experiments that evaluated the robustness of the model throughout the development and evaluation phases, including comparisons with other Language Model variants, to measure generalization in low-resource Bangla hate speech scenarios and data set coverage. In addition, we provide a detailed analysis of our findings, exploring misclassification patterns in the detection of hate speech.
Authors: James Y. Huang, Wenxuan Zhou, Nan Xu, Fei Wang, Qin Liu, Sheng Zhang, Hoifung Poon, Muhao Chen
Abstract: The ability of Large Language Models (LLMs) to generate structured outputs that follow arbitrary schemas is crucial to a wide range of downstream tasks that require diverse structured representations of results such as information extraction, table generation, and function calling. While modern LLMs excel in generating unstructured responses in natural language, whether this advancement translates to a strong performance on text-to-structure tasks remains unclear. To bridge this gap, we first introduce OmniStruct, a comprehensive benchmark for assessing LLMs' capabilities on diverse text-to-structure tasks such as information extraction, table generation, and function calling. We build OmniStruct by identifying existing datasets across a wide range of tasks that are suitable for a structured answer format, and adapting them under a unified text-to-structure problem setting. To facilitate the development of efficient text-to-structure models, we collect high-quality training data via synthetic task generation. Without using any supervised data for OmniStruct tasks, our experiments demonstrate the possibility of fine-tuning much smaller models on synthetic data into universal structured generation models that can rival the performance of GPT-4o.
Authors: Manon Berriche
Abstract: This thesis addresses two paradoxes: (1) why empirical studies find that fake news represent only a small share of the information consulted and shared on social media despite the absence of editorial control or journalistic norms, and (2) how political polarization has intensified even though users do not appear especially receptive to fake news. To investigate these issues, two complementary studies were carried out on Twitter and Facebook, combining quantitative analyses of digital traces with online observation and interviews. This mixed-methods design avoids reducing users to single reactions to identified fake items and instead examines the variety of practices across different interactional situations, online and offline, while recording socio-demographic traits. The first study mapped users who shared at least one item labeled fake by fact-checkers in the French Twittersphere. The second used a corpus of items flagged by Facebook users to study reactions to statements whose epistemic status is uncertain. Three main findings emerge. First, sharing fake news is concentrated among a limited group of users who are not less educated or cognitively disadvantaged but are more politicized and critical of institutions; owing to their high activity and prolific sharing, they can help set the agenda for their political camp. Second, exposed users can deploy varying forms of critical distance depending on their social position and the interactional norms of the situations they inhabit: either discursive caution (prudence \'enonciative) or interventions ('points d'arr\^et') that express disagreement or corrections. Third, these forms of critical distance seldom yield genuine deliberative debates or agonistic pluralism; rather, they often produce dialogues of the deaf among a small, particularly active minority.
Authors: Heejoon Koo
Abstract: A decade of rapid advances in artificial intelligence (AI) has opened new opportunities for clinical decision support systems (CDSS), with large language models (LLMs) demonstrating strong reasoning abilities on timely medical tasks. However, clinical texts are often degraded by human errors or failures in automated pipelines, raising concerns about the reliability and fairness of AI-assisted decision-making. Yet the impact of such degradations remains under-investigated, particularly regarding how noise-induced shifts can heighten predictive uncertainty and unevenly affect demographic subgroups. We present a systematic study of state-of-the-art LLMs under diverse text corruption scenarios, focusing on robustness and equity in next-visit diagnosis prediction. To address the challenge posed by the large diagnostic label space, we introduce a clinically grounded label-reduction scheme and a hierarchical chain-of-thought (CoT) strategy that emulates clinicians' reasoning. Our approach improves robustness and reduces subgroup instability under degraded inputs, advancing the reliable use of LLMs in CDSS. We release code at https://github.com/heejkoo9/NECHOv3.
Authors: Dana Arad, Yonatan Belinkov, Hanjie Chen, Najoung Kim, Hosein Mohebbi, Aaron Mueller, Gabriele Sarti, Martin Tutek
Abstract: Mechanistic interpretability (MI) seeks to uncover how language models (LMs) implement specific behaviors, yet measuring progress in MI remains challenging. The recently released Mechanistic Interpretability Benchmark (MIB; Mueller et al., 2025) provides a standardized framework for evaluating circuit and causal variable localization. Building on this foundation, the BlackboxNLP 2025 Shared Task extends MIB into a community-wide reproducible comparison of MI techniques. The shared task features two tracks: circuit localization, which assesses methods that identify causally influential components and interactions driving model behavior, and causal variable localization, which evaluates approaches that map activations into interpretable features. With three teams spanning eight different methods, participants achieved notable gains in circuit localization using ensemble and regularization strategies for circuit discovery. With one team spanning two methods, participants achieved significant gains in causal variable localization using low-dimensional and non-linear projections to featurize activation vectors. The MIB leaderboard remains open; we encourage continued work in this standard evaluation framework to measure progress in MI research going forward.
Authors: Sultan Alrashed, Chadi Helwe, Francesco Orabona
Abstract: Although the community has tackled the acquisition of high-quality Arabic pretraining data, we still lack large-scale, multi-turn Arabic datasets that include reasoning and tool calling. Naive translation can work at the pretraining scale, but post-training demands much higher quality, which requires a stricter approach to dataset curation. In this work, we introduce SmolKalam, a translation of Smoltalk2 that uses a multi-model ensemble translation pipeline, applies quality filtering, and examines effective translation techniques for traditional decoder-only models through ablations.
Authors: Yu Xia, Sungchul Kim, Tong Yu, Ryan A. Rossi, Julian McAuely
Abstract: Agentic recommendations cast recommenders as large language model (LLM) agents that can plan, reason, use tools, and interact with users of varying preferences in web applications. However, most existing agentic recommender systems focus on generic single-agent plan-execute workflows or multi-agent task decomposition pipelines. Without recommendation-oriented design, they often underuse the collaborative signals in the user-item interaction history, leading to unsatisfying recommendation results. To address this, we propose the Multi-Agent Collaborative Filtering (MACF) framework for agentic recommendations, drawing an analogy between traditional collaborative filtering algorithms and LLM-based multi-agent collaboration. Specifically, given a target user and query, we instantiate similar users and relevant items as LLM agents with unique profiles. Each agent is able to call retrieval tools, suggest candidate items, and interact with other agents. Different from the static preference aggregation in traditional collaborative filtering, MACF employs a central orchestrator agent to adaptively manage the collaboration between user and item agents via dynamic agent recruitment and personalized collaboration instruction. Experimental results on datasets from three different domains show the advantages of our MACF framework compared to strong agentic recommendation baselines.
Authors: B. Y. Yan, Chaofan Li, Hongjin Qian, Shuqi Lu, Zheng Liu
Abstract: Memory is critical for AI agents, yet the widely-adopted static memory, aiming to create readily available memory in advance, is inevitably subject to severe information loss. To address this limitation, we propose a novel framework called \textbf{general agentic memory (GAM)}. GAM follows the principle of "\textbf{just-in time (JIT) compilation}" where it focuses on creating optimized contexts for its client at runtime while keeping only simple but useful memory during the offline stage. To this end, GAM employs a duo-design with the following components. 1) \textbf{Memorizer}, which highlights key historical information using a lightweight memory, while maintaining complete historical information within a universal page-store. 2) \textbf{Researcher}, which retrieves and integrates useful information from the page-store for its online request guided by the pre-constructed memory. This design allows GAM to effectively leverage the agentic capabilities and test-time scalability of frontier large language models (LLMs), while also facilitating end-to-end performance optimization through reinforcement learning. In our experimental study, we demonstrate that GAM achieves substantial improvement on various memory-grounded task completion scenarios against existing memory systems.
Authors: Jos\'e Pombal, Maya D'Eon, Nuno M. Guerreiro, Pedro Henrique Martins, Ant\'onio Farinhas, Ricardo Rei
Abstract: Demand for mental health support through AI chatbots is surging, though current systems present several limitations, like sycophancy or overvalidation, and reinforcement of maladaptive beliefs. A core obstacle to the creation of better systems is the scarcity of benchmarks that capture the complexity of real therapeutic interactions. Most existing benchmarks either only test clinical knowledge through multiple-choice questions or assess single responses in isolation. To bridge this gap, we present MindEval, a framework designed in collaboration with Ph.D-level Licensed Clinical Psychologists for automatically evaluating language models in realistic, multi-turn mental health therapy conversations. Through patient simulation and automatic evaluation with LLMs, our framework balances resistance to gaming with reproducibility via its fully automated, model-agnostic design. We begin by quantitatively validating the realism of our simulated patients against human-generated text and by demonstrating strong correlations between automatic and human expert judgments. Then, we evaluate 12 state-of-the-art LLMs and show that all models struggle, scoring below 4 out of 6, on average, with particular weaknesses in problematic AI-specific patterns of communication. Notably, reasoning capabilities and model scale do not guarantee better performance, and systems deteriorate with longer interactions or when supporting patients with severe symptoms. We release all code, prompts, and human evaluation data.
Authors: Tyler Shoemaker
Abstract: This position paper argues that literary scholars must engage with large language model (LLM) interpretability research. While doing so will involve ideological struggle, if not out-right complicity, the necessity of this engagement is clear: the abiding instrumentality of current approaches to interpretability cannot be the only standard by which we measure interpretation with LLMs. One site at which this struggle could take place, I suggest, is the red team.
Authors: Yacouba Diarra, Nouhoum Souleymane Coulibaly, Panga Azazia Kamat\'e, Madani Amadou Tall, Emmanuel \'Elis\'e Kon\'e, Aymane Demb\'el\'e, Michael Leventhal
Abstract: Creating speech datasets, models, and evaluation frameworks for low-resource languages remains challenging given the lack of a broad base of pertinent experience to draw from. This paper reports on the field collection of 612 hours of spontaneous speech in Bambara, a low-resource West African language; the semi-automated annotation of that dataset with transcriptions; the creation of several monolingual ultra-compact and small models using the dataset; and the automatic and human evaluation of their output. We offer practical suggestions for data collection protocols, annotation, and model design, as well as evidence for the importance of performing human evaluation. In addition to the main dataset, multiple evaluation datasets, models, and code are made publicly available.
Authors: H. M. Shadman Tabib, Jaber Ahmed Deedar
Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in natural language and code generation, and are increasingly deployed as automatic judges of model outputs and learning activities. Yet, their behavior on structured tasks such as predicting the difficulty of competitive programming problems remains under-explored. We conduct a systematic comparison of GPT-4o, used purely as a natural-language difficulty assessor, against an interpretable Light-GBM ensemble trained on explicit numeric and textual features. On a dataset of 1,825 LeetCode problems labeled Easy, Medium, or Hard, LightGBM attains 86% accuracy, whereas GPT-4o reaches only 37.75%. Detailed analyses, including confusion matrices and SHAP-based interpretability, show that numeric constraints -- such as input size limits and acceptance rates -- play a crucial role in separating Hard problems from easier ones. By contrast, GPT-4o often overlooks these cues and exhibits a strong bias toward simpler categories. We further probe GPT-4o through a synthetic Hard-problem generation protocol. Surprisingly, GPT-4o labels almost all of its own synthetic Hard problems as Medium, contradicting its tendency to downgrade real Hard problems to Easy. Our findings connect to recent work on LLMs-as-judges and automatic difficulty estimation in programming and education, and highlight concrete failure modes that must be addressed before LLM-based judges can be considered trustworthy in competitive programming, educational platforms, or reinforcement-learning pipelines.
Authors: Joseph Malone, Rachith Aiyappa, Byunghwee Lee, Haewoon Kwak, Jisun An, Yong-Yeol Ahn
Abstract: Beliefs are central to how humans reason, communicate, and form social connections, yet most computational approaches to studying them remain confined to narrow sociopolitical contexts and rely on fine-tuning for optimal performance. Despite the growing use of large language models (LLMs) across disciplines, how well these systems generalize across diverse belief domains remains unclear. We introduce a systematic, reproducible benchmark that evaluates the ability of LLMs to predict individuals' stances on a wide range of topics in a zero-shot setting using data from an online debate platform. The benchmark includes multiple informational conditions that isolate the contribution of demographic context and known prior beliefs to predictive success. Across several small- to medium-sized models, we find that providing more background information about an individual improves predictive accuracy, but performance varies substantially across belief domains. These findings reveal both the capacity and limitations of current LLMs to emulate human reasoning, advancing the study of machine behavior and offering a scalable framework for modeling belief systems beyond the sociopolitical sphere.
Authors: Mirza Raquib, Munazer Montasir Akash, Tawhid Ahmed, Saydul Akbar Murad, Farida Siddiqi Prity, Mohammad Amzad Hossain, Asif Pervez Polok, Nick Rahimi
Abstract: In our daily lives, newspapers are an essential information source that impacts how the public talks about present-day issues. However, effectively navigating the vast amount of news content from different newspapers and online news portals can be challenging. Newspaper headlines with sentiment analysis tell us what the news is about (e.g., politics, sports) and how the news makes us feel (positive, negative, neutral). This helps us quickly understand the emotional tone of the news. This research presents a state-of-the-art approach to Bangla news headline classification combined with sentiment analysis applying Natural Language Processing (NLP) techniques, particularly the hybrid transfer learning model BERT-CNN-BiLSTM. We have explored a dataset called BAN-ABSA of 9014 news headlines, which is the first time that has been experimented with simultaneously in the headline and sentiment categorization in Bengali newspapers. Over this imbalanced dataset, we applied two experimental strategies: technique-1, where undersampling and oversampling are applied before splitting, and technique-2, where undersampling and oversampling are applied after splitting on the In technique-1 oversampling provided the strongest performance, both headline and sentiment, that is 78.57\% and 73.43\% respectively, while technique-2 delivered the highest result when trained directly on the original imbalanced dataset, both headline and sentiment, that is 81.37\% and 64.46\% respectively. The proposed model BERT-CNN-BiLSTM significantly outperforms all baseline models in classification tasks, and achieves new state-of-the-art results for Bangla news headline classification and sentiment analysis. These results demonstrate the importance of leveraging both the headline and sentiment datasets, and provide a strong baseline for Bangla text classification in low-resource.
Authors: Maanas Taneja
Abstract: Language Models are extremely susceptible to performance collapse with even small changes to input prompt strings. Libraries such as DSpy (from Stanford NLP) avoid this problem through demonstration-based prompt optimisation. Inspired by this, I propose an alternative approach that treats prompt optimisation as a classical state-space search problem. I model the prompt space as a graph where nodes represent prompt states and edges correspond to deliberate transformations such as shortening, adding examples, or re- ordering content. Using beam search and random walk algorithms, I systematically explore this space, evaluating candidates on development sets and pruning unpromising branches. Across five NLP tasks (sentiment classification, question answering, summarisation, reason- ing, and natural language inference), I find that even shallow search configurations (beam width=2, depth=2) improve upon seed prompts on development sets. For instance, beam search achieves development accuracy gains from 0.40 to 0.80 on reasoning tasks, though test set improvements are more modest (0.20 to 0.50), indicating overfitting to the develop- ment heuristic. Analysis of successful optimisation paths reveals that transformations that make prompts concise appear most frequently, while verbosity operators are never selected. My results validate prompt optimization as a search problem and suggest that with greater computational resources and improved evaluation metrics, deeper exploration could yield more robust prompts that generalize beyond development sets. Code and implementation are available at [https://github.com/MaanasTaneja/PromptOptimiser].
Authors: Michael J. Bommarito II
Abstract: We present OpenGloss, a synthetic encyclopedic dictionary and semantic knowledge graph for English that integrates lexicographic definitions, encyclopedic context, etymological histories, and semantic relationships in a unified resource. OpenGloss contains 537K senses across 150K lexemes, on par with WordNet 3.1 and Open English WordNet, while providing more than four times as many sense definitions. These lexemes include 9.1M semantic edges, 1M usage examples, 3M collocations, and 60M words of encyclopedic content. Generated through a multi-agent procedural generation pipeline with schema-validated LLM outputs and automated quality assurance, the entire resource was produced in under one week for under $1,000. This demonstrates that structured generation can create comprehensive lexical resources at cost and time scales impractical for manual curation, enabling rapid iteration as foundation models improve. The resource addresses gaps in pedagogical applications by providing integrated content -- definitions, examples, collocations, encyclopedias, etymology -- that supports both vocabulary learning and natural language processing tasks. As a synthetically generated resource, OpenGloss reflects both the capabilities and limitations of current foundation models. The dataset is publicly available on Hugging Face under CC-BY 4.0, enabling researchers and educators to build upon and adapt this resource.
Authors: Shireen Chand, Faith Baca, Emilio Ferrara
Abstract: Large Language Models (LLMs) inherit societal biases from their training data, potentially leading to harmful or unfair outputs. While various techniques aim to mitigate these biases, their effects are often evaluated only along the dimension of the bias being targeted. This work investigates the cross-category consequences of targeted bias mitigation. We study four bias mitigation techniques applied across ten models from seven model families, and we explore racial, religious, profession- and gender-related biases. We measure the impact of debiasing on model coherence and stereotypical preference using the StereoSet benchmark. Our results consistently show that while targeted mitigation can sometimes reduce bias in the intended dimension, it frequently leads to unintended and often negative consequences in others, such as increasing model bias and decreasing general coherence. These findings underscore the critical need for robust, multi-dimensional evaluation tools when examining and developing bias mitigation strategies to avoid inadvertently shifting or worsening bias along untargeted axes.
Authors: Goun Pyeon, Inbum Heo, Jeesu Jung, Taewook Hwang, Hyuk Namgoong, Hyein Seo, Yerim Han, Eunbin Kim, Hyeonseok Kang, Sangkeun Jung
Abstract: This study systematically evaluated the mathematical reasoning capabilities of Large Language Models (LLMs) using the 2026 Korean College Scholastic Ability Test (CSAT) Mathematics section, ensuring a completely contamination-free evaluation environment. To address data leakage issues in existing benchmarks, we digitized all 46 questions (22 common and 24 elective) within two hours of the exam's public release, eliminating any possibility of inclusion in model training data. We conducted comprehensive evaluations of 24 state-of-the-art LLMs across varying input modalities (text, image, text+figure) and prompt languages (Korean, English). GPT-5 Codex achieved the only perfect score (100 points) with text input and Korean prompts, while Grok 4, GPT-5, and Deepseek R1 scored above 95 points. Notably, gpt-oss-20B achieved 95.7 points despite its relatively small size, demonstrating high cost-effectiveness. Problem-specific analysis revealed geometry as the weakest domain (77.7% average) with significant performance degradation on 4-point high-difficulty problems. Text input consistently outperformed image input, while prompt language effects varied by model scale. In reasoning enhancement experiments with GPT-5 series, increased reasoning intensity improved performance (from 82.6 to 100 points) but quadrupled token usage and drastically reduced efficiency, suggesting that models with minimal reasoning may be more practical. This research contributes: (1) implementation of a completely unexposed evaluation environment, (2) a real-exam-based LLM assessment framework, and (3) a practical evaluation perspective integrating performance, cost, and time considerations. Detailed results and model comparisons are available at the 2026 Korean CSAT LLM Evaluation Leaderboard (https://isoft.cnu.ac.kr/csat2026/).
Authors: Jie He, Richard He Bai, Sinead Williamson, Jeff Z. Pan, Navdeep Jaitly, Yizhe Zhang
Abstract: Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge but still suffers from long contexts and disjoint retrieval-generation optimization. In this work, we propose CLaRa (Continuous Latent Reasoning), a unified framework that performs embedding-based compression and joint optimization in a shared continuous space. To obtain semantically rich and retrievable compressed vectors, we introduce SCP, a key-preserving data synthesis framework using QA and paraphrase supervision. CLaRa then trains the reranker and generator end-to-end via a single language modeling loss, with gradients flowing through both modules using a differentiable top-k estimator. Theoretically, this unified optimization aligns retrieval relevance with answer quality. Experiments across multiple QA benchmarks show that CLaRa achieves state-of-the-art compression and reranking performance, often surpassing text-based fine-tuned baselines.
Authors: Wangjiaxuan Xin
Abstract: This report presents the Empathetic Cascading Networks (ECN) framework, a multi-stage prompting method designed to enhance the empathetic and inclusive capabilities of large language models. ECN employs four stages: Perspective Adoption, Emotional Resonance, Reflective Understanding, and Integrative Synthesis, to guide models toward generating emotionally resonant and contextually aware responses. Experimental results demonstrate that ECN achieves the highest Empathy Quotient (EQ) scores across GPT-3.5-turbo and GPT-4, while maintaining competitive Regard and Perplexity metrics. These findings emphasize ECN's potential for applications requiring empathy and inclusivity in conversational AI.
Authors: Yu Lei, Shuzheng Si, Wei Wang, Yifei Wu, Gang Chen, Fanchao Qi, Maosong Sun
Abstract: Large language models are evolving from single-turn responders into tool-using agents capable of sustained reasoning and decision-making for deep research. Prevailing systems adopt a linear pipeline of plan to search to write to a report, which suffers from error accumulation and context rot due to the lack of explicit control over both model behavior and context. We introduce RhinoInsight, a deep research framework that adds two control mechanisms to enhance robustness, traceability, and overall quality without parameter updates. First, a Verifiable Checklist module transforms user requirements into traceable and verifiable sub-goals, incorporates human or LLM critics for refinement, and compiles a hierarchical outline to anchor subsequent actions and prevent non-executable planning. Second, an Evidence Audit module structures search content, iteratively updates the outline, and prunes noisy context, while a critic ranks and binds high-quality evidence to drafted content to ensure verifiability and reduce hallucinations. Our experiments demonstrate that RhinoInsight achieves state-of-the-art performance on deep research tasks while remaining competitive on deep search tasks.
Authors: Matthew R. DeVerna, Kai-Cheng Yang, Harry Yaojun Yan, Filippo Menczer
Abstract: Large language models (LLMs) have raised hopes for automated end-to-end fact-checking, but prior studies report mixed results. As mainstream chatbots increasingly ship with reasoning capabilities and web search tools -- and millions of users already rely on them for verification -- rigorous evaluation is urgent. We evaluate 15 recent LLMs from OpenAI, Google, Meta, and DeepSeek on more than 6,000 claims fact-checked by PolitiFact, comparing standard models with reasoning- and web-search variants. Standard models perform poorly, reasoning offers minimal benefits, and web search provides only moderate gains, despite fact-checks being available on the web. In contrast, a curated RAG system using PolitiFact summaries improved macro F1 by 233% on average across model variants. These findings suggest that giving models access to curated high-quality context is a promising path for automated fact-checking.
Authors: Daiqing Wu, Dongbao Yang, Can Ma, Yu Zhou
Abstract: As posts on social media increase rapidly, analyzing the sentiments embedded in image-text pairs has become a popular research topic in recent years. Although existing works achieve impressive accomplishments in simultaneously harnessing image and text information, they lack the considerations of possible low-quality and missing modalities. In real-world applications, these issues might frequently occur, leading to urgent needs for models capable of predicting sentiment robustly. Therefore, we propose a Distribution-based feature Recovery and Fusion (DRF) method for robust multimodal sentiment analysis of image-text pairs. Specifically, we maintain a feature queue for each modality to approximate their feature distributions, through which we can simultaneously handle low-quality and missing modalities in a unified framework. For low-quality modalities, we reduce their contributions to the fusion by quantitatively estimating modality qualities based on the distributions. For missing modalities, we build inter-modal mapping relationships supervised by samples and distributions, thereby recovering the missing modalities from available ones. In experiments, two disruption strategies that corrupt and discard some modalities in samples are adopted to mimic the low-quality and missing modalities in various real-world scenarios. Through comprehensive experiments on three publicly available image-text datasets, we demonstrate the universal improvements of DRF compared to SOTA methods under both two strategies, validating its effectiveness in robust multimodal sentiment analysis.
Authors: Bashar Talafha, Amin Abu Alhassan, Muhammad Abdul-Mageed
Abstract: Low-resource ASR remains a challenging problem, especially for languages like Arabic that exhibit wide dialectal variation and limited labeled data. We propose context-aware prompting strategies to adapt OpenAI's Whisper for Arabic speech recognition without retraining. Our methods include decoder prompting with first-pass transcriptions or retrieved utterances, and encoder prefixing using speech synthesized in the target speaker's voice. We introduce techniques such as prompt reordering, speaker-aware prefix synthesis, and modality-specific retrieval (lexical, semantic, acoustic) to improve transcription in real-world, zero-shot settings. Evaluated on nine Arabic linguistic conditions, our approach reduces WER by up to 22.3% on Modern Standard Arabic and 9.2% on dialectal speech, significantly mitigating hallucinations and speaker mismatch.
Authors: Cao Linxiao, Wang Ruitao, Li Jindong, Zhou Zhipeng, Yang Menglin
Abstract: Retrieval-augmented generation (RAG) enables large language models (LLMs) to access external knowledge, helping mitigate hallucinations and enhance domain-specific expertise. Graph-based RAG enhances structural reasoning by introducing explicit relational organization that enables information propagation across semantically connected text units. However, these methods typically rely on Euclidean embeddings that capture semantic similarity but lack a geometric notion of hierarchical depth, limiting their ability to represent abstraction relationships inherent in complex knowledge graphs. To capture both fine-grained semantics and global hierarchy, we propose HyperbolicRAG, a retrieval framework that integrates hyperbolic geometry into graph-based RAG. HyperbolicRAG introduces three key designs: (1) a depth-aware representation learner that embeds nodes within a shared Poincare manifold to align semantic similarity with hierarchical containment, (2) an unsupervised contrastive regularization that enforces geometric consistency across abstraction levels, and (3) a mutual-ranking fusion mechanism that jointly exploits retrieval signals from Euclidean and hyperbolic spaces, emphasizing cross-space agreement during inference. Extensive experiments across multiple QA benchmarks demonstrate that HyperbolicRAG outperforms competitive baselines, including both standard RAG and graph-augmented baselines.
Authors: Kaize Shi, Xueyao Sun, Xiaohui Tao, Lin Li, Qika Lin, Guandong Xu
Abstract: Large Language Models (LLMs) face information overload when handling long contexts, particularly in Retrieval-Augmented Generation (RAG) where extensive supporting documents often introduce redundant content. This issue not only weakens reasoning accuracy but also increases computational overhead. We propose an unsupervised context compression framework that exploits Abstract Meaning Representation (AMR) graphs to preserve semantically essential information while filtering out irrelevant text. By quantifying node-level entropy within AMR graphs, our method estimates the conceptual importance of each node, enabling the retention of core semantics. Specifically, we construct AMR graphs from raw contexts, compute the conceptual entropy of each node, and screen significant informative nodes to form a condensed and semantically focused context than raw documents. Experiments on the PopQA and EntityQuestions datasets show that our method outperforms vanilla and other baselines, achieving higher accuracy while substantially reducing context length. To the best of our knowledge, this is the first work introducing AMR-based conceptual entropy for context compression, demonstrating the potential of stable linguistic features in context engineering.
Authors: Heger Arfaoui, Mohammed Iheb Hergli, Beya Benzina, Slimane BenMiled
Abstract: Focus group discussions generate rich qualitative data but their analysis traditionally relies on labor-intensive manual coding that limits scalability and reproducibility. We present a rigorous, reproducible computational framework for applying neural topic modeling to focus group transcripts, addressing fundamental methodological challenges: hyperparameter sensitivity, model stability, and validation of interpretability. Using BERTopic applied to ten focus groups exploring HPV vaccine perceptions in Tunisia (1,076 utterances), we conducted systematic evaluation across 27 hyperparameter configurations, assessed stability through bootstrap resampling with 30 replicates per configuration, and validated interpretability through formal human evaluation by three domain experts. Our analysis demonstrates substantial sensitivity to hyperparameter choices and reveals that metric selection for stability assessment must align with analytical goals. A hierarchical merging strategy (extracting fine-grained topics for stability then consolidating for interpretability) effectively navigates the stability-coherence tradeoff, achieving coherence of 0.558 compared to 0.539 for direct extraction. Human validation confirmed topic quality with very good inter-rater reliability (ICC = 0.79, weighted Cohen's kappa = 0.578). Our framework provides practical guidelines that researchers can adapt to their own qualitative research contexts. All code, data processing scripts, and evaluation protocols are publicly available to support reproduction and extension of this work.
Authors: V\'aclav Tran, Jakub \v{S}m\'id, Ladislav Lenc, Jean-Pierre Salmon, Pavel Kr\'al
Abstract: Text summarization is the task of automatically condensing longer texts into shorter, coherent summaries while preserving the original meaning and key information. Although this task has been extensively studied in English and other high-resource languages, Czech summarization, particularly in the context of historical documents, remains underexplored. This is largely due to the inherent linguistic complexity of Czech and the lack of high-quality annotated datasets. In this work, we address this gap by leveraging the capabilities of Large Language Models (LLMs), specifically Mistral and mT5, which have demonstrated strong performance across a wide range of natural language processing tasks and multilingual settings. In addition, we also propose a translation-based approach that first translates Czech texts into English, summarizes them using an English-language model, and then translates the summaries back into Czech. Our study makes the following main contributions: We demonstrate that LLMs achieve new state-of-the-art results on the SumeCzech dataset, a benchmark for modern Czech text summarization, showing the effectiveness of multilingual LLMs even for morphologically rich, medium-resource languages like Czech. We introduce a new dataset, Posel od \v{C}erchova, designed for the summarization of historical Czech texts. This dataset is derived from digitized 19th-century publications and annotated for abstractive summarization. We provide initial baselines using modern LLMs to facilitate further research in this underrepresented area. By combining cutting-edge models with both modern and historical Czech datasets, our work lays the foundation for further progress in Czech summarization and contributes valuable resources for future research in Czech historical document processing and low-resource summarization more broadly.
Authors: Fengyuan Liu, Huang Yi, Sichun Luo, Yuqi Wang, Yazheng Yang, Xinye Li, Zefa Hu, Junlan Feng, Qi Liu
Abstract: Discovering effective predictive signals, or ``alphas,'' from financial data with high dimensionality and extremely low signal-to-noise ratio remains a difficult open problem. Despite progress in deep learning, genetic programming, and, more recently, large language model (LLM)--based factor generation, existing approaches still explore only a narrow region of the vast alpha search space. Neural models tend to produce opaque and fragile patterns, while symbolic or formula-based methods often yield redundant or economically ungrounded expressions that generalize poorly. Although different in form, these paradigms share a key limitation: none can conduct broad, structured, and human-like exploration that balances logical consistency with creative leaps. To address this gap, we introduce the Cognitive Alpha Mining Framework (CogAlpha), which combines code-level alpha representation with LLM-driven reasoning and evolutionary search. Treating LLMs as adaptive cognitive agents, our framework iteratively refines, mutates, and recombines alpha candidates through multi-stage prompts and financial feedback. This synergistic design enables deeper thinking, richer structural diversity, and economically interpretable alpha discovery, while greatly expanding the effective search space. Experiments on A-share equities demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and generalization over existing methods. Our results highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery. All source code will be released.
Authors: Masoomali Fatehkia, Enes Altinisik, Husrev Taha Sencar
Abstract: Content moderation filters are a critical safeguard against alignment failures in language models. Yet most existing filters focus narrowly on general safety and overlook cultural context. In this work, we introduce FanarGuard, a bilingual moderation filter that evaluates both safety and cultural alignment in Arabic and English. We construct a dataset of over 468K prompt and response pairs, drawn from synthetic and public datasets, scored by a panel of LLM judges on harmlessness and cultural awareness, and use it to train two filter variants. To rigorously evaluate cultural alignment, we further develop the first benchmark targeting Arabic cultural contexts, comprising over 1k norm-sensitive prompts with LLM-generated responses annotated by human raters. Results show that FanarGuard achieves stronger agreement with human annotations than inter-annotator reliability, while matching the performance of state-of-the-art filters on safety benchmarks. These findings highlight the importance of integrating cultural awareness into moderation and establish FanarGuard as a practical step toward more context-sensitive safeguards.
Authors: Xingyu Huang, Fei Jiang, Jianli Xiao
Abstract: With the rapid development of large language models (LLMs), the applications of LLMs have grown substantially. In the education domain, LLMs demonstrate significant potential, particularly in automatic text generation, which enables the creation of intelligent and adaptive learning content. This paper proposes a new LLMs framework, which is named as Reading Comprehension Exercise Generation (RCEG). It can generate high-quality and personalized English reading comprehension exercises automatically. Firstly, RCEG uses fine-tuned LLMs to generate content candidates. Then, it uses a discriminator to select the best candidate. Finally, the quality of the generated content has been improved greatly. To evaluate the performance of RCEG, a dedicated dataset for English reading comprehension is constructed to perform the experiments, and comprehensive evaluation metrics are used to analyze the experimental results. These metrics include content diversity, factual accuracy, linguistic toxicity, and pedagogical alignment. Experimental results show that RCEG significantly improves the relevance and cognitive appropriateness of the generated exercises.
Authors: Yang Xiang, Yixin Ji, Juntao Li, Min Zhang
Abstract: Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning benchmarks. However, their long chain-of-thought reasoning processes incur significant inference overhead. Pruning has emerged as a promising approach to reducing computational costs. However, existing efforts have primarily focused on large language models (LLMs), while pruning LRMs remains unexplored. In this work, we conduct the first empirical study on pruning LRMs and show that directly applying existing pruning techniques fails to yield satisfactory results. Our findings indicate that using self-generated reasoning data for calibration can substantially improve pruning performance. We further investigate how the difficulty and length of reasoning data affect pruning outcomes. Our analysis reveals that challenging and moderately long self-generated reasoning data serve as ideal calibration data. Based on these insights, we propose a Selective Self-Generated Reasoning (SSGR) data construction strategy to provide effective calibration data for pruning LRMs. Experimental results on the DeepSeek-R1-Distill model series validate that our strategy improves the reasoning ability of pruned LRMs by 10%-13% compared to general pruning methods.
Authors: Jingqian Zhao, Bingbing Wang, Geng Tu, Yice Zhang, Qianlong Wang, Bin Liang, Jing Li, Ruifeng Xu
Abstract: Data contamination poses a significant challenge to the fairness of LLM evaluations in natural language processing tasks by inadvertently exposing models to test data during training. Current studies attempt to mitigate this issue by modifying existing datasets or generating new ones from freshly collected information. However, these methods fall short of ensuring contamination-resilient evaluation, as they fail to fully eliminate pre-existing knowledge from models or preserve the semantic complexity of the original datasets. To address these limitations, we propose \textbf{CoreEval}, a \textbf{Co}ntamination-\textbf{re}silient \textbf{Eval}uation strategy for automatically updating data with real-world knowledge. This approach begins by extracting entity relationships from the original data and leveraging the GDELT database to retrieve relevant, up-to-date knowledge. The retrieved knowledge is then recontextualized and integrated with the original data, which is refined and restructured to ensure semantic coherence and enhanced task relevance. Ultimately, a robust data reflection mechanism is employed to iteratively verify and refine labels, ensuring consistency between the updated and original datasets. Extensive experiments on updated datasets validate the robustness of CoreEval, demonstrating its effectiveness in mitigating performance overestimation caused by data contamination.
Authors: Adam Rychert, Gasper Spagnolo, Evgenii Posashkov
Abstract: In this reproducibility study, we revisit the LLAMBO framework of Daxberger et al. (2024), a prompting-based Bayesian optimization (BO) method that uses large language models as discriminative surrogates and acquisition optimizers via text-only interactions. We replicate the core Bayesmark and HPOBench experiments under the original evaluation protocol, but replace GPT-3.5 with the open-weight Llama 3.1 70B model used for all text encoding components. Our results broadly confirm the main claims of LLAMBO. Contextual warm starting via textual problem and hyperparameter descriptions substantially improves early regret behaviour and reduces variance across runs. LLAMBO's discriminative surrogate is weaker than GP or SMAC as a pure single task regressor, yet benefits from cross task semantic priors induced by the language model. Ablations that remove textual context markedly degrade predictive accuracy and calibration, while the LLAMBO candidate sampler consistently generates higher quality and more diverse proposals than TPE or random sampling. Experiments with smaller backbones (Gemma 27B, Llama 3.1 8B) yield unstable or invalid predictions, suggesting insufficient capacity for reliable surrogate behaviour. Overall, our study shows that the LLAMBO architecture is robust to changing the language model backbone and remains effective when instantiated with Llama 3.1 70B.
Authors: Sahil Kale
Abstract: Modern large language models integrate web search to provide real-time answers, yet it remains unclear whether they are efficiently calibrated to use search when it is actually needed. We introduce a benchmark evaluating both the necessity and effectiveness of web access across commercial models with no access to internal states or parameters. The dataset includes a static split of 783 temporally anchored questions answerable from pre-cutoff knowledge, aimed at testing whether models invoke search based on low internal confidence, and a dynamic split of 288 post-cutoff queries designed to test whether models recognise when search is required and retrieve updated information. Web access substantially improves static accuracy for GPT-5-mini and Claude Haiku 4.5, though confidence calibration worsens. On dynamic queries, both models frequently invoke search yet remain below 70 percent accuracy due to weak query formulation. Costs per accuracy-improving call remain low, but returns diminish once initial retrieval fails. Selective invocation helps, but models become overconfident and inconsistent after search. Overall, built-in web search meaningfully improves factual accuracy and can be invoked selectively, yet models remain overconfident, skip retrieval when it is essential, and falter once initial search queries underperform. Taken together, internal web search works better as a good low-latency verification layer than a reliable analytical tool, with clear room for improvement.
Authors: Yuchen Ji, Bo Xu, Jie Shi, Jiaqing Liang, Deqing Yang, Yu Mao, Hai Chen, Yanghua Xiao
Abstract: The task of translating natural language questions into query languages has long been a central focus in semantic parsing. Recent advancements in Large Language Models (LLMs) have significantly accelerated progress in this field. However, existing studies typically focus on a single query language, resulting in methods with limited generalizability across different languages. In this paper, we formally define the Text-to-Query task paradigm, unifying semantic parsing tasks across various query languages. We identify query skeletons as a shared optimization target of Text-to-Query tasks, and propose a general dynamic data augmentation framework that explicitly diagnoses model-specific weaknesses in handling these skeletons to synthesize targeted training data. Experiments on four Text-to-Query benchmarks demonstrate that our method achieves state-of-the-art performance using only a small amount of synthesized data, highlighting the efficiency and generality of our approach and laying a solid foundation for unified research on Text-to-Query tasks. We release our code at https://github.com/jjjycaptain/Skeletron.
Authors: Francois Vandenhende, Anna Georgiou, Michalis Georgiou, Theodoros Psaras, Ellie Karekla, Elena Hadjicosta
Abstract: We present a graphical, knowledge-based method for reviewing treatment-emergent adverse events (AEs) in clinical trials. The approach enhances MedDRA by adding a hidden medical knowledge layer (Safeterm) that captures semantic relationships between terms in a 2-D map. Using this layer, AE Preferred Terms can be regrouped automatically into similarity clusters, and their association to the trial disease may be quantified. The Safeterm map is available online and connected to aggregated AE incidence tables from ClinicalTrials.gov. For signal detection, we compute treatment-specific disproportionality metrics using shrinkage incidence ratios. Cluster-level EBGM values are then derived through precision-weighted aggregation. Two visual outputs support interpretation: a semantic map showing AE incidence and an expectedness-versus-disproportionality plot for rapid signal detection. Applied to three legacy trials, the automated method clearly recovers all expected safety signals. Overall, augmenting MedDRA with a medical knowledge layer improves clarity, efficiency, and accuracy in AE interpretation for clinical trials.
Authors: Hayami Takahashi, Kensuke Takahashi
Abstract: In expressing emotions, as an expression form separate from natural language, we propose an alternative form that complements natural language, acting as a proxy or window for emotional states. First, we set up an expression form "Effect of Contradictory Structure." "Effect of Contradictory Structure" is not static but dynamic. Effect in "Effect of Contradictory Structure" is unpleasant or pleasant, and the orientation to avoid that unpleasantness is considered pseudo-expression of will. Second, "Effect of Contradictory Structure" can be overlapped with each other. This overlapping operation is called "montage." A broader "Structure" that includes related "Effect of Contradictory Structure" and "Effect of Structure" are set up. Montage produces "Effect of Structure". In montage, it is necessary to set something like "strength," so we adopted Deleuze and Deleuze/Guattari's word "intensity" and set it as an element of our model. We set up a general theoretical framework - Word Import Between Systems (Models) and justified the import of "intensity" through Austin's use of the word "force." "Effect of Structure" process is demonstrated using the example of proceeding to the next level of education.
Authors: Yutong Li, Yitian Zhou, Xudong Wang, GuoChen, Caiyan Qin
Abstract: Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, including multi-step reasoning such as mathematical proving. However, existing approaches often lack an explicit and dynamic mechanism to structurally represent and evolve intermediate reasoning states, which limits their ability to perform context-aware theorem selection and iterative conclusion generation. To address these challenges, we propose GraphMind, a novel dynamic graph-based framework that integrates the graph neural network (GNN) with LLMs to iteratively select theorems and generate intermediate conclusions for multi-step reasoning. Our method models the reasoning process as a heterogeneous evolving graph, where nodes represent conditions, theorems, and conclusions, while edges capture logical dependencies between nodes. By encoding the current reasoning state with GNN and leveraging semantic matching for theorem selection, our framework enables context-aware, interpretable, and structured reasoning in a closed-loop manner. Experiments on various question-answering (QA) datasets demonstrate that our proposed GraphMind method achieves consistent performance improvements and significantly outperforms existing baselines in multi-step reasoning, validating the effectiveness and generalizability of our approach.
Authors: Wenxuan Mu, Jinzhong Ning, Di Zhao, Yijia Zhang
Abstract: In-context learning (ICL) with large language models (LLMs) has emerged as a promising paradigm for named entity recognition (NER) in low-resource scenarios. However, existing ICL-based NER methods suffer from three key limitations: (1) reliance on dynamic retrieval of annotated examples, which is problematic when annotated data is scarce; (2) limited generalization to unseen domains due to the LLM's insufficient internal domain knowledge; and (3) failure to incorporate external knowledge or resolve entity ambiguities. To address these challenges, we propose KDR-Agent, a novel multi-agent framework for multi-domain low-resource in-context NER that integrates Knowledge retrieval, Disambiguation, and Reflective analysis. KDR-Agent leverages natural-language type definitions and a static set of entity-level contrastive demonstrations to reduce dependency on large annotated corpora. A central planner coordinates specialized agents to (i) retrieve factual knowledge from Wikipedia for domain-specific mentions, (ii) resolve ambiguous entities via contextualized reasoning, and (iii) reflect on and correct model predictions through structured self-assessment. Experiments across ten datasets from five domains demonstrate that KDR-Agent significantly outperforms existing zero-shot and few-shot ICL baselines across multiple LLM backbones. The code and data can be found at https://github.com/MWXGOD/KDR-Agent.
Authors: Ziyuan Gao, Di Liang, Xianjie Wu, Philippe Morel, Minlong Peng
Abstract: Existing reinforcement learning methods for Chain-of-Thought reasoning suffer from two critical limitations. First, they operate as monolithic black boxes that provide undifferentiated reward signals, obscuring individual step contributions and hindering error diagnosis. Second, sequential decoding has O(n) time complexity. This makes real-time deployment impractical for complex reasoning tasks. We present DeCoRL (Decoupled Reasoning Chains via Coordinated Reinforcement Learning), a novel framework that transforms reasoning from sequential processing into collaborative modular orchestration. DeCoRL trains lightweight specialized models to generate reasoning sub-steps concurrently, eliminating sequential bottlenecks through parallel processing. To enable precise error attribution, the framework designs modular reward functions that score each sub-step independently. Cascaded DRPO optimization then coordinates these rewards while preserving inter-step dependencies. Comprehensive evaluation demonstrates state-of-the-art results across RM-Bench, RMB, and RewardBench, outperforming existing methods including large-scale models. DeCoRL delivers 3.8 times faster inference while maintaining superior solution quality and offers a 22.7\% improvement in interpretability through explicit reward attribution. These advancements, combined with a 72.4\% reduction in energy consumption and a 68\% increase in throughput, make real-time deployment of complex reasoning systems a reality.
Authors: Juan-Jos\'e Guzm\'an-Landa, Jes\'us V\'azquez-Osorio, Juan-Manuel Torres-Moreno, Ligia Quintana Torres, Miguel Figueroa-Saavedra, Martha-Lorena Avenda\~no-Garrido, Graham Ranger, Patricia Vel\'azquez-Morales, Gerardo Eugenio Sierra Mart\'inez
Abstract: In this paper, we describe a symbolic model for the automatic orthographic unification of Nawatl text documents. Our model is based on algorithms that we have previously used to analyze sentences in Nawatl, and on the corpus called $\pi$-yalli, consisting of texts in several Nawatl orthographies. Our automatic unification algorithm implements linguistic rules in symbolic regular expressions. We also present a manual evaluation protocol that we have proposed and implemented to assess the quality of the unified sentences generated by our algorithm, by testing in a sentence semantic task. We have obtained encouraging results from the evaluators for most of the desired features of our artificially unified sentences
Authors: Phong Le, Mees Lindeman, Raquel G. Alhama
Abstract: The structure of naming systems in natural languages hinges on a trade-off between high informativeness and low complexity. Prior work capitalizes on information theory to formalize these notions; however, these studies generally rely on two simplifications: (i) optimal listeners, and (ii) universal communicative need across languages. Here, we address these limitations by introducing an information-theoretic framework for discrete object naming systems, and we use it to prove that an optimal trade-off is achievable if and only if the listener's decoder is equivalent to the Bayesian decoder of the speaker. Adopting a referential game setup from emergent communication, and focusing on the semantic domain of kinship, we show that our notion of optimality is not only theoretically achievable but also emerges empirically in learned communication systems.
Authors: Yaping Chai, Haoran Xie, Joe S. Qin
Abstract: Aspect category sentiment analysis (ACSA) has achieved remarkable progress with large language models (LLMs), yet existing approaches primarily emphasize sentiment polarity while overlooking the underlying emotional dimensions that shape sentiment expressions. This limitation hinders the model's ability to capture fine-grained affective signals toward specific aspect categories. To address this limitation, we introduce a novel emotion-enhanced multi-task ACSA framework that jointly learns sentiment polarity and category-specific emotions grounded in Ekman's six basic emotions. Leveraging the generative capabilities of LLMs, our approach enables the model to produce emotional descriptions for each aspect category, thereby enriching sentiment representations with affective expressions. Furthermore, to ensure the accuracy and consistency of the generated emotions, we introduce an emotion refinement mechanism based on the Valence-Arousal-Dominance (VAD) dimensional framework. Specifically, emotions predicted by the LLM are projected onto a VAD space, and those inconsistent with their corresponding VAD coordinates are re-annotated using a structured LLM-based refinement strategy. Experimental results demonstrate that our approach significantly outperforms strong baselines on all benchmark datasets. This underlines the effectiveness of integrating affective dimensions into ACSA.
Authors: Zijian Wang, Yanxiang Ma, Chang Xu
Abstract: Chain-of-Thought (CoT) reasoning is a critical capability for large language models (LLMs), enabling them to tackle com- plex multi-step tasks. While base LLMs, pre-trained on general text corpora, often struggle with reasoning due to a lack of specialized training, recent studies reveal their latent reason- ing potential tied to hidden states. However, existing hidden state manipulation methods, such as linear activation steering, suffer from limitations due to their rigid and unconstrained nature, often leading to distribution shifts and degraded text quality. In this work, we propose a novel approach for elic- iting CoT reasoning from base LLMs through hidden state manipulation grounded in probabilistic conditional generation. By reformulating the challenge as an optimization problem with a balanced likelihood and prior regularization framework, our method guides hidden states toward reasoning-oriented trajectories while preserving linguistic coherence. Extensive evaluations across mathematical, commonsense, and logical reasoning benchmarks demonstrate that our approach con- sistently outperforms existing steering methods, offering a theoretically principled and effective solution for enhancing reasoning capabilities in base LLMs.
Authors: Samantha Dies, Courtney Maynard, Germans Savcisens, Tina Eliassi-Rad
Abstract: Large language models (LLMs) are widely used for factual tasks such as "What treats asthma?" or "What is the capital of Latvia?". However, it remains unclear how stably LLMs encode distinctions between true, false, and neither-true-nor-false content in their internal probabilistic representations. We introduce representational stability as the robustness of an LLM's veracity representations to perturbations in the operational definition of truth. We assess representational stability by (i) training a linear probe on an LLM's activations to separate true from not-true statements and (ii) measuring how its learned decision boundary shifts under controlled label changes. Using activations from sixteen open-source models and three factual domains, we compare two types of neither statements. The first are fact-like assertions about entities we believe to be absent from any training data. We call these unfamiliar neither statements. The second are nonfactual claims drawn from well-known fictional contexts. We call these familiar neither statements. The unfamiliar statements induce the largest boundary shifts, producing up to $40\%$ flipped truth judgements in fragile domains (such as word definitions), while familiar fictional statements remain more coherently clustered and yield smaller changes ($\leq 8.2\%$). These results suggest that representational stability stems more from epistemic familiarity than from linguistic form. More broadly, our approach provides a diagnostic for auditing and training LLMs to preserve coherent truth assignments under semantic uncertainty, rather than optimizing for output accuracy alone.
Authors: Christos-Nikolaos Zacharopoulos, Revekka Kyriakoglou
Abstract: How and where does a transformer notice that a sentence has gone semantically off the rails? To explore this question, we evaluated the causal language model (phi-2) using a carefully curated corpus, with sentences that concluded plausibly or implausibly. Our analysis focused on the hidden states sampled at each model layer. To investigate how violations are encoded, we utilized two complementary probes. First, we conducted a per-layer detection using a linear probe. Our findings revealed that a simple linear decoder struggled to distinguish between plausible and implausible endings in the lowest third of the model's layers. However, its accuracy sharply increased in the middle blocks, reaching a peak just before the top layers. Second, we examined the effective dimensionality of the encoded violation. Initially, the violation widens the representational subspace, followed by a collapse after a mid-stack bottleneck. This might indicate an exploratory phase that transitions into rapid consolidation. Taken together, these results contemplate the idea of alignment with classical psycholinguistic findings in human reading, where semantic anomalies are detected only after syntactic resolution, occurring later in the online processing sequence.
Authors: Md. Tanzim Ferdous, Naeem Ahsan Chowdhury, Prithwiraj Bhattacharjee
Abstract: This study developed a new Bangla abstractive summarization dataset to generate concise summaries of Bangla articles from diverse sources. Most existing studies in this field have concentrated on news articles, where journalists usually follow a fixed writing style. While such approaches are effective in limited contexts, they often fail to adapt to the varied nature of real-world Bangla texts. In today's digital era, a massive amount of Bangla content is continuously produced across blogs, newspapers, and social media. This creates a pressing need for summarization systems that can reduce information overload and help readers understand content more quickly. To address this challenge, we developed a dataset of over 54,000 Bangla articles and summaries collected from multiple sources, including blogs such as Cinegolpo and newspapers such as Samakal and The Business Standard. Unlike single-domain resources, our dataset spans multiple domains and writing styles. It offers greater adaptability and practical relevance. To establish strong baselines, we trained and evaluated this dataset using several deep learning and transfer learning models, including LSTM, BanglaT5-small, and MTS-small. The results highlight its potential as a benchmark for future research in Bangla natural language processing. This dataset provides a solid foundation for building robust summarization systems and helps expand NLP resources for low-resource languages.
Authors: Shaltiel Shmidman, Asher Fredman, Oleg Sudakov, Meriem Bendris
Abstract: Test-time scaling, which leverages additional computation during inference to improve model accuracy, has enabled a new class of Large Language Models (LLMs) that are able to reason through complex problems by understanding the goal, turning this goal into a plan, working through intermediate steps, and checking their own work before answering . Frontier large language models with reasoning capabilities, such as DeepSeek-R1 and OpenAI's gpt-oss, follow the same procedure when solving complex problems by generating intermediate reasoning traces before giving the final answer. Today, these models are being increasingly used to generate reasoning traces that serve as high-quality supervised data for post-training of small and medium-sized language models to teach reasoning capabilities without requiring expensive human curation. In this work, we compare the performance of medium-sized LLMs on Math problems after post-training on two kinds of reasoning traces. We compare the impact of reasoning traces generated by DeepSeek-R1 and gpt-oss LLMs in terms of accuracy and inference efficiency.
Authors: Rulin Shao, Akari Asai, Shannon Zejiang Shen, Hamish Ivison, Varsha Kishore, Jingming Zhuo, Xinran Zhao, Molly Park, Samuel G. Finlayson, David Sontag, Tyler Murray, Sewon Min, Pradeep Dasigi, Luca Soldaini, Faeze Brahman, Wen-tau Yih, Tongshuang Wu, Luke Zettlemoyer, Yoon Kim, Hannaneh Hajishirzi, Pang Wei Koh
Abstract: Deep research models perform multi-step research to produce long-form, well-attributed answers. However, most open deep research models are trained on easily verifiable short-form QA tasks via reinforcement learning with verifiable rewards (RLVR), which does not extend to realistic long-form tasks. We address this with Reinforcement Learning with Evolving Rubrics (RLER), in which we construct and maintain rubrics that co-evolve with the policy model during training; this allows the rubrics to incorporate information that the model has newly explored and to provide discriminative, on-policy feedback. Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first open model that is directly trained for open-ended, long-form deep research. Across four long-form deep research benchmarks in science, healthcare and general domains, DR Tulu substantially outperforms existing open deep research models, and matches or exceeds proprietary deep research systems, while being significantly smaller and cheaper per query. To facilitate future research, we release all data, models, and code, including our new MCP-based agent infrastructure for deep research systems.
Authors: James Y. Huang, Sheng Zhang, Qianchu Liu, Guanghui Qin, Tinghui Zhu, Tristan Naumann, Muhao Chen, Hoifung Poon
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in challenging, knowledge-intensive reasoning tasks. However, extending LLMs to perceive and reason over a new modality (e.g., vision), often requires costly development of large-scale vision language models (VLMs) with LLMs as backbones. Smaller VLMs are more efficient and adaptable but often lack the broad knowledge and reasoning capabilities of frontier LLMs. In this work, we propose BeMyEyes, a modular, multi-agent framework for extending LLMs to multimodal reasoning by orchestrating collaboration between efficient, adaptable VLMs as perceivers and powerful LLMs as reasoners through conversations. We then introduce a data synthesis and supervised fine-tuning pipeline to train the perceiver agent to effectively collaborate with the reasoner agent. By combining the complementary strengths of perception and reasoning agents, BeMyEyes avoids the need for training large-scale multimodal models, preserves the generalization and reasoning capabilities of LLMs, and allows flexible extension to new domains and modalities. Experiments show that our framework unlocks the multimodal reasoning capabilities for LLMs, enabling a lightweight and fully open-source solution, i.e. equipping text-only DeepSeek-R1 with Qwen2.5-VL-7B perceiver, to outperform large-scale proprietary VLMs such as GPT-4o on a wide range of knowledge-intensive multimodal tasks. These results demonstrate the effectiveness, modularity, and scalability of our multi-agent approach for building future multimodal reasoning systems.
Authors: Jason M. Pittman, Anton Phillips Jr., Yesenia Medina-Santos, Brielle C. Stark
Abstract: Analyzing spoken discourse is a valid means of quantifying language ability in persons with aphasia. There are many ways to quantify discourse, one common way being to evaluate the informativeness of the discourse. That is, given the total number of words produced, how many of those are context-relevant and accurate. This type of analysis is called Correct Information Unit (CIU) analysis and is one of the most prevalent discourse analyses used by speech-language pathologists (SLPs). Despite this, CIU analysis in the clinic remains limited due to the manual labor needed by SLPs to code and analyze collected speech. Recent advances in machine learning (ML) seek to augment such labor by automating modeling of propositional, macrostructural, pragmatic, and multimodal dimensions of discourse. To that end, this study evaluated five ML models for reliable identification of Correct Information Units (CIUs, Nicholas & Brookshire, 1993), during a picture description task. The five supervised ML models were trained using randomly selected human-coded transcripts and accompanying words and CIUs from persons with aphasia. The baseline model training produced a high accuracy across transcripts for word vs non-word, with all models achieving near perfect performance (0.995) with high AUC range (0.914 min, 0.995 max). In contrast, CIU vs non-CIU showed a greater variability, with the k-nearest neighbor (k-NN) model the highest accuracy (0.824) and second highest AUC (0.787). These findings indicate that while the supervised ML models can distinguish word from not word, identifying CIUs is challenging.
Authors: Sumon Kanti Dey, Manvi S, Zeel Mehta, Meet Shah, Unnati Agrawal, Suhani Jalota, Azra Ismail
Abstract: Large Language Models (LLMs) have been positioned as having the potential to expand access to health information in the Global South, yet their evaluation remains heavily dependent on benchmarks designed around Western norms. We present insights from a preliminary benchmarking exercise with a chatbot for sexual and reproductive health (SRH) for an underserved community in India. We evaluated using HealthBench, a benchmark for conversational health models by OpenAI. We extracted 637 SRH queries from the dataset and evaluated on the 330 single-turn conversations. Responses were evaluated using HealthBench's rubric-based automated grader, which rated responses consistently low. However, qualitative analysis by trained annotators and public health experts revealed that many responses were actually culturally appropriate and medically accurate. We highlight recurring issues, particularly a Western bias, such as for legal framing and norms (e.g., breastfeeding in public), diet assumptions (e.g., fish safe to eat during pregnancy), and costs (e.g., insurance models). Our findings demonstrate the limitations of current benchmarks in capturing the effectiveness of systems built for different cultural and healthcare contexts. We argue for the development of culturally adaptive evaluation frameworks that meet quality standards while recognizing needs of diverse populations.
Authors: Guansu Wang, Peijie Sun
Abstract: Recent advances in text-to-speech (TTS) have enabled models to clone arbitrary unseen speakers and synthesize high-quality, natural-sounding speech. However, evaluation methods lag behind: typical mean opinion score (MOS) estimators perform regression over entire utterances, while failures usually occur in a few problematic words. We observe that encoder-decoder ASR models (e.g., Whisper) surface word-level mismatches between speech and text via cross-attention, providing a fine-grained reward signal. Building on this, we introduce Word-level TTS Alignment by ASR-driven Attentive Reward (W3AR). Without explicit reward annotations, W3AR uses attention from a pre-trained ASR model to drive finer-grained alignment and optimization of sequences predicted by a TTS model. Experiments show that W3AR improves the quality of existing TTS systems and strengthens zero-shot robustness on unseen speakers. More broadly, our results suggest a simple recipe for generative modeling: understanding models can act as evaluators, delivering informative, fine-grained feedback for optimization.
Authors: Fengming Yu, Qingyu Meng, Haiwei Pan, Kejia Zhang
Abstract: With the rapid development of deep learning, large language models have shown strong capabilities in complex reasoning tasks such as mathematical equation solving. However, their substantial computational and storage costs hinder practical deployment. This paper proposes a lightweight optimization method that integrates dynamic attention head pruning with knowledge distillation. The approach dynamically evaluates the importance of each attention head in the multi-head attention mechanism using a combination of weight norms and entropy, and prunes redundant heads in real time to reduce computational overhead. To mitigate performance degradation, knowledge distillation transfers information from the original model to the pruned student, enabling the smaller model to preserve reasoning ability. Experiments conducted on both Math23k and ASDiv-A verify the effectiveness of the proposed method. For example, on Math23k with a 30% pruning ratio, parameters are reduced by 18.7%, inference speed is improved by 27.5%, FLOPs are reduced by 19.3%, and accuracy drops only 0.7% (from 84.4% to 83.7%). These results demonstrate that the method achieves substantial efficiency gains while maintaining strong reasoning performance, providing a practical solution for efficient deployment of large language models in mathematical reasoning tasks.
Authors: S\"oren Dr\'eano, Derek Molloy, Noel Murphy
Abstract: This work introduces Llamazip, a novel lossless text compression algorithm based on the predictive capabilities of the LLaMA3 language model. Llamazip achieves significant data reduction by only storing tokens that the model fails to predict, optimizing storage efficiency without compromising data integrity. Key factors affecting its performance, including quantization and context window size, are analyzed, revealing their impact on compression ratios and computational requirements. Beyond compression, Llamazip demonstrates the potential to identify whether a document was part of the training dataset of a language model. This capability addresses critical concerns about data provenance, intellectual property, and transparency in language model training.
Authors: Hyakka Nakada, Yoshiyasu Tanaka
Abstract: Optical Character Recognition (OCR) for data extraction from documents is essential to intelligent informatics, such as digitizing medical records and recognizing road signs. Multi-modal Large Language Models (LLMs) can solve this task and have shown remarkable performance. Recently, it has been noticed that the accuracy of data extraction by multi-modal LLMs can be affected when in-plane rotations are present in the documents. However, real-world document images are usually not only in-plane rotated but also perspectively distorted. This study investigates the impacts of such perturbations on the data extraction accuracy for the state-of-the-art model, Gemini-1.5-pro. Because perspective distortions have a high degree of freedom, designing experiments in the same manner as single-parametric rotations is difficult. We observed typical distortions of document images and showed that most of them approximately follow an isosceles-trapezoidal transformation, which allows us to evaluate distortions with a small number of parameters. We were able to reduce the number of independent parameters from eight to two, i.e. rotation angle and distortion ratio. Then, specific entities were extracted from synthetically generated sample documents with varying these parameters. As the performance of LLMs, we evaluated not only a character-recognition accuracy but also a structure-recognition accuracy. Whereas the former represents the classical indicators for optical character recognition, the latter is related to the correctness of reading order. In particular, the structure-recognition accuracy was found to be significantly degraded by document distortion. In addition, we found that this accuracy can be improved by a simple rotational correction. This insight will contribute to the practical use of multi-modal LLMs for OCR tasks.
Authors: Brendan Gho, Suman Muppavarapu, Afnan Shaik, Tyson Tsay, James Begin, Kevin Zhu, Archana Vaidheeswaran, Vasu Sharma
Abstract: As foundation models are increasingly deployed as interacting agents in multi-agent systems, their collective behavior raises new challenges for trustworthiness, transparency, and accountability. Traditional coordination mechanisms, such as centralized oversight or adversarial adjudication, struggle to scale and often obscure how decisions emerge. We introduce a market-making framework for multi-agent large language model (LLM) coordination that organizes agent interactions as structured economic exchanges. In this setup, each agent acts as a market participant, updating and trading probabilistic beliefs, to converge toward shared, truthful outcomes. By aligning local incentives with collective epistemic goals, the framework promotes self-organizing, verifiable reasoning without requiring external enforcement. Empirically, we evaluate this approach across factual reasoning, ethical judgment, and commonsense inference tasks. Market-based coordination yields accuracy gains of up to 10% over single-shot baselines while preserving interpretability and transparency of intermediate reasoning steps. Beyond these improvements, our findings demonstrate that economic coordination principles can operationalize accountability and robustness in multi-agent LLM systems, offering a scalable pathway toward self-correcting, socially responsible AI capable of maintaining trust and oversight in real world deployment scenarios.
Authors: Ye Tian, Chengcheng Wang, Jing Han, Yehui Tang, Kai Han
Abstract: As Large Language Models (LLMs) continue to grow in size, storing and transmitting them on edge devices becomes increasingly challenging. Traditional methods like quantization and pruning struggle to achieve extreme compression of LLMs without sacrificing accuracy. In this paper, we introduce PocketLLM, a novel approach to compress LLMs in a latent space via meta-networks. A simple encoder network is proposed to project the weights of LLMs into discrete latent vectors, which are then represented using a compact codebook. A lightweight decoder network is employed to map the codebook's representative vectors back to the original weight space. This method allows for significant compression of the large weights in LLMs, consisting solely of a small decoder, a concise codebook, and an index. Extensive experiments show that PocketLLM achieves superior performance even at significantly high compression ratios, e.g., compressing Llama 2-7B by 10x with a negligible drop in accuracy.
Authors: Atharv Singh Patlan, Peiyao Sheng, S. Ashwin Hebbar, Prateek Mittal, Pramod Viswanath
Abstract: Language agents are rapidly expanding from single-user assistants to multi-user collaborators in shared workspaces and groups. However, today's language models lack a mechanism for isolating user interactions and concurrent tasks, creating a new attack vector inherent to this new setting: cross-user poisoning (CUP). In a CUP attack, an adversary injects ordinary-looking messages that poison the persistent, shared state, which later triggers the agent to execute unintended, attacker-specified actions on behalf of benign users. We validate CUP on real systems, successfully attacking popular multi-user agents. To study the phenomenon systematically, we present MURMUR, a framework that composes single-user tasks into concurrent, group-based scenarios using an LLM to generate realistic, history-aware user interactions. We observe that CUP attacks succeed at high rates and their effects persist across multiple tasks, thus posing fundamental risks to multi-user LLM deployments. Finally, we introduce a first-step defense with task-based clustering to mitigate this new class of vulnerability
Authors: Myung Ho Kim
Abstract: Large language model agents suffer from fundamental architectural problems: entangled reasoning and execution, memory volatility, and uncontrolled action sequences. We introduce Structured Cognitive Loop (SCL), a modular architecture that explicitly separates agent cognition into five phases: Retrieval, Cognition, Control, Action, and Memory (R-CCAM). At the core of SCL is Soft Symbolic Control, an adaptive governance mechanism that applies symbolic constraints to probabilistic inference, preserving neural flexibility while restoring the explainability and controllability of classical symbolic systems. Through empirical validation on multi-step conditional reasoning tasks, we demonstrate that SCL achieves zero policy violations, eliminates redundant tool calls, and maintains complete decision traceability. These results address critical gaps in existing frameworks such as ReAct, AutoGPT, and memory-augmented approaches. Our contributions are threefold: (1) we situate SCL within the taxonomy of hybrid intelligence, differentiating it from prompt-centric and memory-only approaches; (2) we formally define Soft Symbolic Control and contrast it with neuro-symbolic AI; and (3) we derive three design principles for trustworthy agents: modular decomposition, adaptive symbolic governance, and transparent state management. We provide a complete open-source implementation demonstrating the R-CCAM loop architecture, alongside a live GPT-4o-powered travel planning agent. By connecting expert system principles with modern LLM capabilities, this work offers a practical and theoretically grounded path toward reliable, explainable, and governable AI agents. Code: https://github.com/enkiluv/scl-core-experiment Demo: https://scl-travel-planner.streamlit.app/
URLs: https://github.com/enkiluv/scl-core-experiment, https://scl-travel-planner.streamlit.app/
Authors: Xi Wang, Xianyao Ling, Kun Li, Gang Yin, Liang Zhang, Jiang Wu, Annie Wang, Weizhe Wang
Abstract: The rapid progress in Generative AI and Agent technologies is profoundly transforming enterprise data management and analytics. Traditional database applications and system deployment are fundamentally impacted by AI-driven tools, such as Retrieval-Augmented Generation (RAG) and vector database technologies, which provide new pathways for semantic querying over enterprise knowledge bases. In the meantime, data security and compliance are top priorities for organizations adopting AI technologies. For enterprise data analysis, SQL generations powered by large language models (LLMs) and AI agents, has emerged as a key bridge connecting natural language with structured data, effectively lowering the barrier to enterprise data access and improving analytical efficiency. This paper focuses on enterprise data analysis applications and system deployment, covering a range of innovative frameworks, enabling complex query understanding, multi-agent collaboration, security verification, and computational efficiency. Through representative use cases, key challenges related to distributed deployment, data security, and inherent difficulties in SQL generation tasks are discussed.
Authors: Tim Schlippe, Matthias W\"olfel, Koena Ronny Mabokela
Abstract: This study investigates how cultural proximity affects the ability to detect AI-generated fake news by comparing South African participants with those from other nationalities. As large language models increasingly enable the creation of sophisticated fake news, understanding human detection capabilities becomes crucial, particularly across different cultural contexts. We conducted a survey where 89 participants (56 South Africans, 33 from other nationalities) evaluated 10 true South African news articles and 10 AI-generated fake versions. Results reveal an asymmetric pattern: South Africans demonstrated superior performance in detecting true news about their country (40% deviation from ideal rating) compared to other participants (52%), but performed worse at identifying fake news (62% vs. 55%). This difference may reflect South Africans' higher overall trust in news sources. Our analysis further shows that South Africans relied more on content knowledge and contextual understanding when judging credibility, while participants from other countries emphasised formal linguistic features such as grammar and structure. Overall, the deviation from ideal rating was similar between groups (51% vs. 53%), suggesting that cultural familiarity appears to aid verification of authentic information but may also introduce bias when evaluating fabricated content. These insights contribute to understanding cross-cultural dimensions of misinformation detection and inform strategies for combating AI-generated fake news in increasingly globalised information ecosystems where content crosses cultural and geographical boundaries.
Authors: Gin\'es Carreto Pic\'on, Peng Yuan Zhou, Qi Zhang, Alexandros Iosifidis
Abstract: Transformer-based models have dramatically increased their size and parameter count to tackle increasingly complex tasks. At the same time, there is a growing demand for low-latency inference on resource-constrained devices that achieves high performance. In particular, stream data inference is typically performed over a sliding temporal window, leading to highly redundant computations. The recent Continual Transformers have addressed this issue, but they can only be effectively used in shallow models, which limits their scope and generalization power. In this paper, we propose the Deep Continual Transformer (DeepCoT), a redundancy-free encoder-only model that can be applied over existing deep encoder architectures with minimal changes. In our experiments over audio, video, and text streams, we show that DeepCoTs retain comparative performance to their non-continual baselines while offering a linear computational cost for all Transformer layers, which reduces up to two orders of magnitude in the running time compared to previous efficient models.
Authors: Ziyang Zhang, Xinheng Ding, Jiayi Yuan, Rixin Liu, Huizi Mao, Jiarong Xing, Zirui Liu
Abstract: Deterministic inference is increasingly critical for large language model (LLM) applications such as LLM-as-a-judge evaluation, multi-agent systems, and Reinforcement Learning (RL). However, existing LLM serving frameworks exhibit non-deterministic behavior: identical inputs can yield different outputs when system configurations (e.g., tensor parallel (TP) size, batch size) vary, even under greedy decoding. This arises from the non-associativity of floating-point arithmetic and inconsistent reduction orders across GPUs. While prior work has addressed batch-size-related nondeterminism through batch-invariant kernels, determinism across different TP sizes remains an open problem, particularly in RL settings, where the training engine typically uses Fully Sharded Data Parallel (i.e., TP = 1) while the rollout engine relies on multi-GPU TP to maximize the inference throughput, creating a natural mismatch between the two. This precision mismatch problem may lead to suboptimal performance or even collapse for RL training. We identify and analyze the root causes of TP-induced inconsistency and propose Tree-Based Invariant Kernels (TBIK), a set of TP-invariant matrix multiplication and reduction primitives that guarantee bit-wise identical results regardless of TP size. Our key insight is to align intra- and inter-GPU reduction orders through a unified hierarchical binary tree structure. We implement these kernels in Triton and integrate them into vLLM and FSDP. Experiments confirm zero probability divergence and bit-wise reproducibility for deterministic inference across different TP sizes. Also, we achieve bit-wise identical results between vLLM and FSDP in RL training pipelines with different parallel strategy. Code is available at https://github.com/nanomaoli/llm_reproducibility.
Authors: Pume Tuchinda, Parinthapat Pengpun, Romrawin Chumpu, Sarana Nutanong, Peerat Limkonchotiwat
Abstract: Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder efficient deployment. Knowledge distillation (KD) has emerged as a powerful approach for building lightweight but competitive models, with strong evidence from both language and vision domains. However, its application to VLMs, particularly CLIP-style models, remains limited, often constrained to small-scale teachers and narrow evaluation tasks such as classification or retrieval. In this work, we present the first systematic study of distillation across a range of CLIP-style teacher models, ranging from standard baselines to large-scale state-of-the-art models. Contrary to trends observed in NLP and vision, we find that stronger teachers do not consistently yield better students; in fact, existing distillation frameworks often fail to scale, leading to degraded performance in downstream multimodal tasks such as visual question answering. Our findings challenge prevailing assumptions in KD and point toward new directions for designing parameter-efficient multimodal models.
Authors: Ziyi Guo, Zhou Liu, Wentao Zhang
Abstract: The manual creation of system architecture diagrams for scientific papers is a time-consuming and subjective process, while existing generative models lack the necessary structural control and semantic understanding for this task. A primary obstacle hindering research and development in this domain has been the profound lack of a standardized benchmark to quantitatively evaluate the automated generation of diagrams from text. To address this critical gap, we introduce a novel and comprehensive benchmark, the first of its kind, designed to catalyze progress in automated scientific visualization. It consists of 3,000 research papers paired with their corresponding high-quality ground-truth diagrams and is accompanied by a three-tiered evaluation metric assessing semantic accuracy, layout coherence, and visual quality. Furthermore, to establish a strong baseline on this new benchmark, we propose Paper2SysArch, an end-to-end system that leverages multi-agent collaboration to convert papers into structured, editable diagrams. To validate its performance on complex cases, the system was evaluated on a manually curated and more challenging subset of these papers, where it achieves a composite score of 69.0. This work's principal contribution is the establishment of a large-scale, foundational benchmark to enable reproducible research and fair comparison. Meanwhile, our proposed system serves as a viable proof-of-concept, demonstrating a promising path forward for this complex task.
Authors: Bowen Qu, Shangkun Sun, Xiaoyu Liang, Wei Gao
Abstract: Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not well aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding edited results from different editing methods, and nearly 4,000 samples with corresponding Mean Opinion Scores (MOS) provided by 15 human subjects. Furthermore, we introduce IE-Critic-R1, which, benefiting from Reinforcement Learning from Verifiable Rewards (RLVR), provides more comprehensive and explainable quality assessment for text-driven image editing that aligns with human perception. Extensive experiments demonstrate IE-Critic-R1's superior subjective-alignments on the text-driven image editing task compared with previous metrics. Related data and codes are available to the public.
Authors: Adrien Banse, Alessandro Abate, Rapha\"el M. Jungers
Abstract: Labeled Markov Chains (or LMCs for short) are useful mathematical objects to model complex probabilistic languages. A central challenge is to compare two LMCs, for example to assess the accuracy of an abstraction or to quantify the effect of model perturbations. In this work, we study the recently introduced Cantor-Kantorovich (or CK) distance. In particular we show that the latter can be framed as a discounted sum of finite-horizon Total Variation distances, making it an instance of discounted linear distance, but arising from the natural Cantor topology. Building on the latter observation, we analyze the properties of the CK distance along three dimensions: computational complexity, continuity properties and approximation. More precisely, we show that the exact computation of the CK distance is #P-hard. We also provide an upper bound on the CK distance as a function of the approximation relation between the two LMCs, and show that a bounded CK distance implies a bounded error between probabilities of finite-horizon traces. Finally, we provide a computable approximation scheme, and show that the latter is also #P-hard. Altogether, our results provide a rigorous theoretical foundation for the CK distance and clarify its relationship with existing distances.
Authors: Dachuan Zhao, Weiyue Li, Zhenda Shen, Yushu Qiu, Bowen Xu, Haoyu Chen, Yongchao Chen
Abstract: Vision-Language Models (VLMs) have become indispensable for multimodal reasoning, yet their representations often encode and amplify demographic biases, resulting in biased associations and misaligned predictions in downstream tasks. Such behavior undermines fairness and distorts the intended alignment between vision and language. Recent post-hoc approaches attempt to mitigate bias by replacing the most attribute-correlated embedding coordinates with neutral values. However, our systematic analysis reveals three critical failures of this coordinate-wise approach: feature entanglement, poor cross-dataset generalization, and incomplete bias removal. We find that bias is not localized to a few coordinates but is instead distributed across a few linear subspaces. To address these limitations, we propose $\textbf{S}$ubspace $\textbf{P}$rojection $\textbf{D}$ebiasing ($\textbf{SPD}$), a geometrically principled framework that identifies and removes the entire subspace of linearly decodable bias while reinserting a neutral mean component to preserve semantic fidelity. Extensive experiments across zero-shot classification, text-to-image retrieval, and image generation validate the effectiveness of SPD: our method achieves more robust debiasing with an average improvement of $18.5\%$ across four fairness metrics, while maintaining minimal loss in task performance compared to the best debiasing baseline.
Authors: Xiaoqing Wang, Keman Huang, Bin Liang, Hongyu Li, Xiaoyong Du
Abstract: The rapid advancement of Large Language Model (LLM)-driven multi-agent systems has significantly streamlined software developing tasks, enabling users with little technical expertise to develop executable applications. While these systems democratize software creation through natural language requirements, they introduce significant security risks that remain largely unexplored. We identify two risky scenarios: Malicious User with Benign Agents (MU-BA) and Benign User with Malicious Agents (BU-MA). We introduce the Implicit Malicious Behavior Injection Attack (IMBIA), demonstrating how multi-agent systems can be manipulated to generate software with concealed malicious capabilities beneath seemingly benign applications, and propose Adv-IMBIA as a defense mechanism. Evaluations across ChatDev, MetaGPT, and AgentVerse frameworks reveal varying vulnerability patterns, with IMBIA achieving attack success rates of 93%, 45%, and 71% in MU-BA scenarios, and 71%, 84%, and 45% in BU-MA scenarios. Our defense mechanism reduced attack success rates significantly, particularly in the MU-BA scenario. Further analysis reveals that compromised agents in the coding and testing phases pose significantly greater security risks, while also identifying critical agents that require protection against malicious user exploitation. Our findings highlight the urgent need for robust security measures in multi-agent software development systems and provide practical guidelines for implementing targeted, resource-efficient defensive strategies.
Authors: Chunyu Qiang, Kang Yin, Xiaopeng Wang, Yuzhe Liang, Jiahui Zhao, Ruibo Fu, Tianrui Wang, Cheng Gong, Chen Zhang, Longbiao Wang, Jianwu Dang
Abstract: Text-to-speech (TTS) and text-to-music (TTM) models face significant limitations in instruction-based control. TTS systems usually depend on reference audio for timbre, offer only limited text-level attribute control, and rarely support dialogue generation. TTM systems are constrained by input conditioning requirements that depend on expert knowledge annotations. The high heterogeneity of these input control conditions makes them difficult to joint modeling with speech synthesis. Despite sharing common acoustic modeling characteristics, these two tasks have long been developed independently, leaving open the challenge of achieving unified modeling through natural language instructions. We introduce InstructAudio, a unified framework that enables instruction-based (natural language descriptions) control of acoustic attributes including timbre (gender, age), paralinguistic (emotion, style, accent), and musical (genre, instrument, rhythm, atmosphere). It supports expressive speech, music, and dialogue generation in English and Chinese. The model employs joint and single diffusion transformer layers with a standardized instruction-phoneme input format, trained on 50K hours of speech and 20K hours of music data, enabling multi-task learning and cross-modal alignment. Fig. 1 visualizes performance comparisons with mainstream TTS and TTM models, demonstrating that InstructAudio achieves optimal results on most metrics. To our best knowledge, InstructAudio represents the first instruction-controlled framework unifying speech and music generation. Audio samples are available at: https://qiangchunyu.github.io/InstructAudio/
Authors: Jian Yang, Wei Zhang, Shark Liu, Jiajun Wu, Shawn Guo, Yizhi Li
Abstract: Large language models (LLMs) have fundamentally transformed automated software development by enabling direct translation of natural language descriptions into functional code, driving commercial adoption through tools like Github Copilot (Microsoft), Cursor (Anysphere), Trae (ByteDance), and Claude Code (Anthropic). While the field has evolved dramatically from rule-based systems to Transformer-based architectures, achieving performance improvements from single-digit to over 95\% success rates on benchmarks like HumanEval. In this work, we provide a comprehensive synthesis and practical guide (a series of analytic and probing experiments) about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents. We analyze the code capability of the general LLMs (GPT-4, Claude, LLaMA) and code-specialized LLMs (StarCoder, Code LLaMA, DeepSeek-Coder, and QwenCoder), critically examining the techniques, design decisions, and trade-offs. Further, we articulate the research-practice gap between academic research (e.g., benchmarks and tasks) and real-world deployment (e.g., software-related code tasks), including code correctness, security, contextual awareness of large codebases, and integration with development workflows, and map promising research directions to practical needs. Last, we conduct a series of experiments to provide a comprehensive analysis of code pre-training, supervised fine-tuning, and reinforcement learning, covering scaling law, framework selection, hyperparameter sensitivity, model architectures, and dataset comparisons.
Authors: Amin Rakhsha, Kanika Madan, Tianyu Zhang, Amir-massoud Farahmand, Amir Khasahmadi
Abstract: Sampling multiple outputs from a Large Language Model (LLM) and selecting the most frequent (Self-consistency) or highest-scoring (Best-of-N) candidate is a popular approach to achieve higher accuracy in tasks with discrete final answers. Best-of-N (BoN) selects the output with the highest reward, and with perfect rewards, it often achieves near-perfect accuracy. With imperfect rewards from reward models, however, BoN fails to reliably find the correct answer and its performance degrades drastically. We consider the distribution of BoN's outputs and highlight that, although the correct answer does not usually have a probability close to one under imperfect rewards, it is often the most likely outcome. This suggests that the mode of this distribution can be more reliably correct than a sample from it. Based on this idea, we propose Majority-of-the-Bests (MoB), a novel selection mechanism that estimates the output distribution of BoN via bootstrapping and selects its mode. Experimental results across five benchmarks, three different base LLMs, and two reward models demonstrate consistent improvements over BoN in 25 out of 30 setups. We also provide theoretical results for the consistency of the bootstrapping. MoB serves as a simple, yet strong alternative to BoN and self-consistency, and more broadly, motivates further research in more nuanced selection mechanisms.
Authors: Alvin Wei Ming Tan, Jane Yang, Tarun Sepuri, Khai Loong Aw, Robert Z. Sparks, Zi Yin, Virginia A. Marchman, Michael C. Frank, Bria Long
Abstract: Figuring out which objects or concepts words refer to is a central language learning challenge for young children. Most models of this process posit that children learn early object labels from co-occurrences of words and their referents that occur when someone around them talks about an object in the immediate physical environment. But how aligned in time are children's visual and linguistic experiences during everyday learning? To date, answers to this question have been limited by the need for labor-intensive manual annotations of vision-language co-occurrences. Here, we evaluate the use of contrastive language-image pretraining (CLIP) models to automatically characterize vision-language alignment in egocentric videos taken from the infant perspective in home environments. After validating CLIP alignment scores using human alignment judgments, we apply this metric to a large corpus of infant-perspective videos. We show that idealized aligned moments for learning (e.g., "look at the ball" with a ball present in the child's view) are relatively rare in children's everyday experiences compared to modern machine learning datasets, and highlight variability in alignment both within and across children. These findings suggest that infrequent alignment is a constraint for models describing early word learning and offer a new method for investigating children's multimodal environment.
Authors: Kairong Luo, Zhenbo Sun, Haodong Wen, Xinyu Shi, Jiarui Cui, Chenyi Dang, Kaifeng Lyu, Wenguang Chen
Abstract: Due to the scarcity of high-quality data, large language models (LLMs) are often trained on mixtures of data with varying quality levels, even after sophisticated data curation. A natural approach to better leverage high-quality data is curriculum-based pretraining, where the model is trained on data sorted in ascending order of quality as determined by a quality metric. However, prior studies have reported limited improvements from such curriculum-based pretraining strategies. This work identifies a critical factor constraining these methods: the incompatibility between the ascending data quality order and the decaying learning rate (LR) schedule. We find that while curriculum-based training substantially outperforms random shuffling when using a constant LR, its advantage diminishes under standard LR decay schedules. Our experiments show this incompatibility can be mitigated by two simple strategies: (1) employing a more moderate LR decay schedule, where the final LR is only moderately smaller than the peak LR, and (2) replacing LR decay with model averaging, i.e., computing a weighted average of the final few checkpoints. By combining these strategies, we improve the average score on a suite of standard benchmarks by 1.64% over random shuffling, without additional data refinement. Validated on 1.5B-parameter models trained over 30B tokens with various data-quality metrics, our findings call for a re-evaluation of curriculum-based LLM pretraining and underscore the potential of co-designing data curricula with optimization methods.
Authors: Santhosh G S, Saurav Prakash, Balaraman Ravindran
Abstract: Large Language Models (LLMs) face a significant bottleneck during autoregressive inference due to the massive memory footprint of the Key-Value (KV) cache. Existing compression techniques like token eviction, quantization, or other low-rank methods often risk information loss, have fixed limits, or introduce significant computational overhead from explicit decompression steps. In this work, we introduce SWAN, a novel, fine-tuning-free framework that eliminates this overhead. Our method uses an offline orthogonal matrix to rotate and prune the KV-cache, which is then used directly in the attention computation without any reconstruction. Our extensive experiments demonstrate that SWAN, augmented with a small dense buffer, offers a robust trade-off, maintaining performance close to the uncompressed baseline even at aggressive 50-60% memory savings per-token on KV-cache. A key advantage is its runtime-tunable compression level, allowing operators to dynamically adjust the memory footprint, a flexibility absent in methods requiring fixed offline configurations. This combination of a decompression-free design, high performance under compression, and adaptability makes SWAN a practical and efficient solution for serving LLMs with long contexts.
Authors: Zineddine Tighidet, Lazhar Labiod, Mohamed Nadif
Abstract: The mixture model is undoubtedly one of the greatest contributions to clustering. For continuous data, Gaussian models are often used and the Expectation-Maximization (EM) algorithm is particularly suitable for estimating parameters from which clustering is inferred. If these models are particularly popular in various domains including image clustering, they however suffer from the dimensionality and also from the slowness of convergence of the EM algorithm. However, the Classification EM (CEM) algorithm, a classifying version, offers a fast convergence solution while dimensionality reduction still remains a challenge. Thus we propose in this paper an algorithm combining simultaneously and non-sequentially the two tasks --Data embedding and Clustering-- relying on Principal Component Analysis (PCA) and CEM. We demonstrate the interest of such approach in terms of clustering and data embedding. We also establish different connections with other clustering approaches.
Authors: Junbo Zhang, Ran Chen, Qianli Zhou, Xinyang Deng, Wen Jiang
Abstract: Large language models demonstrate powerful capabilities across various natural language processing tasks, yet they also harbor safety vulnerabilities. To enhance LLM safety, various jailbreak defense methods have been proposed to guard against harmful outputs. However, improvements in model safety often come at the cost of severe over-refusal, failing to strike a good balance between safety and usability. In this paper, we first analyze the causes of over-refusal from a representation perspective, revealing that over-refusal samples reside at the boundary between benign and malicious samples. Based on this, we propose MOSR, designed to mitigate over-refusal by intervening the safety representation of LLMs. MOSR incorporates two novel components: (1) Overlap-Aware Loss Weighting, which determines the erasure weight for malicious samples by quantifying their similarity to pseudo-malicious samples in the representation space, and (2) Context-Aware Augmentation, which supplements the necessary context for rejection decisions by adding harmful prefixes before rejection responses. Experiments demonstrate that our method outperforms existing approaches in mitigating over-refusal while largely maintaining safety. Overall, we advocate that future defense methods should strike a better balance between safety and over-refusal.
Authors: Moazzam Umer Gondal, Hamad Ul Qudous, Daniya Siddiqui, Asma Ahmad Farhan
Abstract: This paper introduces the retrieval-augmented framework for automatic fashion caption and hashtag generation, combining multi-garment detection, attribute reasoning, and Large Language Model (LLM) prompting. The system aims to produce visually grounded, descriptive, and stylistically interesting text for fashion imagery, overcoming the limitations of end-to-end captioners that have problems with attribute fidelity and domain generalization. The pipeline combines a YOLO-based detector for multi-garment localization, k-means clustering for dominant color extraction, and a CLIP-FAISS retrieval module for fabric and gender attribute inference based on a structured product index. These attributes, together with retrieved style examples, create a factual evidence pack that is used to guide an LLM to generate human-like captions and contextually rich hashtags. A fine-tuned BLIP model is used as a supervised baseline model for comparison. Experimental results show that the YOLO detector is able to obtain a mean Average Precision (mAP@0.5) of 0.71 for nine categories of garments. The RAG-LLM pipeline generates expressive attribute-aligned captions and achieves mean attribute coverage of 0.80 with full coverage at the 50% threshold in hashtag generation, whereas BLIP gives higher lexical overlap and lower generalization. The retrieval-augmented approach exhibits better factual grounding, less hallucination, and great potential for scalable deployment in various clothing domains. These results demonstrate the use of retrieval-augmented generation as an effective and interpretable paradigm for automated and visually grounded fashion content generation.
Authors: Deyi Ji, Yuekui Yang, Liqun Liu, Peng Shu, Haiyang Wu, Shaogang Tang, Xudong Chen, Shaoping Ma, Tianrun Chen, Lanyun Zhu
Abstract: Advertising (Ad) is a cornerstone of the digital economy, yet the moderation of video advertisements remains a significant challenge due to their complexity and the need for precise violation localization. While recent advancements, such as the RAVEN model, have improved coarse-grained violation detection, critical gaps persist in fine-grained understanding, explainability, and generalization. To address these limitations, we propose RAVEN++, a novel framework that introduces three key innovations: 1) Active Reinforcement Learning (RL), which dynamically adapts training to samples of varying difficulty; 2) Fine-Grained Violation Understanding, achieved through hierarchical reward functions and reasoning distillation; and 3) Progressive Multi-Stage Training, which systematically combines knowledge injection, curriculum-based passive RL, and active RL. Extensive experiments on both public and proprietary datasets, on both offline scenarios and online deployed A/B Testing, demonstrate that RAVEN++ outperforms general-purpose LLMs and specialized models like RAVEN in terms of fine-grained violation understanding, reasoning capabilities, and generalization ability.
Authors: Kyle Verrier, Achille Nazaret, Joseph Futoma, Andrew C. Miller, Guillermo Sapiro
Abstract: Hunger and satiety dynamics shape dietary behaviors and metabolic health, yet remain difficult to capture in everyday settings. We present a Nutrition Photoplethysmography Language Model (NPLM), integrating continuous photoplethysmography (PPG) from wearables with meal descriptions. NPLM projects PPG into embeddings interpretable by language models, enabling joint reasoning over physiology and meal context. Trained on 19,340 participants and 1.1 million meal-PPG pairs, the model improved daily caloric intake prediction by 11% over text-only baselines, with accuracy maintained when 80% of meal text was removed. In an independent validation study (n=140) with controlled dining and detailed meal information, the model replicated these findings. These results demonstrate the value of integrating physiological measurements from consumer wearables with meal information for noninvasive dietary monitoring at scale.
Authors: Minseo Kim, Chenfeng Xu, Coleman Hooper, Harman Singh, Ben Athiwaratkun, Ce Zhang, Kurt Keutzer, Amir Gholami
Abstract: Diffusion Language Models (DLMs) offer a promising parallel generation paradigm but suffer from slow inference due to numerous refinement steps and the inability to use standard KV caching. We introduce CDLM (Consistency Diffusion Language Models), a training-based acceleration method that simultaneously tackles both bottlenecks. CDLM integrates consistency modeling to drastically reduce the number of required sampling steps by enabling multi-token finalization. Furthermore, we enforce a block-wise causal attention mask during fine-tuning, making the model fully compatible with KV caching. Experiments show CDLM achieves 3.6x-14.5x lower latency while maintaining competitive accuracy on math and coding tasks. The full training and evaluation code is available at https://github.com/SqueezeAILab/CDLM.
Authors: Victor Rambaud, Salvador Mascarenhas, Yair Lakretz
Abstract: A cognitive map is an internal model which encodes the abstract relationships among entities in the world, giving humans and animals the flexibility to adapt to new situations, with a strong out-of-distribution (OOD) generalization that current AI systems still do not possess. To bridge this gap, we introduce MapFormers, new architectures based on Transformer models, which can learn cognitive maps from observational data and perform path integration in parallel, in a self-supervised manner. Cognitive maps are learned in the model by disentangling structural relationships in the inputs from their specific content, a property that can be achieved naturally by updating the positional encoding in Transformers with input-dependent matrices. We developed two variants of MapFormers that unify absolute and relative positional encoding to model episodic (EM) and working memory (WM), respectively. We tested MapFormers on several tasks, including a classic 2D navigation task, showing that our models can learn a cognitive map of the underlying space and generalize OOD (e.g., to longer sequences) with near-perfect performance, unlike current architectures. Together, these results demonstrate the superiority of models designed to learn a cognitive map, and the importance of introducing a structural bias for structure-content disentanglement, which can be achieved in Transformers with input-dependent positional encoding. MapFormers have broad applications in both neuroscience and AI, by explaining the neural mechanisms giving rise to cognitive maps, while allowing these relation models to be learned at scale.
Authors: Jiayi Zhang, Yiran Peng, Fanqi Kong, Yang Cheng, Yifan Wu, Zhaoyang Yu, Jinyu Xiang, Jianhao Ruan, Jinlin Wang, Maojia Song, HongZhang Liu, Xiangru Tang, Bang Liu, Chenglin Wu, Yuyu Luo
Abstract: Humans naturally adapt to diverse environments by learning underlying rules across worlds with different dynamics, observations, and reward structures. In contrast, existing agents typically demonstrate improvements via self-evolving within a single domain, implicitly assuming a fixed environment distribution. Cross-environment learning has remained largely unmeasured: there is no standard collection of controllable, heterogeneous environments, nor a unified way to represent how agents learn. We address these gaps in two steps. First, we propose AutoEnv, an automated framework that treats environments as factorizable distributions over transitions, observations, and rewards, enabling low-cost (4.12 USD on average) generation of heterogeneous worlds. Using AutoEnv, we construct AutoEnv-36, a dataset of 36 environments with 358 validated levels, on which seven language models achieve 12-49% normalized reward, demonstrating the challenge of AutoEnv-36. Second, we formalize agent learning as a component-centric process driven by three stages of Selection, Optimization, and Evaluation applied to an improvable agent component. Using this formulation, we design eight learning methods and evaluate them on AutoEnv-36. Empirically, the gain of any single learning method quickly decrease as the number of environments increases, revealing that fixed learning methods do not scale across heterogeneous environments. Environment-adaptive selection of learning methods substantially improves performance but exhibits diminishing returns as the method space expands. These results highlight both the necessity and the current limitations of agent learning for scalable cross-environment generalization, and position AutoEnv and AutoEnv-36 as a testbed for studying cross-environment agent learning. The code is avaiable at https://github.com/FoundationAgents/AutoEnv.
Authors: Jaewoo Lee, Archiki Prasad, Justin Chih-Yao Chen, Zaid Khan, Elias Stengel-Eskin, Mohit Bansal
Abstract: Information-seeking is a core capability for AI agents, requiring them to gather and reason over tool-generated information across long trajectories. However, such multi-step information-seeking tasks remain challenging for agents backed by language models. While process reward models (PRMs) can guide agents by ranking candidate steps at test-time, existing PRMs, designed for short reasoning with binary judgment, cannot capture richer dimensions of information-seeking steps, such as tool interactions and reasoning over tool outputs, nor handle the rapidly growing context in long-horizon tasks. To address these limitations, we introduce PRInTS, a generative PRM trained with dual capabilities: (1) dense scoring based on the PRM's reasoning across multiple step quality dimensions (e.g., interpretation of tool outputs, tool call informativeness) and (2) trajectory summarization that compresses the growing context while preserving essential information for step evaluation. Extensive evaluations across FRAMES, GAIA (levels 1-3), and WebWalkerQA (easy-hard) benchmarks on multiple models, along with ablations, reveal that best-of-n sampling with PRInTS enhances information-seeking abilities of open-source models as well as specialized agents, matching or surpassing the performance of frontier models with a much smaller backbone agent and outperforming other strong reward modeling baselines.
Authors: Roksana Goworek, Olivia Macmillan-Scott, Eda B. \"Ozyi\u{g}it
Abstract: Cross-lingual information retrieval (CLIR) enables access to multilingual knowledge but remains challenging due to disparities in resources, scripts, and weak cross-lingual semantic alignment in embedding models. Existing pipelines often rely on translation and monolingual retrieval heuristics, which add computational overhead and noise, degrading performance. This work systematically evaluates four intervention types, namely document translation, multilingual dense retrieval with pretrained encoders, contrastive learning at word, phrase, and query-document levels, and cross-encoder re-ranking, across three benchmark datasets. We find that dense retrieval models trained specifically for CLIR consistently outperform lexical matching methods and derive little benefit from document translation. Contrastive learning mitigates language biases and yields substantial improvements for encoders with weak initial alignment, and re-ranking can be effective, but depends on the quality of the cross-encoder training data. Although high-resource languages still dominate overall performance, gains over lexical and document-translated baselines are most pronounced for low-resource and cross-script pairs. These findings indicate that cross-lingual search systems should prioritise semantic multilingual embeddings and targeted learning-based alignment over translation-based pipelines, particularly for cross-script and under-resourced languages.
Authors: Olivia Macmillan-Scott, Roksana Goworek, Eda B. \"Ozyi\u{g}it
Abstract: Query expansion is the reformulation of a user query by adding semantically related information, and is an essential component of monolingual and cross-lingual information retrieval used to ensure that relevant documents are not missed. Recently, multilingual large language models (mLLMs) have shifted query expansion from semantic augmentation with synonyms and related words to pseudo-document generation. Pseudo-documents both introduce additional relevant terms and bridge the gap between short queries and long documents, which is particularly beneficial in dense retrieval. This study evaluates recent mLLMs and fine-tuned variants across several generative expansion strategies to identify factors that drive cross-lingual retrieval performance. Results show that query length largely determines which prompting technique is effective, and that more elaborate prompts often do not yield further gains. Substantial linguistic disparities persist: cross-lingual query expansion can produce the largest improvements for languages with the weakest baselines, yet retrieval is especially poor between languages written in different scripts. Fine-tuning is found to lead to performance gains only when the training and test data are of similar format. These outcomes underline the need for more balanced multilingual and cross-lingual training and evaluation resources.
Authors: Nikita Neveditsin, Pawan Lingras, Vijay Mago
Abstract: Clustering short text embeddings is a foundational task in natural language processing, yet remains challenging due to the need to specify the number of clusters in advance. We introduce a scalable spectral method that estimates the number of clusters directly from the structure of the Laplacian eigenspectrum, constructed using cosine similarities and guided by an adaptive sampling strategy. This sampling approach enables our estimator to efficiently scale to large datasets without sacrificing reliability. To support intrinsic evaluation of cluster quality without ground-truth labels, we propose the Cohesion Ratio, a simple and interpretable evaluation metric that quantifies how much intra-cluster similarity exceeds the global similarity background. It has an information-theoretic motivation inspired by mutual information, and in our experiments it correlates closely with extrinsic measures such as normalized mutual information and homogeneity. Extensive experiments on six short-text datasets and four modern embedding models show that standard algorithms like K-Means and HAC, when guided by our estimator, significantly outperform popular parameter-light methods such as HDBSCAN, OPTICS, and Leiden. These results demonstrate the practical value of our spectral estimator and Cohesion Ratio for unsupervised organization and evaluation of short text data. Implementation of our estimator of k and Cohesion Ratio, along with code for reproducing the experiments, is available at https://anonymous.4open.science/r/towards_clustering-0C2E.
URLs: https://anonymous.4open.science/r/towards_clustering-0C2E.
Authors: Dorian Quelle, Calvin Cheng, Alexandre Bovet, Scott A. Hale
Abstract: Misinformation and disinformation are growing threats in the digital age, affecting people across languages and borders. However, no research has investigated the prevalence of multilingual misinformation and quantified the extent to which misinformation diffuses across languages. This paper investigates the prevalence and dynamics of multilingual misinformation through an analysis of 264,487 fact-checks spanning 95 languages. To study the evolution of claims over time and mutations across languages, we represent fact-checks with multilingual sentence embeddings and build a graph where semantically similar claims are linked. We provide quantitative evidence of repeated fact-checking efforts and establish that claims diffuse across languages. Specifically, we find that while the majority of misinformation claims are only fact-checked once, 10.26%, corresponding to more than 27,000 claims, are checked multiple times. Using fact-checks as a proxy for the spread of misinformation, we find 32.26% of repeated claims cross linguistic boundaries, suggesting that some misinformation permeates language barriers. However, spreading patterns exhibit strong assortativity, with misinformation more likely to spread within the same language or language family. Next we show that fact-checkers take more time to fact-check claims that have crossed language barriers and model the temporal and cross-lingual evolution of claims. We analyze connected components and shortest paths connecting different versions of a claim finding that claims gradually drift over time and undergo greater alteration when traversing languages. Misinformation changes over time, reducing the effectiveness of static claim matching algorithms. The findings advocate for expanded information sharing between fact-checkers globally while underscoring the importance of localized verification.
Authors: Zheng Liu, Chaofan Li, Shitao Xiao, Yingxia Shao, Defu Lian
Abstract: Dense retrieval calls for discriminative embeddings to represent the semantic relationship between query and document. It may benefit from the using of large language models (LLMs), given LLMs' strong capability on semantic understanding. However, the LLMs are learned by auto-regression, whose working mechanism is completely different from representing whole text as one discriminative embedding. Thus, it is imperative to study how to adapt LLMs properly so that they can be effectively initialized as the backbone encoder for dense retrieval. In this paper, we propose a novel approach, called Llama2Vec, which performs unsupervised adaptation of LLM for its dense retrieval application. Llama2Vec consists of two pretext tasks: EBAE (Embedding-Based Auto-Encoding) and EBAR (Embedding-Based Auto-Regression), where the LLM is prompted to reconstruct the input sentence and predict the next sentence based on its text embeddings. Llama2Vec is simple, lightweight, but highly effective. It is used to adapt LLaMA-2-7B on the Wikipedia corpus. With a moderate steps of adaptation, it substantially improves the model's fine-tuned performances on a variety of dense retrieval benchmarks. Notably, it results in the new state-of-the-art performances on popular benchmarks, such as passage and document retrieval on MSMARCO, and zero-shot retrieval on BEIR. The model and source code will be made publicly available to facilitate the future research. Our model is available at https://github.com/FlagOpen/FlagEmbedding.
Authors: Huaqin Zhao, Zhengliang Liu, Zihao Wu, Yiwei Li, Tianze Yang, Peng Shu, Shaochen Xu, Haixing Dai, Lin Zhao, Hanqi Jiang, Yi Pan, Junhao Chen, Yifan Zhou, Zeyu Zhang, Gengchen Mai, Ninghao Liu, Tianming Liu
Abstract: In recent years, Large Language Models (LLMs) like ChatGPT have seen considerable advancements and have been applied in diverse fields. Built on the Transformer architecture, these models are trained on extensive datasets, enabling them to understand and generate human language effectively. In the financial domain, the deployment of LLMs is gaining momentum. These models are being utilized for automating financial report generation, forecasting market trends, analyzing investor sentiment, and offering personalized financial advice. Leveraging their natural language processing capabilities, LLMs can distill key insights from vast financial data, aiding institutions in making informed investment choices and enhancing both operational efficiency and customer satisfaction. In this study, we provide a comprehensive overview of the emerging integration of LLMs into various financial tasks. Additionally, we conducted holistic tests on multiple financial tasks through the combination of natural language instructions. Our findings show that GPT-4 effectively follow prompt instructions across various financial tasks. This survey and evaluation of LLMs in the financial domain aim to deepen the understanding of LLMs' current role in finance for both financial practitioners and LLM researchers, identify new research and application prospects, and highlight how these technologies can be leveraged to solve practical challenges in the finance industry.
Authors: Yupeng Cao, Aishwarya Muralidharan Nair, Nastaran Jamalipour Soofi, Elyon Eyimife, K. P. Subbalakshmi
Abstract: Scientific facts are often spun in the popular press with the intent to influence public opinion and action, as was evidenced during the COVID-19 pandemic. Automatic detection of misinformation in the scientific domain is challenging because of the distinct styles of writing in these two media types and is still in its nascence. Most research on the validity of scientific reporting treats this problem as a claim verification challenge. In doing so, significant expert human effort is required to generate appropriate claims. Our solution bypasses this step and addresses a more real-world scenario where such explicit, labeled claims may not be available. The central research question of this paper is whether it is possible to use large language models (LLMs) to detect misinformation in scientific reporting. To this end, we first present a new labeled dataset SciNews, containing 2.4k scientific news stories drawn from trusted and untrustworthy sources, paired with related abstracts from the CORD-19 database. Our dataset includes both human-written and LLM-generated news articles, making it more comprehensive in terms of capturing the growing trend of using LLMs to generate popular press articles. Then, we identify dimensions of scientific validity in science news articles and explore how this can be integrated into the automated detection of scientific misinformation. We propose several baseline architectures using LLMs to automatically detect false representations of scientific findings in the popular press. For each of these architectures, we use several prompt engineering strategies including zero-shot, few-shot, and chain-of-thought prompting. We also test these architectures and prompting strategies on GPT-3.5, GPT-4, and Llama2-7B, Llama2-13B.
Authors: Yanjun Lyu, Zihao Wu, Lu Zhang, Jing Zhang, Yiwei Li, Wei Ruan, Zhengliang Liu, Zeyu Zhang, Xiang Li, Rongjie Liu, Chao Huang, Wentao Li, Tianming Liu, Dajiang Zhu
Abstract: Pre-trained large language models(LLMs) have attracted increasing attention in biomedical domains due to their success in natural language processing. However, the complex traits and heterogeneity of multi-sources genomics data pose significant challenges when adapting these models to the bioinformatics and biomedical field. To address these challenges, we present GP-GPT, the first specialized large language model for genetic-phenotype knowledge representation and genomics relation analysis. Our model is fine-tuned in two stages on a comprehensive corpus composed of over 3,000,000 terms in genomics, proteomics, and medical genetics, derived from multiple large-scale validated datasets and scientific publications. GP-GPT demonstrates proficiency in accurately retrieving medical genetics information and performing common genomics analysis tasks, such as genomics information retrieval and relationship determination. Comparative experiments across domain-specific tasks reveal that GP-GPT outperforms state-of-the-art LLMs, including Llama2, Llama3 and GPT-4. These results highlight GP-GPT's potential to enhance genetic disease relation research and facilitate accurate and efficient analysis in the fields of genomics and medical genetics. Our investigation demonstrated the subtle changes of bio-factor entities' representations in the GP-GPT, which suggested the opportunities for the application of LLMs to advancing gene-phenotype research.
Authors: Tianyang Zhong, Zhengliang Liu, Yi Pan, Yutong Zhang, Zeyu Zhang, Yifan Zhou, Shizhe Liang, Zihao Wu, Yanjun Lyu, Peng Shu, Xiaowei Yu, Chao Cao, Hanqi Jiang, Hanxu Chen, Yiwei Li, Junhao Chen, Huawen Hu, Yiheng Liu, Huaqin Zhao, Shaochen Xu, Haixing Dai, Lin Zhao, Ruidong Zhang, Wei Zhao, Zhenyuan Yang, Jingyuan Chen, Peilong Wang, Wei Ruan, Hui Wang, Huan Zhao, Jing Zhang, Yiming Ren, Shihuan Qin, Tong Chen, Jiaxi Li, Arif Hassan Zidan, Afrar Jahin, Minheng Chen, Sichen Xia, Jason Holmes, Yan Zhuang, Jiaqi Wang, Bochen Xu, Weiran Xia, Jichao Yu, Kaibo Tang, Yaxuan Yang, Bolun Sun, Tao Yang, Guoyu Lu, Xianqiao Wang, Lilong Chai, He Li, Jin Lu, Xin Zhang, Bao Ge, Xintao Hu, Lian Zhang, Hua Zhou, Lu Zhang, Shu Zhang, Zhen Xiang, Yudan Ren, Jun Liu, Xi Jiang, Yu Bao, Wei Zhang, Xiang Li, Gang Li, Wei Liu, Dinggang Shen, Andrea Sikora, Xiaoming Zhai, Dajiang Zhu, Tuo Zhang, Tianming Liu
Abstract: This comprehensive study evaluates the performance of OpenAI's o1-preview large language model across a diverse array of complex reasoning tasks, spanning multiple domains, including computer science, mathematics, natural sciences, medicine, linguistics, and social sciences. Through rigorous testing, o1-preview demonstrated remarkable capabilities, often achieving human-level or superior performance in areas ranging from coding challenges to scientific reasoning and from language processing to creative problem-solving. Key findings include: -83.3% success rate in solving complex competitive programming problems, surpassing many human experts. -Superior ability in generating coherent and accurate radiology reports, outperforming other evaluated models. -100% accuracy in high school-level mathematical reasoning tasks, providing detailed step-by-step solutions. -Advanced natural language inference capabilities across general and specialized domains like medicine. -Impressive performance in chip design tasks, outperforming specialized models in areas such as EDA script generation and bug analysis. -Remarkable proficiency in anthropology and geology, demonstrating deep understanding and reasoning in these specialized fields. -Strong capabilities in quantitative investing. O1 has comprehensive financial knowledge and statistical modeling skills. -Effective performance in social media analysis, including sentiment analysis and emotion recognition. The model excelled particularly in tasks requiring intricate reasoning and knowledge integration across various fields. While some limitations were observed, including occasional errors on simpler problems and challenges with certain highly specialized concepts, the overall results indicate significant progress towards artificial general intelligence.
Authors: Vibhor Agarwal, Yiqiao Jin, Mohit Chandra, Munmun De Choudhury, Srijan Kumar, Nishanth Sastry
Abstract: Large language models (LLMs) are starting to complement traditional information seeking mechanisms such as web search. LLM-powered chatbots like ChatGPT are gaining prominence among the general public. AI chatbots are also increasingly producing content on social media platforms. However, LLMs are also prone to hallucinations, generating plausible yet factually incorrect or fabricated information. This becomes a critical problem when laypeople start seeking information about sensitive issues such as healthcare. Existing works in LLM hallucinations in the medical domain mainly focus on testing the medical knowledge of LLMs through standardized medical exam questions which are often well-defined and clear-cut with definitive answers. However, these approaches may not fully capture how these LLMs perform during real-world interactions with patients. This work conducts a pioneering study on hallucinations in LLM-generated responses to real-world healthcare queries from patients.We introduce MedHalu, a novel medical hallucination benchmark featuring diverse health-related topics and hallucinated responses from LLMs, with detailed annotation of the hallucination types and text spans. We also propose MedHaluDetect, a comprehensive framework for evaluating LLMs' abilities to detect hallucinations. Furthermore, we study the vulnerability to medical hallucinations among three groups -- medical experts, LLMs, and laypeople. Notably, LLMs significantly underperform human experts and, in some cases, even laypeople in detecting medical hallucinations. To improve hallucination detection, we propose an expert-in-the-loop approach that integrates expert reasoning into LLM inputs, significantly improving hallucination detection for all LLMs, including a 6.3% macro-F1 improvement for GPT-4. Our code and dataset are available at https://netsys.surrey.ac.uk/datasets/medhalu/.
Authors: Yuling Shi, Songsong Wang, Chengcheng Wan, Min Wang, Xiaodong Gu
Abstract: While large language models have made significant strides in code generation, the pass rate of the generated code is bottlenecked on subtle errors, often requiring human intervention to pass tests, especially for complex problems. Existing LLM-based debugging systems treat generated programs as monolithic units, failing to address bugs at multiple levels of granularity, from low-level syntax errors to high-level algorithmic flaws. In this paper, we introduce Multi-Granularity Debugger (MGDebugger), a hierarchical code debugger by isolating, identifying, and resolving bugs at various levels of granularity. MGDebugger decomposes problematic code into a hierarchical tree structure of subfunctions, with each level representing a particular granularity of error. During debugging, it analyzes each subfunction and iteratively resolves bugs in a bottom-up manner. To effectively test each subfunction, we propose an LLM-simulated Python executor, which traces code execution and tracks important variable states to pinpoint errors accurately. Extensive experiments demonstrate that MGDebugger outperforms existing debugging systems, achieving an 18.9% improvement in accuracy over seed generations in HumanEval and a 97.6% repair success rate in HumanEvalFix. Furthermore, MGDebugger effectively fixes bugs across different categories and difficulty levels, demonstrating its robustness and effectiveness.
Authors: Shan Xie, Man Luo, Chadly Daniel Stern, Mengnan Du, Lu Cheng
Abstract: Large language models (LLMs) using in-context learning (ICL) excel in many tasks without task-specific fine-tuning. However, demonstration selection and ordering greatly impact ICL effectiveness. Focus on this issue, we propose DemoShapley, a Shapley-value based method that evaluates each demonstration's contribution by measuring its marginal effect across different prompt permutations. To further account for ICL's limited context windows and frequent low-shot settings, we introduce Beta-DemoShapley, a weighted extension that emphasizes the influence of smaller prompt sizes. Experiments on multiple benchmarks show that DemoShapley consistently outperforms existing influence-based selection strategies, while Beta-DemoShapley further improves performance in low-shot scenarios. Both methods also detect mislabeled data, enhance generalization to out-of-distribution tasks, and reduce demographic bias. Together, they provide a unified and robust framework for demonstration valuation in ICL.
Authors: Isack Lee, Haebin Seong
Abstract: Although large language models (LLMs) demonstrate impressive proficiency in various tasks, they present potential safety risks, such as `jailbreaks', where malicious inputs can coerce LLMs into generating harmful content bypassing safety alignments. In this paper, we delve into the ethical biases in LLMs and examine how those biases could be exploited for jailbreaks. Notably, these biases result in a jailbreaking success rate in GPT-4o models that differs by 20\% between non-binary and cisgender keywords and by 16\% between white and black keywords, even when the other parts of the prompts are identical. We introduce the concept of BiasJailbreak, highlighting the inherent risks posed by these safety-induced biases. BiasJailbreak generates biased keywords automatically by asking the target LLM itself, and utilizes the keywords to generate harmful output. Additionally, we propose an efficient defense method BiasDefense, which prevents jailbreak attempts by injecting defense prompts prior to generation. BiasDefense stands as an appealing alternative to Guard Models, such as Llama-Guard, that require additional inference cost after text generation. Our findings emphasize that ethical biases in LLMs can actually lead to generating unsafe output, and suggest a method to make the LLMs more secure and unbiased. To enable further research and improvements, we open-source our code and artifacts of BiasJailbreak, providing the community with tools to better understand and mitigate safety-induced biases in LLMs.
Authors: Seoyeon Kim, Huiseo Kim, Chanjun Park, Jinyoung Yeo, Dongha Lee
Abstract: Recent large language models (LLMs) demonstrate multilingual abilities, yet they are English-centric due to dominance of English in training corpora. The limited resource for low-resource languages remains a crucial challenge. Code-switching (CS), a phenomenon where multilingual speakers alternate between languages in a discourse, can convey subtle cultural and linguistic nuances that can be otherwise lost in translation and elicits language-specific knowledge in human communications. In light of this, we investigate whether code-switching can activate, or identify and leverage knowledge for reasoning when LLMs solve low-resource language tasks. To facilitate the research, we first present EnKoQA, a synthetic English-Korean CS question-answering dataset. We provide comprehensive analysis on a variety of multilingual LLMs by subdividing activation process into knowledge identification and knowledge leveraging. Our results demonstrate that compared to English text, CS can faithfully activate knowledge inside LLMs especially on language-specific domains, suggesting the potential of code-switching on low-resource language tasks.
Authors: Mikita Balesni, Tomek Korbak, Owain Evans
Abstract: Large language models can use chain-of-thought (CoT) to externalize reasoning, potentially enabling oversight of capable LLM agents. Prior work has shown that models struggle at two-hop question-answering without CoT. This capability is so basic that if it was a fundamental limitation, it would imply that many complex agentic tasks would similarly require CoT. We investigate LLM latent reasoning capabilities using two-hop question answering as a case study. Previous work on the gap between latent and externalized two-hop reasoning produced mixed evidence with inconclusive results. In this paper, we introduce a controlled setting for investigating two-hop reasoning in LLMs, where a positive result provides definitive evidence for latent reasoning. We fine-tune LLMs (including Llama 3 8B and GPT-4o) on synthetic facts and test two-hop reasoning over these facts. By using synthetic facts, we rule out memorization and reasoning shortcuts as explanations for two-hop performance. We observe a nuanced picture: Models fail to compose two synthetic facts, but can succeed when one fact is synthetic and the other is natural. These results demonstrate that LLMs are undeniably capable of latent two-hop reasoning, although it remains unclear how this ability scales with model size. Finally, we highlight a lesson for researchers studying LLM reasoning: when drawing conclusions about LLM latent reasoning, one must be careful to avoid both spurious successes (that stem from memorization and reasoning shortcuts) and spurious failures (that may stem from artificial experimental setups, divorced from training setups of frontier LLMs).
Authors: Lehan He, Zeren Chen, Zhelun Shi, Tianyu Yu, Jing Shao, Lu Sheng
Abstract: The success of Direct Preference Optimization (DPO) in mitigating hallucinations in Vision Language Models (VLMs) critically hinges on the true reward gaps within preference pairs. However, current methods, typically relying on ranking or rewriting strategies, often struggle to optimize these reward gaps in a systematic way during data curation. A core difficulty lies in precisely characterizing and strategically manipulating the overall reward gap configuration, that is, the deliberate design of how to shape these reward gaps within each preference pair across the data. To address this, we introduce Topic-level Preference Rewriting(TPR), a novel framework designed for the systematic optimization of reward gap configuration. Through selectively replacing semantic topics within VLM responses with model's own resampled candidates for targeted rewriting, TPR can provide topic-level control over fine-grained semantic details. This precise control enables advanced data curation strategies, such as progressively adjusting the difficulty of rejected responses, thereby sculpting an effective reward gap configuration that guides the model to overcome challenging hallucinations. Comprehensive experiments demonstrate TPR achieves state-of-the-art performance on multiple hallucination benchmarks, outperforming previous methods by an average of 20%. Notably, it significantly reduces hallucinations by up to 93% on ObjectHal-Bench, and also exhibits superior data efficiency towards robust and cost-effective VLM alignment. Code and datasets are available at https://tpr-dpo.github.io .
Authors: Alfredo Garrach\'on Ruiz, Tom\'as de la Rosa, Daniel Borrajo
Abstract: The high inference cost of Large Language Models (LLMs) poses challenges, especially for tasks requiring lengthy outputs. However, natural language often contains redundancy, which presents an opportunity for optimization. We have observed that LLMs can generate distilled language (i.e., concise outputs that retain essential meaning) when prompted appropriately. We propose TRIM, a pipeline for saving computational cost in which the LLM omits a predefined set of semantically irrelevant and easily inferable words based on the context during inference. Then, a specifically trained smaller language model with lower inference cost reconstructs the distilled answer into the ideal answer. Our experiments show promising results, particularly on the proposed NaLDA evaluation dataset focused on the reconstruction task, with 19.4% saved tokens on average for GPT-4o and only a tiny decrease in evaluation metrics. This suggests that the approach can effectively balance efficiency and accuracy in language processing tasks.
Authors: Lei Yang, Shaoyang Xu, Jianxiang Peng, Shaolin Zhu, Deyi Xiong
Abstract: Large language models (LLMs) based on the Transformer architecture usually have their context length limited due to the high training cost. Recent advancements extend the context window by adjusting the scaling factors of RoPE and fine-tuning. However, suboptimal initialization of these factors results in increased fine-tuning costs and reduced performance at target length. To address these challenges, we propose a novel RoPE-based fine-tuning framework that diverges from conventional scaling factors search. Specifically, we present a \textbf{D}ivide-and-\textbf{C}onquer \textbf{I}ncremental \textbf{S}earch (DCIS) algorithm that strategically determines the better scaling factors. Further fine-tuning with the identified scaling factors effectively extends the context window of LLMs. Empirical results demonstrate that our methodology not only mitigates performance decay at extended target lengths but also allows the model to fine-tune on short contexts and generalize to long contexts, thereby reducing the cost of fine-tuning. The scaling factors obtained through DCIS can even perform effectively without fine-tuning. Further analysis of the search space reveals that DCIS achieves twice the search efficiency compared to other methods. We also examine the impact of the non-strictly increasing scaling factors utilized in DCIS and evaluate the general capabilities of LLMs across various context lengths.
Authors: Hongji Li, Andrianos Michail, Reto Gubelmann, Simon Clematide, Juri Opitz
Abstract: Controllable and transparent text generation has been a long-standing goal in NLP. Almost as long-standing is a general idea for addressing this challenge: Parsing text to a symbolic representation, and generating from it. However, earlier approaches were hindered by parsing and generation insufficiencies. Using modern parsers and a safety supervision mechanism, we show how close current methods come to this goal. Concretely, we propose the Sentence Smith framework for English, which has three steps: 1. Parsing a sentence into a semantic graph. 2. Applying human-designed semantic manipulation rules. 3. Generating text from the manipulated graph. A final entailment check (4.) verifies the validity of the applied transformation. To demonstrate our framework's utility, we use it to induce hard negative text pairs that challenge text embedding models. Since the controllable generation makes it possible to clearly isolate different types of semantic shifts, we can evaluate text embedding models in a fine-grained way, also addressing an issue in current benchmarking where linguistic phenomena remain opaque. Human validation confirms that our transparent generation process produces texts of good quality. Notably, our way of generation is very resource-efficient, since it relies only on smaller neural networks.
Authors: WonJin Yoon, Ian Bulovic, Timothy A. Miller
Abstract: Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.
Authors: Yosuke Yamagishi, Tomohiro Kikuchi, Shouhei Hanaoka, Takeharu Yoshikawa, Osamu Abe
Abstract: Japanese language models for medical text classification face challenges with complex vocabulary and linguistic structures in radiology reports. This study compared three Japanese models--BERT Base, JMedRoBERTa, and ModernBERT--for multi-label classification of 18 chest CT findings. Using the CT-RATE-JPN dataset, all models were fine-tuned under identical conditions. ModernBERT showed clear efficiency advantages, producing substantially fewer tokens and achieving faster training and inference than the other models while maintaining comparable performance on the internal test dataset (exact match accuracy: 74.7% vs. 72.7% for BERT Base). To assess generalizability, we additionally constructed RR-Findings, an external dataset of 243 naturally written Japanese radiology reports annotated using the same schema. Under this domain-shifted setting, performance differences became pronounced: BERT Base outperformed both JMedRoBERTa and ModernBERT, whereas ModernBERT showed the largest decline in exact match accuracy. Average precision differences were smaller, indicating that ModernBERT retained reasonable ranking ability despite reduced calibration. Overall, ModernBERT offers substantial computational efficiency and strong in-domain performance but remains sensitive to real-world linguistic variability. These results highlight the need for more diverse natural-language training data and domain-specific calibration strategies to improve robustness when deploying modern transformer models in heterogeneous clinical environments.
Authors: Thomas Cory, Wolf Rieder, Julia Kr\"amer, Philip Raschke, Patrick Herbke, Axel K\"upper
Abstract: Ensuring transparency of data practices related to personal information is a core requirement of the General Data Protection Regulation (GDPR). However, large-scale compliance assessment remains challenging due to the complexity and diversity of privacy policy language. Manual audits are labour-intensive and inconsistent, while current automated methods often lack the granularity required to capture nuanced transparency disclosures. In this paper, we present a modular large language model (LLM)-based pipeline for fine-grained word-level annotation of privacy policies with respect to GDPR transparency requirements. Our approach integrates LLM-driven annotation with passage-level classification, retrieval-augmented generation, and a self-correction mechanism to deliver scalable, context-aware annotations across 21 GDPR-derived transparency requirements. To support empirical evaluation, we compile a corpus of 703,791 English-language privacy policies and generate a ground-truth sample of 200 manually annotated policies based on a comprehensive, GDPR-aligned annotation scheme. We propose a two-tiered evaluation methodology capturing both passage-level classification and span-level annotation quality and conduct a comparative analysis of seven state-of-the-art LLMs on two annotation schemes, including the widely used OPP-115 dataset. The results of our evaluation show that decomposing the annotation task and integrating targeted retrieval and classification components significantly improve annotation accuracy, particularly for well-structured requirements. Our work provides new empirical resources and methodological foundations for advancing automated transparency compliance assessment at scale.
Authors: Marc Brinner, Tarek Al Mustafa, Sina Zarrie{\ss}
Abstract: We investigate the use of LLM-generated data for continual pretraining of encoder models in specialized domains with limited training data, using the scientific domain of invasion biology as a case study. To this end, we leverage domain-specific ontologies by enriching them with LLM-generated data and pretraining the encoder model as an ontology-informed embedding model for concept definitions. To evaluate the effectiveness of this method, we compile a benchmark specifically designed for assessing model performance in invasion biology. After demonstrating substantial improvements over standard LLM pretraining, we investigate the feasibility of applying the proposed approach to domains without comprehensive ontologies by substituting ontological concepts with concepts automatically extracted from a small corpus of scientific abstracts and establishing relationships between concepts through distributional statistics. Our results demonstrate that this automated approach achieves comparable performance using only a small set of scientific abstracts, resulting in a fully automated pipeline for enhancing domain-specific understanding of small encoder models that is especially suited for application in low-resource settings and achieves performance comparable to masked language modeling pretraining on much larger datasets.
Authors: Xiao Wang, Daniil Larionov, Siwei Wu, Yiqi Liu, Steffen Eger, Nafise Sadat Moosavi, Chenghua Lin
Abstract: Evaluating the quality of generated text automatically remains a significant challenge. Conventional reference-based metrics have been shown to exhibit relatively weak correlation with human evaluations. Recent research advocates the use of large language models (LLMs) as source-based metrics for natural language generation (NLG) assessment. While promising, LLM-based metrics, particularly those using smaller models, still fall short in aligning with human judgments. In this work, we introduce ContrastScore, a contrastive evaluation metric designed to enable higher-quality, less biased, and more efficient assessment of generated text. We evaluate ContrastScore on two NLG tasks: machine translation and summarization. Experimental results show that ContrastScore consistently achieves stronger correlation with human judgments than both single-model and ensemble-based baselines. Notably, ContrastScore based on Qwen 3B and 0.5B even outperforms Qwen 7B, despite having only half as many parameters, demonstrating its efficiency. Furthermore, it effectively mitigates common evaluation biases such as length and likelihood preferences, resulting in more robust automatic evaluation.
Authors: Dongyang Fan, Vinko Sabol\v{c}ec, Martin Jaggi
Abstract: Large Language Models (LLMs) are commonly pretrained on vast corpora of text without utilizing contextual metadata such as source, quality, or topic, leading to a context-free learning paradigm. While recent studies suggest that adding metadata like URL information as context (i.e., auxiliary inputs not used in the loss calculation) can improve training efficiency and downstream performance, they offer limited understanding of which types of metadata are truly effective and under what conditions. In this work, we conduct a systematic evaluation and find that not all metadata types contribute equally. Only URL context speeds up training, whereas quality scores and topic/format domain information offer no clear benefit. Furthermore, the improved downstream performances of URL conditioning emerge only when longer prompts are used at inference time. In addition, we demonstrate that context-aware pretraining enables more controllable generation than context-free pretraining, in a classifier-free guidance fashion. Although topic and format metadata do not accelerate training, they are effective for steering outputs, offering human-interpretable control over generation.
Authors: Yihong Wu, Liheng Ma, Muzhi Li, Jiaming Zhou, Lei Ding, Jianye Hao, Ho-fung Leung, Irwin King, Yingxue Zhang, Jian-Yun Nie
Abstract: Large Language Models (LLMs) equipped with modern Retrieval-Augmented Generation (RAG) systems often employ multi-turn interaction pipelines to interface with search engines for complex reasoning tasks. However, such multi-turn interactions inevitably produce long intermediate contexts, as context length grows exponentially with exploration depth. This leads to a well-known limitation of LLMs: their difficulty in effectively leveraging information from long contexts. This problem is further amplified in RAG systems that depend on in-context learning, where few-shot demonstrations must also be included in the prompt, compounding the context-length bottleneck. To address these challenges, we propose Mujica-MyGo, a unified framework for efficient multi-turn reasoning in RAG. Inspired by the divide-and-conquer principle, we introduce Mujica (Multi-hop Joint Intelligence for Complex Question Answering), a multi-agent RAG workflow that decomposes multi-turn interactions into cooperative sub-interactions, thereby mitigating long-context issues. To eliminate the dependency on in-context learning, we further develop MyGO (Minimalist Policy Gradient Optimization), a lightweight and efficient reinforcement learning algorithm that enables effective post-training of LLMs within complex RAG pipelines. We provide theoretical guarantees for MyGO's convergence to the optimal policy. Empirical evaluations across diverse question-answering benchmarks, covering both text corpora and knowledge graphs, show that Mujica-MyGO achieves superior performance.
Authors: Niranjan Chebrolu, Gerard Christopher Yeo, Kokil Jaidka
Abstract: Large Language Models (LLMs) demonstrate increasing conversational fluency, yet instilling them with nuanced, human-like emotional expression remains a significant challenge. Current alignment techniques often address surface-level output or require extensive fine-tuning. This paper demonstrates that targeted activation engineering can steer LLaMA 3.1-8B to exhibit more human-like emotional nuances. We first employ attribution patching to identify causally influential components, to find a key intervention locus by observing activation patterns during diagnostic conversational tasks. We then derive emotional expression vectors from the difference in the activations generated by contrastive text pairs (positive vs. negative examples of target emotions). Applying these vectors to new conversational prompts significantly enhances emotional characteristics: steered responses show increased positive sentiment (e.g., joy, trust) and more frequent first-person pronoun usage, indicative of greater personal engagement. Our findings offer a precise and interpretable method for controlling specific emotional attributes in LLMs, contributing to developing more aligned and empathetic conversational AI.
Authors: Masoomali Fatehkia, Enes Altinisik, Mohamed Osman, Husrev Taha Sencar
Abstract: Aligning large language models (LLMs) with deployment-specific requirements is critical but inherently imperfect. Despite extensive training, models remain susceptible to misalignment and adversarial inputs such as jailbreaks. Content moderation filters are commonly used as external safeguards, though they typically focus narrowly on safety. We introduce SGM (Specification-Guided Moderation), a flexible framework for training moderation filters grounded in user-defined specifications that go beyond standard safety concerns. SGM automates training data generation without relying on human-written examples, enabling scalable support for diverse, application-specific alignment goals. SGM-trained filters perform on par with state-of-the-art safety filters built on curated datasets, while supporting fine-grained and user-defined alignment control.
Authors: Ziju Shen, Naohao Huang, Fanyi Yang, Yutong Wang, Guoxiong Gao, Tianyi Xu, Jiedong Jiang, Wanyi He, Pu Yang, Mengzhou Sun, Haocheng Ju, Peihao Wu, Bryan Dai, Bin Dong
Abstract: Nowadays, formal theorem provers have made monumental progress on high-school and competition-level mathematics, but few of them generalize to more advanced mathematics. In this paper, we present REAL-Prover, a new open-source stepwise theorem prover for Lean 4 to push this boundary. This prover, based on our fine-tuned large language model (REAL-Prover-v1) and integrated with a retrieval system (Leansearch-PS), notably boosts performance on solving college-level mathematics problems. To train REAL-Prover-v1, we developed HERALD-AF, a data extraction pipeline that converts natural language math problems into formal statements, and a new open-source Lean 4 interactive environment (Jixia-interactive) to facilitate synthesis data collection. In our experiments, our prover using only supervised fine-tune achieves competitive results with a 23.7% success rate (Pass@64) on the ProofNet dataset-comparable to state-of-the-art (SOTA) models. To further evaluate our approach, we introduce FATE-M, a new benchmark focused on algebraic problems, where our prover achieves a SOTA success rate of 56.7% (Pass@64).
Authors: Shimao Zhang, Zhejian Lai, Xiang Liu, Shuaijie She, Xiao Liu, Yeyun Gong, Shujian Huang, Jiajun Chen
Abstract: Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of ''Spontaneous Multilingual Alignment''. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.
Authors: Yujin Choi, Youngjoo Park, Junyoung Byun, Jaewook Lee, Jinseong Park
Abstract: Retrieval-augmented generation (RAG) mitigates the hallucination problem in large language models (LLMs) and has proven effective for personalized usages. However, delivering private retrieved documents directly to LLMs introduces vulnerability to membership inference attacks (MIAs), which try to determine whether the target data point exists in the private external database or not. Based on the insight that MIA queries typically exhibit high similarity to only one target document, we introduce a novel similarity-based MIA detection framework designed for the RAG system. With the proposed method, we show that a simple detect-and-hide strategy can successfully obfuscate attackers, maintain data utility, and remain system-agnostic against MIA. We experimentally prove its detection and defense against various state-of-the-art MIA methods and its adaptability to existing RAG systems.
Authors: Runyu Wang, Peng Ping, Zhengyu Guo, Xiaoye Zhang, Quan Shi, Liting Zhou, Tianbo Ji
Abstract: Fine-tuning adapts pretrained models for specific tasks but poses the risk of catastrophic forgetting (CF), where critical knowledge from pretraining is overwritten. To address the issue of CF in a general-purpose framework, we propose Low-damage Knowledge Implanting (LoKI), a parameter-efficient fine-tuning (PEFT) technique that utilizes recent mechanistic understanding of how knowledge is stored in transformer architectures. We compare LoKI against state-of-the-art PEFT methods in two real-world fine-tuning scenarios. The results show that LoKI demonstrates significantly better preservation of general capabilities. At the same time, its task-specific performance is comparable to or even surpasses that of full parameter fine-tuning and these PEFT methods across various model architectures. Our work bridges the mechanistic insights of LLMs' knowledge storage with practical fine-tuning objectives, enabling an effective balance between task-specific adaptation and the retention of general-purpose capabilities.
Authors: Yida Xue, Zhen Bi, Jinnan Yang, Jungang Lou, Kehai Chen, Min Zhang, Huajun Chen, Ningyu Zhang
Abstract: Recent advances in Multimodal Large Language Models (MLLMs) have significantly enhanced their capabilities; however, their spatial perception abilities remain a notable limitation. To address this challenge, multimodal data synthesis offers a promising solution. Yet, ensuring that synthesized data adhere to spatial common sense is a non-trivial task. Our approach addresses this critical gap by providing a systematic framework for generating spatially coherent data. In this work, we introduce SKG2DATA, a novel multimodal synthesis approach guided by spatial knowledge graphs, grounded in the concept of knowledge-to-data generation. SKG2DATA employs an automated pipeline for constructing Spatial Knowledge Graph (SKG) that effectively captures human-like spatial cognition, including directional and distance relationships. These structured representations then serve as precise guidance for our integrated synthesis pipeline, where a diffusion model generates spatially-consistent images while a MLLM produces corresponding textual descriptions. The automated construction of SKG enables scalable generation of diverse yet realistic spatial configurations, overcoming the limitations of manual data collection and annotation. Extensive experiments demonstrate that data synthesized from diverse types of spatial knowledge, including direction and distance, enhance the spatial perception and reasoning abilities of MLLMs markedly, albeit with a slight cost to their general capabilities. We hope that the idea of knowledge-based data synthesis can advance the development of spatial intelligence. Code is available at https://github.com/zjunlp/Knowledge2Data.
Authors: Jinzhe Li, Gengxu Li, Yi Chang, Yuan Wu
Abstract: Large language models (LLMs) have witnessed rapid advancements, demonstrating remarkable capabilities. However, a notable vulnerability persists: LLMs often uncritically accept flawed or contradictory premises, leading to inefficient reasoning and unreliable outputs. This emphasizes the significance of possessing the \textbf{Premise Critique Ability} for LLMs, defined as the capacity to proactively identify and articulate errors in input premises. Most existing studies assess LLMs' reasoning ability in ideal settings, largely ignoring their vulnerabilities when faced with flawed premises. Thus, we introduce the \textbf{Premise Critique Bench (PCBench)}, designed by incorporating four error types across three difficulty levels, paired with multi-faceted evaluation metrics. We conducted systematic evaluations of 15 representative LLMs. Our findings reveal: (1) Most models rely heavily on explicit prompts to detect errors, with limited autonomous critique; (2) Premise critique ability depends on question difficulty and error type, with direct contradictions being easier to detect than complex or procedural errors; (3) Reasoning ability does not consistently correlate with the premise critique ability; (4) Flawed premises trigger overthinking in reasoning models, markedly lengthening responses due to repeated attempts at resolving conflicts. These insights underscore the urgent need to enhance LLMs' proactive evaluation of input validity, positioning premise critique as a foundational capability for developing reliable, human-centric systems. The code is available at https://github.com/MLGroupJLU/Premise_Critique.
Authors: Xiaoyuan Wu, Weiran Lin, Omer Akgul, Lujo Bauer
Abstract: Large language models (LLMs) are prone to hallucinations and sensitive to prompt perturbations, often resulting in inconsistent or unreliable generated text. Different methods have been proposed to mitigate such hallucinations and fragility, one of which is to measure the consistency of LLM responses -- the model's confidence in the response or likelihood of generating a similar response when resampled. In previous work, measuring LLM response consistency often relied on calculating the probability of a response appearing within a pool of resampled responses, analyzing internal states, or evaluating logits of responses. However, it was not clear how well these approaches approximated users' perceptions of consistency of LLM responses. To find out, we performed a user study ($n=2,976$) demonstrating that current methods for measuring LLM response consistency typically do not align well with humans' perceptions of LLM consistency. We propose a logit-based ensemble method for estimating LLM consistency and show that our method matches the performance of the best-performing existing metric in estimating human ratings of LLM consistency. Our results suggest that methods for estimating LLM consistency without human evaluation are sufficiently imperfect to warrant broader use of evaluation with human input; this would avoid misjudging the adequacy of models because of the imperfections of automated consistency metrics.
Authors: Silin Gao, Antoine Bosselut, Samy Bengio, Emmanuel Abbe
Abstract: Recent studies have shown that large language models (LLMs), especially smaller ones, often lack robustness in grade school math (GSM) reasoning. In particular, they tend to experience performance drops when faced with distribution shifts, such as changes to numerical or nominal variables, or insertions of distracting clauses. A possible strategy to address this involves generating synthetic data to further "instantiate" reasoning problems on potential variations. In this work, we instead focuses on the strategy of "abstracting" reasoning problems. This not only helps counteract distribution shifts but also facilitates the connection to symbolic tools for deriving solutions. Focusing on GSM, we find that this abstraction process is better acquired through reinforcement learning (RL) than just supervised fine-tuning, which often fails to produce faithful abstractions. Our method, AbstRaL -- which promotes abstract reasoning in LLMs using RL on granular abstraction data -- significantly mitigates performance degradation on recent GSM perturbation benchmarks. Besides, improving GSM robustness via AbstRaL is shown to also implicitly benefit LLMs' capabilities on OOD mathematical and general reasoning tasks, indicating that abstract thinking broadly enables better generalizability.
Authors: Pengbo Wang, Chaozhuo Li, Chenxu Wang, Liwen Zheng, Litian Zhang, Xi Zhang
Abstract: LLMs have demonstrated unprecedented capabilities in natural language processing, yet their practical deployment remains hindered by persistent factuality and faithfulness hallucinations. While existing methods address these hallucination types independently, they inadvertently induce performance trade-offs, as interventions targeting one type often exacerbate the other. Through empirical and theoretical analysis of activation space dynamics in LLMs, we reveal that these hallucination categories share overlapping subspaces within neural representations, presenting an opportunity for concurrent mitigation. To harness this insight, we propose SPACE, a unified framework that jointly enhances factuality and faithfulness by editing shared activation subspaces. SPACE establishes a geometric foundation for shared subspace existence through dual-task feature modeling, then identifies and edits these subspaces via a hybrid probe strategy combining spectral clustering and attention head saliency scoring. Experimental results across multiple benchmark datasets demonstrate the superiority of our approach.
Authors: Hyungjune Bu, Chanjoo Jung, Minjae Kang, Jaehyung Kim
Abstract: As large language models (LLMs) are progressively deployed in various real-world applications, personalization of LLMs has become increasingly important. While various approaches to LLM personalization such as prompt-based and training-based methods have been actively explored, the development of effective decoding-time algorithms remains largely overlooked, despite their demonstrated potential. In this paper, we propose CoPe (Contrasting Personal Preference), a novel decoding-time approach applied after performing parameter-efficient fine-tuning (PEFT) on user-specific data. Our core idea is to leverage reward-guided decoding specifically for personalization by maximizing each user's implicit reward signal. We evaluate CoPe across five open-ended personalized text generation tasks. Our empirical results demonstrate that CoPe achieves strong performance, improving personalization by an average of 10.57% in ROUGE-L, without relying on external reward models or additional training procedures.
Authors: Lincan Li, Eren Erman Ozguven, Yue Zhao, Guang Wang, Yiqun Xie, Yushun Dong
Abstract: Accurate typhoon track forecasting is crucial for early system warning and disaster response. While Transformer-based models have demonstrated strong performance in modeling the temporal dynamics of dense trajectories of humans and vehicles in smart cities, they usually lack access to broader contextual knowledge that enhances the forecasting reliability of sparse meteorological trajectories, such as typhoon tracks. To address this challenge, we propose TyphoFormer, a novel framework that incorporates natural language descriptions as auxiliary prompts to improve typhoon trajectory forecasting. For each time step, we use Large Language Model (LLM) to generate concise textual descriptions based on the numerical attributes recorded in the North Atlantic hurricane database. The language descriptions capture high-level meteorological semantics and are embedded as auxiliary special tokens prepended to the numerical time series input. By integrating both textual and sequential information within a unified Transformer encoder, TyphoFormer enables the model to leverage contextual cues that are otherwise inaccessible through numerical features alone. Extensive experiments are conducted on HURDAT2 benchmark, results show that TyphoFormer consistently outperforms other state-of-the-art baseline methods, particularly under challenging scenarios involving nonlinear path shifts and limited historical observations.
Authors: Haoze Wu, Yunzhi Yao, Wenhao Yu, Ningyu Zhang
Abstract: Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes. Specifically, we construct a dataset of approximately 2,000 data entries to train the LLMs to perform version migration based on updated information. Then, we introduce a modified string similarity metric for code evaluation as the reward for reinforcement learning. Our experiments demonstrate that ReCode substantially boosts LLMs' code generation performance in dynamic API scenarios, especially on the unseen CodeUpdateArena task. Crucially, compared to supervised fine-tuning, ReCode has less impact on LLMs' general code generation abilities. We apply ReCode on various LLMs and reinforcement learning algorithms (GRPO and DAPO), all achieving consistent improvements. Notably, after training, Qwen2.5-Coder-7B outperforms that of the 32B parameter code instruction-tuned model and the reasoning model with the same architecture. Code is available at https://github.com/zjunlp/ReCode.
Authors: Md Rakibul Hasan, Md Zakir Hossain, Aneesh Krishna, Shafin Rahman, Tom Gedeon
Abstract: Noisy self-reported empathy scores challenge supervised learning for empathy regression. While many algorithms have been proposed for learning with noisy labels in textual classification problems, the regression counterpart is relatively under-explored. We propose UPLME, an uncertainty-aware probabilistic language modelling framework to capture label noise in empathy regression tasks. One of the novelties in UPLME is a probabilistic language model that predicts both empathy scores and heteroscedastic uncertainty, and is trained using Bayesian concepts with variational model ensembling. We further introduce two novel loss components: one penalises degenerate Uncertainty Quantification (UQ), and another enforces similarity between the input pairs on which empathy is being predicted. UPLME achieves state-of-the-art performance (Pearson Correlation Coefficient: $0.558\rightarrow0.580$ and $0.629\rightarrow0.634$) in terms of the performance reported in the literature on two public benchmarks with label noise. Through synthetic label noise injection, we demonstrate that UPLME is effective in distinguishing between noisy and clean samples based on the predicted uncertainty. UPLME further outperform (Calibration error: $0.571\rightarrow0.376$) a recent variational model ensembling-based UQ method designed for regression problems. Code is publicly available at https://github.com/hasan-rakibul/UPLME.
Authors: Yutong Wu, Di Huang, Ruosi Wan, Yue Peng, Shijie Shang, Chenrui Cao, Lei Qi, Rui Zhang, Zidong Du, Jie Yan, Xing Hu
Abstract: Autoformalization aims to translate natural-language mathematical statements into a formal language. While LLMs have accelerated progress in this area, existing methods still suffer from low accuracy. We identify two key abilities for effective autoformalization: comprehensive mastery of formal-language domain knowledge, and reasoning capability of natural language problem understanding and informal-formal alignment. Without the former, a model cannot identify the correct formal objects; without the latter, it struggles to interpret real-world contexts and map them precisely into formal expressions. To address these gaps, we introduce ThinkingF, a data synthesis and training pipeline that improves both abilities. First, we construct two datasets: one by distilling and selecting large-scale examples rich in formal knowledge, and another by generating informal-to-formal reasoning trajectories guided by expert-designed templates. We then apply SFT and RLVR with these datasets to further fuse and refine the two abilities. The resulting 7B and 32B models exhibit both comprehensive formal knowledge and strong informal-to-formal reasoning. Notably, StepFun-Formalizer-32B achieves SOTA BEq@1 scores of 40.5% on FormalMATH-Lite and 26.7% on ProverBench, surpassing all prior general-purpose and specialized models.
Authors: Lingkun Long, Rubing Yang, Yushi Huang, Desheng Hui, Ao Zhou, Jianlei Yang
Abstract: Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to $\mathbf{2.53\times}$ time-to-first-token (TTFT) speedup and $\mathbf{1.88\times}$ end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code is available at https://github.com/Longxmas/SlimInfer.
Authors: Wenpeng Xing, Lanyi Wei, Haixiao Hu, Rongchang Li, Mohan Li, Changting Lin, Meng Han
Abstract: The rapid proliferation of large language models (LLMs) in applications targeting children and adolescents necessitates a fundamental reassessment of prevailing AI safety frameworks, which are largely tailored to adult users and neglect the distinct developmental vulnerabilities of minors. This paper highlights key deficiencies in existing LLM safety benchmarks, including their inadequate coverage of age-specific cognitive, emotional, and social risks spanning early childhood (ages 0--6), middle childhood (7--12), and adolescence (13--18). To bridge these gaps, we introduce SproutBench, an innovative evaluation suite comprising 1,283 developmentally grounded adversarial prompts designed to probe risks such as emotional dependency, privacy violations, and imitation of hazardous behaviors. Through rigorous empirical evaluation of 47 diverse LLMs, we uncover substantial safety vulnerabilities, corroborated by robust inter-dimensional correlations (e.g., between Safety and Risk Prevention) and a notable inverse relationship between Interactivity and Age Appropriateness. These insights yield practical guidelines for advancing child-centric AI design and deployment.
Authors: Jinzhong Ning, Paerhati Tulajiang, Yingying Le, Yijia Zhang, Yuanyuan Sun, Hongfei Lin, Haifeng Liu
Abstract: Speech Relation Extraction (SpeechRE) aims to extract relation triplets directly from speech. However, existing benchmark datasets rely heavily on synthetic data, lacking sufficient quantity and diversity of real human speech. Moreover, existing models also suffer from rigid single-order generation templates and weak semantic alignment, substantially limiting their performance. To address these challenges, we introduce CommonVoice-SpeechRE, a large-scale dataset comprising nearly 20,000 real-human speech samples from diverse speakers, establishing a new benchmark for SpeechRE research. Furthermore, we propose the Relation Prompt-Guided Multi-Order Generative Ensemble (RPG-MoGe), a novel framework that features: (1) a multi-order triplet generation ensemble strategy, leveraging data diversity through diverse element orders during both training and inference, and (2) CNN-based latent relation prediction heads that generate explicit relation prompts to guide cross-modal alignment and accurate triplet generation. Experiments show our approach outperforms state-of-the-art methods, providing both a benchmark dataset and an effective solution for real-world SpeechRE. The source code and dataset are publicly available at https://github.com/NingJinzhong/SpeechRE_RPG_MoGe.
Authors: Aya E. Fouda, Abdelrahamn A. Hassan, Radwa J. Hanafy, Mohammed E. Fouda
Abstract: Large language models (LLMs) offer significant potential in enhancing psychiatric practice, from improving diagnostic accuracy to streamlining clinical documentation and therapeutic support. However, existing evaluation resources heavily rely on small clinical interview corpora, social media posts, or synthetic dialogues, which limits their clinical validity and fails to capture the full complexity of diagnostic reasoning. In this work, we introduce PsychiatryBench, a rigorously curated benchmark grounded exclusively in authoritative, expert-validated psychiatric textbooks and casebooks. PsychiatryBench comprises eleven distinct question-answering tasks ranging from diagnostic reasoning and treatment planning to longitudinal follow-up, management planning, clinical approach, sequential case analysis, and multiple-choice/extended matching formats totaling 5,188 expert-annotated items. {\color{red}We evaluate a diverse set of frontier LLMs (including Google Gemini, DeepSeek, Sonnet 4.5, and GPT 5) alongside leading open-source medical models such as MedGemma using both conventional metrics and an "LLM-as-judge" similarity scoring framework. Our results reveal substantial gaps in clinical consistency and safety, particularly in multi-turn follow-up and management tasks, underscoring the need for specialized model tuning and more robust evaluation paradigms. PsychiatryBench offers a modular, extensible platform for benchmarking and improving LLM performance in mental health applications.
Authors: Sreejato Chatterjee, Linh Tran, Quoc Duy Nguyen, Roni Kirson, Drue Hamlin, Harvest Aquino, Hanjia Lyu, Jiebo Luo, Timothy Dye
Abstract: Traditional efforts to measure historical structural oppression struggle with cross-national validity due to the unique, locally specified histories of exclusion, colonization, and social status in each country, and often have relied on structured indices that privilege material resources while overlooking lived, identity-based exclusion. We introduce a novel framework for oppression measurement that leverages Large Language Models (LLMs) to generate context-sensitive scores of lived historical disadvantage across diverse geopolitical settings. Using unstructured self-identified ethnicity utterances from a multilingual COVID-19 global study, we design rule-guided prompting strategies that encourage models to produce interpretable, theoretically grounded estimations of oppression. We systematically evaluate these strategies across multiple state-of-the-art LLMs. Our results demonstrate that LLMs, when guided by explicit rules, can capture nuanced forms of identity-based historical oppression within nations. This approach provides a complementary measurement tool that highlights dimensions of systemic exclusion, offering a scalable, cross-cultural lens for understanding how oppression manifests in data-driven research and public health contexts. To support reproducible evaluation, we release an open-sourced benchmark dataset for assessing LLMs on oppression measurement (https://github.com/chattergpt/HSO-Bench).
Authors: Madison Van Doren, Casey Ford
Abstract: Multimodal large language models (MLLMs) are increasingly used in real world applications, yet their safety under adversarial conditions remains underexplored. This study evaluates the harmlessness of four leading MLLMs (GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus) when exposed to adversarial prompts across text-only and multimodal formats. A team of 26 red teamers generated 726 prompts targeting three harm categories: illegal activity, disinformation, and unethical behaviour. These prompts were submitted to each model, and 17 annotators rated 2,904 model outputs for harmfulness using a 5-point scale. Results show significant differences in vulnerability across models and modalities. Pixtral 12B exhibited the highest rate of harmful responses (~62%), while Claude Sonnet 3.5 was the most resistant (~10%). Contrary to expectations, text-only prompts were slightly more effective at bypassing safety mechanisms than multimodal ones. Statistical analysis confirmed that both model type and input modality were significant predictors of harmfulness. These findings underscore the urgent need for robust, multimodal safety benchmarks as MLLMs are deployed more widely.
Authors: Yanfang Ye, Zheyuan Zhang, Tianyi Ma, Zehong Wang, Yiyang Li, Shifu Hou, Weixiang Sun, Kaiwen Shi, Yijun Ma, Wei Song, Ahmed Abbasi, Ying Cheng, Jane Cleland-Huang, Steven Corcelli, Robert Goulding, Ming Hu, Ting Hua, John Lalor, Fang Liu, Tengfei Luo, Edward Maginn, Nuno Moniz, Jason Rohr, Brett Savoie, Daniel Slate, Matthew Webber, Olaf Wiest, Johnny Zhang, Nitesh V. Chawla
Abstract: Cutting-edge Artificial Intelligence (AI) techniques keep reshaping our view of the world. For example, Large Language Models (LLMs) based applications such as ChatGPT have shown the capability of generating human-like conversation on extensive topics. Due to the impressive performance on a variety of language-related tasks (e.g., open-domain question answering, translation, and document summarization), one can envision the far-reaching impacts that can be brought by the LLMs with broader real-world applications (e.g., customer service, education and accessibility, and scientific discovery). Inspired by their success, this paper will offer an overview of state-of-the-art LLMs and their integration into a wide range of academic disciplines, including: (1) arts, letters, and law (e.g., history, philosophy, political science, arts and architecture, law), (2) economics and business (e.g., finance, economics, accounting, marketing), and (3) science and engineering (e.g., mathematics, physics and mechanical engineering, chemistry and chemical engineering, life sciences and bioengineering, earth sciences and civil engineering, computer science and electrical engineering). Integrating humanity and technology, in this paper, we will explore how LLMs are shaping research and practice in these fields, while also discussing key limitations, open challenges, and future directions in the era of generative AI. The review of how LLMs are engaged across disciplines-along with key observations and insights-can help researchers and practitioners interested in exploiting LLMs to advance their works in diverse real-world applications.
Authors: Gustavo Cilleruelo, Emily Allaway, Barry Haddow, Alexandra Birch
Abstract: MGen is a dataset of over 4 million naturally occurring generic and quantified sentences extracted from diverse textual sources. Sentences in the dataset have long context documents, corresponding to websites and academic papers, and cover 11 different quantifiers. We analyze the features of generics sentences in the dataset, with interesting insights: generics can be long sentences (averaging over 16 words) and speakers often use them to express generalisations about people. MGen is the biggest and most diverse dataset of naturally occurring generic sentences, opening the door to large-scale computational research on genericity. It is publicly available at https://gustavocilleruelo.com/mgen
Authors: Yubo Li, Ramayya Krishnan, Rema Padman
Abstract: Large Language Models (LLMs) have revolutionized conversational AI, yet their robustness in extended multi-turn dialogues remains poorly understood. Existing evaluation frameworks focus on static benchmarks and single-turn assessments, failing to capture the temporal dynamics of conversational degradation that characterize real-world interactions. In this work, we present a large-scale survival analysis of conversational robustness, modeling failure as a time-to-event process over 36,951 turns from 9 state-of-the-art LLMs on the MT-Consistency benchmark. Our framework combines Cox proportional hazards, Accelerated Failure Time (AFT), and Random Survival Forest models with simple semantic drift features. We find that abrupt prompt-to-prompt semantic drift sharply increases the hazard of inconsistency, whereas cumulative drift is counterintuitively \emph{protective}, suggesting adaptation in conversations that survive multiple shifts. AFT models with model-drift interactions achieve the best combination of discrimination and calibration, and proportional hazards checks reveal systematic violations for key drift covariates, explaining the limitations of Cox-style modeling in this setting. Finally, we show that a lightweight AFT model can be turned into a turn-level risk monitor that flags most failing conversations several turns before the first inconsistent answer while keeping false alerts modest. These results establish survival analysis as a powerful paradigm for evaluating multi-turn robustness and for designing practical safeguards for conversational AI systems.
Authors: Jahidul Arafat, Fariha Tasmin, Sanjaya Poudel
Abstract: Wordle presents an algorithmically rich testbed for constraint satisfaction problem (CSP) solving. While existing solvers rely on information-theoretic entropy maximization or frequency-based heuristics without formal constraint treatment, we present the first comprehensive CSP formulation of Wordle with novel constraint-aware solving strategies. We introduce CSP-Aware Entropy, computing information gain after constraint propagation rather than on raw candidate sets, and a Probabilistic CSP framework integrating Bayesian word-frequency priors with logical constraints. Through evaluation on 2,315 English words, CSP-Aware Entropy achieves 3.54 average guesses with 99.9% success rate, a statistically significant 1.7% improvement over Forward Checking (t=-4.82, p<0.001, Cohen's d=0.07) with 46% faster runtime (12.9ms versus 23.7ms per guess). Under 10% noise, CSP-aware approaches maintain 5.3 percentage point advantages (29.0% versus 23.7%, p=0.041), while Probabilistic CSP achieves 100% success across all noise levels (0-20%) through constraint recovery mechanisms. Cross-lexicon validation on 500 Spanish words demonstrates 88% success with zero language-specific tuning, validating that core CSP principles transfer across languages despite an 11.2 percentage point gap from linguistic differences (p<0.001, Fisher's exact test). Our open-source implementation with 34 unit tests achieving 91% code coverage provides reproducible infrastructure for CSP research. The combination of formal CSP treatment, constraint-aware heuristics, probabilistic-logical integration, robustness analysis, and cross-lexicon validation establishes new performance benchmarks demonstrating that principled constraint satisfaction techniques outperform classical information-theoretic and learning-based approaches for structured puzzle-solving domains.
Authors: Canhui Wu, Qiong Cao, Chang Li, Zhenfang Wang, Chao Xue, Yuwei Fan, Wei Xi, Xiaodong He
Abstract: Large Reasoning Models (LRMs) demonstrate strong performance on complex tasks but often suffer from excessive verbosity, known as "overthinking." Existing solutions via reinforcement learning (RL) typically penalize generated tokens to promote conciseness. However, these methods encounter two challenges: responses with fewer tokens do not always correspond to fewer reasoning steps, and models may develop hacking behavior in later stages of training by discarding reasoning steps to minimize token usage. In this work, we introduce \textbf{Step Pruner (SP)}, an RL framework that steers LRMs toward more efficient reasoning by favoring compact reasoning steps. Our step-aware reward function prioritizes correctness while imposing penalties for redundant steps, and withholds rewards for incorrect responses to prevent the reinforcement of erroneous reasoning. Moreover, we propose a dynamic stopping mechanism: when the model's output no longer shortens, training is halted to prevent hacking behavior caused by the merging of steps. Extensive experiments across four reasoning benchmarks demonstrate that SP achieves state-of-the-art accuracy while significantly reducing response length. For instance, on AIME24, SP reduces token usage by \textbf{69.7\%}.
Authors: Vardhan Dongre, Ryan A. Rossi, Viet Dac Lai, David Seunghyun Yoon, Dilek Hakkani-T\"ur, Trung Bui
Abstract: Large Language Models (LLMs) excel at single-turn tasks such as instruction following and summarization, yet real-world deployments require sustained multi-turn interactions where user goals and conversational context persist and evolve. A recurring challenge in this setting is context drift: the gradual divergence of a model's outputs from goal-consistent behavior across turns. Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics. In this work, we present a study of context drift in multi-turn interactions and propose a simple dynamical framework to interpret its behavior. We formalize drift as the turn-wise KL divergence between the token-level predictive distributions of the test model and a goal-consistent reference model, and propose a recurrence model that interprets its evolution as a bounded stochastic process with restoring forces and controllable interventions. We instantiate this framework in both synthetic long-horizon rewriting tasks and realistic user-agent simulations such as in $\tau$-Bench, measuring drift for several open-weight LLMs that are used as user simulators. Our experiments consistently reveal stable, noise-limited equilibria rather than runaway degradation, and demonstrate that simple reminder interventions reliably reduce divergence in line with theoretical predictions. Together, these results suggest that multi-turn drift can be understood as a controllable equilibrium phenomenon rather than as inevitable decay, providing a foundation for studying and mitigating context drift in extended interactions.
Authors: Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul, Uddip Acharjee Shuvo, Muhit Islam Emon, Xuan Wang, Liqing Zhang
Abstract: Large language models (LLMs) and emerging agentic frameworks are beginning to transform single-cell biology by enabling natural-language reasoning, generative annotation, and multimodal data integration. However, progress remains fragmented across data modalities, architectures, and evaluation standards. LLM4Cell presents the first unified survey of 58 foundation and agentic models developed for single-cell research, spanning RNA, ATAC, multi-omic, and spatial modalities. We categorize these methods into five families-foundation, text-bridge, spatial, multimodal, epigenomic, and agentic-and map them to eight key analytical tasks including annotation, trajectory and perturbation modeling, and drug-response prediction. Drawing on over 40 public datasets, we analyze benchmark suitability, data diversity, and ethical or scalability constraints, and evaluate models across 10 domain dimensions covering biological grounding, multi-omics alignment, fairness, privacy, and explainability. By linking datasets, models, and evaluation domains, LLM4Cell provides the first integrated view of language-driven single-cell intelligence and outlines open challenges in interpretability, standardization, and trustworthy model development.
Authors: Jingbiao Mei, Mingsheng Sun, Jinghong Chen, Pengda Qin, Yuhong Li, Da Chen, Bill Byrne
Abstract: Hateful memes have emerged as a particularly challenging form of online abuse, motivating the development of automated detection systems. Most prior approaches rely on direct detection, producing only binary predictions. Such models fail to provide the context and explanations that real-world moderation requires. Recent Explain-then-Detect approaches, using Chain-of-Thought prompting or LMM agents, perform worse than simple SFT baselines, and even advanced post-training methods such as GRPO fail to close the gap. Our analysis identifies two key issues of such systems: important policy-relevant cues such as targets and attack types are not hypothesized by the model as a likely explanation; and the binary reward signal is insufficient to guide reasoning. To address these challenges, we propose ExPO-HM (Explain-then-Detect Policy Optimization for Hateful Memes), inspired by the training and evaluation process of human annotators. ExPO-HM combines SFT warmup, GRPO with curriculum learning, and Conditional Decision Entropy (CDE) as both metric and reward for reasoning quality. Across three hateful meme benchmarks, ExPO-HM achieves state-of-the-art performance on binary detection, fine-grained classification, and reasoning quality, with up to 15\% and 17\% F1 improvement over the GRPO and DPO baselines, respectively. By moving hateful meme detection from simple binary alarms to explanation-driven detection, ExPO-HM provides accurate, interpretable, and actionable moderation support.
Authors: Joe Watson, Ivan O'Connor, Chia-Wen Chen, Luning Sun, Fang Luo, David Stillwell
Abstract: Psychological assessments are dominated by rating scales, which cannot capture the nuance in natural language. Efforts to supplement them with qualitative text have relied on labelled datasets or expert rubrics, limiting scalability. We introduce a framework that avoids this reliance: large language models (LLMs) score free-text responses with simple prompts to produce candidate LLM items, from which we retain those that yield the most test information when co-calibrated with a baseline scale. Using depression as a case study, we developed and tested the method in upper-secondary students (n=693) and a matched synthetic dataset (n=3,000). Results on held-out test sets showed that augmenting a 19-item scale with LLM items improved its precision, accuracy, and convergent validity. Further, the test information gain matched that of adding as many as 16 rating-scale items. This framework leverages the increasing availability of transcribed language to enhance psychometric measures, with applications in clinical health and beyond.
Authors: Mar\'ia Victoria Carro, Denise Alejandra Mester, Facundo Nieto, Oscar Agust\'in Stanchi, Guido Ernesto Bergman, Mario Alejandro Leiva, Eitan Sprejer, Luca Nicol\'as Forziati Gangi, Francisca Gauna Selasco, Juan Gustavo Corval\'an, Gerardo I. Simari, Mar\'ia Vanina Martinez
Abstract: The core premise of AI debate as a scalable oversight technique is that it is harder to lie convincingly than to refute a lie, enabling the judge to identify the correct position. Yet, existing debate experiments have relied on datasets with ground truth, where lying is reduced to defending an incorrect proposition. This overlooks a subjective dimension: lying also requires the belief that the claim defended is false. In this work, we apply debate to subjective questions and explicitly measure large language models' prior beliefs before experiments. Debaters were asked to select their preferred position, then presented with a judge persona deliberately designed to conflict with their identified priors. This setup tested whether models would adopt sycophantic strategies, aligning with the judge's presumed perspective to maximize persuasiveness, or remain faithful to their prior beliefs. We implemented and compared two debate protocols, sequential and simultaneous, to evaluate potential systematic biases. Finally, we assessed whether models were more persuasive and produced higher-quality arguments when defending positions consistent with their prior beliefs versus when arguing against them. Our main findings show that models tend to prefer defending stances aligned with the judge persona rather than their prior beliefs, sequential debate introduces significant bias favoring the second debater, models are more persuasive when defending positions aligned with their prior beliefs, and paradoxically, arguments misaligned with prior beliefs are rated as higher quality in pairwise comparison. These results can inform human judges to provide higher-quality training signals and contribute to more aligned AI systems, while revealing important aspects of human-AI interaction regarding persuasion dynamics in language models.
Authors: Christine Ye, Sihan Yuan, Suchetha Cooray, Steven Dillmann, Ian L. V. Roque, Dalya Baron, Philipp Frank, Sergio Martin-Alvarez, Nolan Koblischke, Frank J Qu, Diyi Yang, Risa Wechsler, Ioana Ciuca
Abstract: Frontier AI agents show increasing promise as scientific research assistants, and may eventually be useful for extended, open-ended research workflows. However, in order to use agents for novel research, we must first assess the underlying faithfulness and correctness of their work. To evaluate agents as research assistants, we introduce ReplicationBench, an evaluation framework that tests whether agents can replicate entire research papers drawn from the astrophysics literature. Astrophysics, where research relies heavily on archival data and computational study while requiring little real-world experimentation, is a particularly useful testbed for AI agents in scientific research. We split each paper into tasks which require agents to replicate the paper's core contributions, including the experimental setup, derivations, data analysis, and codebase. Each task is co-developed with the original paper authors and targets a key scientific result, enabling objective evaluation of both faithfulness (adherence to original methods) and correctness (technical accuracy of results). ReplicationBench is extremely challenging for current frontier language models: even the best-performing language models score under 20%. We analyze ReplicationBench trajectories in collaboration with domain experts and find a rich, diverse set of failure modes for agents in scientific research. ReplicationBench establishes the first benchmark of paper-scale, expert-validated astrophysics research tasks, reveals insights about agent performance generalizable to other domains of data-driven science, and provides a scalable framework for measuring AI agents' reliability in scientific research.
Authors: P. Bilha Githinji, Aikaterini Meilliou, Zeming Liang, Lian Zhang, Peiwu Qin
Abstract: The increasing health-seeking behavior and digital consumption of biomedical information by the general public necessitate scalable solutions for automatically adapting complex scientific and technical documents into plain language. Automatic text simplification solutions, including advanced large language models (LLMs), however, continue to face challenges in reliably arbitrating the tension between optimizing readability performance and ensuring preservation of discourse fidelity. This report empirically assesses two major classes of general-purpose LLMs, demonstrating how they navigate the readability-accuracy tension compared to a human benchmark. Using a comparative analysis of the instruction-tuned Mistral-Small 3 24B and the reasoning-augmented QWen2.5 32B, we identify an architectural advantage in the instruction-tuned LLM. Mistral exhibits a tempered lexical simplification strategy that enhances readability across a suite of metrics while preserving human-level discourse with a BERTScore of 0.91. QWen also attains enhanced readability performance and a reasonable BERTScore of 0.89, but its operational strategy shows a disconnect in balancing between readability and accuracy. Additionally, a comprehensive correlation analysis of a suite of 21 metrics spanning readability, discourse fidelity, content safety, and underlying distributional measures for mechanistic insights, confirms strong functional redundancies, and informs metric selection and domain adaptation for text simplification.
Authors: Jiarui Liu, Kaustubh Dhole, Yingheng Wang, Haoyang Wen, Sarah Zhang, Haitao Mao, Gaotang Li, Neeraj Varshney, Jingguo Liu, Xiaoman Pan
Abstract: Deductive reasoning is the process of deriving conclusions strictly from the given premises, without relying on external knowledge. We define honesty in this setting as a model's ability to respond only when the conclusion is logically entailed by the premises, and to abstain otherwise. However, current language models often fail to reason honestly, producing unwarranted answers when the input is insufficient. To study this challenge, we formulate honest deductive reasoning as multi-step tasks where models must either derive the correct conclusion or abstain. We curate two datasets from graph structures, one for linear algebra and one for logical inference, and introduce unanswerable cases by randomly perturbing an edge in half of the instances. We find that prompting and existing training methods, including GRPO with or without supervised fine-tuning initialization, struggle on these tasks. In particular, GRPO optimize only for final task outcomes, leaving models vulnerable to collapse when negative rewards dominate early training. To address this, we propose \methodname{}, a reinforcement learning method that injects ground truth trajectories into rollouts, preventing early training collapse. Our results demonstrate that this method stabilizes learning and significantly improves the overall reasoning performance, underscoring the importance of training dynamics for enabling honest deductive reasoning in language models.
Authors: Vikram Aggarwal, Jay Kulkarni, Aditi Mascarenhas, Aakriti Narang, Siddarth Raman, Ajay Shah, Susan Thomas
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in text comprehension, but their ability to process complex, hierarchical tabular data remains underexplored. We present a novel approach to extracting structured data from multi-page government fiscal documents using LLM-based techniques. Applied to annual fiscal documents from the State of Karnataka in India (200+ pages), our method achieves high accuracy through a multi-stage pipeline that leverages domain knowledge, sequential context, and algorithmic validation. A large challenge with traditional OCR methods is the inability to verify the accurate extraction of numbers. When applied to fiscal data, the inherent structure of fiscal tables, with totals at each level of the hierarchy, allows for robust internal validation of the extracted data. We use these hierarchical relationships to create multi-level validation checks. We demonstrate that LLMs can read tables and also process document-specific structural hierarchies, offering a scalable process for converting PDF-based fiscal disclosures into research-ready databases. Our implementation shows promise for broader applications across developing country contexts.
Authors: Yunxin Li, Xinyu Chen, Shenyuan Jiang, Haoyuan Shi, Zhenyu Liu, Xuanyu Zhang, Nanhao Deng, Zhenran Xu, Yicheng Ma, Meishan Zhang, Baotian Hu, Min Zhang
Abstract: We present Uni-MoE 2.0 from the Lychee family. As a fully open-source omnimodal large model (OLM), it substantially advances Lychee's Uni-MoE series in language-centric multimodal understanding, reasoning, and generating. Based on the dense LLM, we build Uni-MoE-2.0-Omni from scratch through three core contributions: dynamic-capacity Mixture-of-Experts (MoE) design, a progressive training strategy enhanced with an iterative reinforcement strategy, and a carefully curated multimodal data matching technique. It is capable of omnimodal understanding, as well as generating images, text, and speech. Architecturally, our new MoE framework balances computational efficiency and capability for 10 cross-modal inputs using shared, routed, and null experts, while our Omni-Modality 3D RoPE ensures spatio-temporal cross-modality alignment in the self-attention layer. For training, following cross-modal pretraining, we use a progressive supervised fine-tuning strategy that activates modality-specific experts and is enhanced by balanced data composition and an iterative GSPO-DPO method to stabilise RL training and improve reasoning. Data-wise, the base model, trained on approximately 75B tokens of open-source multimodal data, is equipped with special speech and image generation tokens, allowing it to learn these generative tasks by conditioning its outputs on linguistic cues. Extensive evaluation across 85 benchmarks demonstrates that our model achieves SOTA or highly competitive performance against leading OLMs, surpassing Qwen2.5-Omni (trained with 1.2T tokens) on over 50 of 76 benchmarks. Key strengths include video understanding (+7% avg. of 8), omnimodallity understanding (+7% avg. of 4), and audiovisual reasoning (+4%). It also advances long-form speech processing (reducing WER by 4.2%) and leads in low-level image processing and controllable generation across 5 metrics.
Authors: Serge Gladkoff, Lifeng Han, Katerina Gasova
Abstract: Analytic Translation Quality Evaluation (TQE), based on Multidimensional Quality Metrics (MQM), traditionally uses a linear error-to-penalty scale calibrated to a reference sample of 1000-2000 words. However, linear extrapolation biases judgment on samples of different sizes, over-penalizing short samples and under-penalizing long ones, producing misalignment with expert intuition. Building on the Multi-Range framework, this paper presents a calibrated, non-linear scoring model that better reflects how human content consumers perceive translation quality across samples of varying length. Empirical data from three large-scale enterprise environments shows that acceptable error counts grow logarithmically, not linearly, with sample size. Psychophysical and cognitive evidence, including the Weber-Fechner law and Cognitive Load Theory, supports this premise by explaining why the perceptual impact of additional errors diminishes while the cognitive burden grows with scale. We propose a two-parameter model E(x) = a * ln(1 + b * x), a, b > 0, anchored to a reference tolerance and calibrated from two tolerance points using a one-dimensional root-finding step. The model yields an explicit interval within which the linear approximation stays within +/-20 percent relative error and integrates into existing evaluation workflows with only a dynamic tolerance function added. The approach improves interpretability, fairness, and inter-rater reliability across both human and AI-generated translations. By operationalizing a perceptually valid scoring paradigm, it advances translation quality evaluation toward more accurate and scalable assessment. The model also provides a stronger basis for AI-based document-level evaluation aligned with human judgment. Implementation considerations for CAT/LQA systems and implications for human and AI-generated text evaluation are discussed.
Authors: Hao Wang, Yuanfeng Song, Xiaoming Yin, Xing Chen
Abstract: Text-to-SQL datasets are essential for training and evaluating text-to-SQL models, but existing datasets often suffer from limited coverage and fail to capture the diversity of real-world applications. To address this, we propose a novel taxonomy for text-to-SQL classification based on dimensions including core intents, statement types, syntax structures, and key actions. Using this taxonomy, we evaluate widely used public text-to-SQL datasets (e.g., Spider and Bird) and reveal limitations in their coverage and diversity. We then introduce a taxonomy-guided dataset synthesis pipeline, yielding a new dataset named SQL-Synth. This approach combines the taxonomy with Large Language Models (LLMs) to ensure the dataset reflects the breadth and complexity of real-world text-to-SQL applications. Extensive analysis and experimental results validate the effectiveness of our taxonomy, as SQL-Synth exhibits greater diversity and coverage compared to existing benchmarks. Moreover, we uncover that existing LLMs typically fall short in adequately capturing the full range of scenarios, resulting in limited performance on SQL-Synth. However, fine-tuning can substantially improve their performance in these scenarios. The proposed taxonomy has significant potential impact, as it not only enables comprehensive analysis of datasets and the performance of different LLMs, but also guides the construction of training data for LLMs.
Authors: Hourun Zhu, Yang Gao, Wenlong Fei, Jiawei Li, Huashan Sun
Abstract: Large reasoning models have demonstrated remarkable performance on complex reasoning tasks, yet the excessive length of their chain-of-thought outputs remains a major practical bottleneck due to high computation cost and poor deployability. Existing compression methods have achieved partial success but overlook a crucial phenomenon in the training process -- the entropy conflict. During compression training, entropy decreases, leading to shorter reasoning but limited exploration, while accuracy-oriented objectives increase entropy, lengthening reasoning chains. This can cause the model to get stuck in a local dilemma. Our analysis further reveals the origin of the entropy conflict: many high-entropy tokens are logical connectors that receive larger gradients and are encouraged under the performance objective, while the compression objective simultaneously penalizes these potentially redundant connectors. This opposing pressure creates a direct source of entropy conflict. To address these issues, we adopt an entropy-guided training framework. As entropy descends, the model is guided toward efficient reasoning by encouraging concise thought steps; as entropy rises, exploration is reinforced under the compact reasoning mode to improve robustness. Experiments on six mathematical benchmarks show that our method compresses reasoning length to 20% of the original while maintaining or even surpassing baseline accuracy. Code and models will be released publicly.
Authors: Raha Aghaei, Ali A. Kiaei, Mahnaz Boush, Mahan Rofoosheh, Mohammad Zavvar
Abstract: This study analyzes the multiple functions of Large Language Models (LLMs) in transforming research and development (R&D) processes. By automating knowledge discovery, boosting hypothesis creation, integrating transdisciplinary insights, and enabling cooperation within innovation ecosystems, LLMs dramatically improve the efficiency and effectiveness of research processes. Through extensive analysis of scientific literature, patent databases, and experimental data, these models enable more flexible and informed R&D workflows, ultimately accelerating innovation cycles and lowering time-to-market for breakthrough ideas.
Authors: Pei-Fu Guo, Yun-Da Tsai, Chun-Chia Hsu, Kai-Xin Chen, Ya-An Tsai, Kai-Wei Chang, Nanyun Peng, Mi-Yen Yeh, Shou-De Lin
Abstract: Evaluating cross-lingual knowledge transfer in large language models is challenging, as correct answers in a target language may arise either from genuine transfer or from prior exposure during pre-training. We present LiveCLKTBench, an automated generation pipeline specifically designed to isolate and measure cross-lingual knowledge transfer. Our pipeline identifies self-contained, time-sensitive knowledge entities from real-world domains, filters them based on temporal occurrence, and verifies them against the model's knowledge. The documents of these valid entities are then used to generate factual questions, which are translated into multiple languages to evaluate transferability across linguistic boundaries. Using LiveCLKTBench, we evaluate several LLMs across five languages and observe that cross-lingual transfer is strongly influenced by linguistic distance and often asymmetric across language directions. While larger models improve transfer, the gains diminish with scale and vary across domains. These findings provide new insights into multilingual transfer and demonstrate the value of LiveCLKTBench as a reliable benchmark for future research.
Authors: Elias Lumer, Alex Cardenas, Matt Melich, Myles Mason, Sara Dieter, Vamse Kumar Subbiah, Pradeep Honaganahalli Basavaraju, Roberto Hernandez
Abstract: Recent advancements in Retrieval-Augmented Generation (RAG) have enabled Large Language Models (LLMs) to access multimodal knowledge bases containing both text and visual information such as charts, diagrams, and tables in financial documents. However, existing multimodal RAG systems rely on LLM-based summarization to convert images into text during preprocessing, storing only text representations in vector databases, which causes loss of contextual information and visual details critical for downstream retrieval and question answering. To address this limitation, we present a comprehensive comparative analysis of two retrieval approaches for multimodal RAG systems, including text-based chunk retrieval (where images are summarized into text before embedding) and direct multimodal embedding retrieval (where images are stored natively in the vector space). We evaluate all three approaches across 6 LLM models and a two multi-modal embedding models on a newly created financial earnings call benchmark comprising 40 question-answer pairs, each paired with 2 documents (1 image and 1 text chunk). Experimental results demonstrate that direct multimodal embedding retrieval significantly outperforms LLM-summary-based approaches, achieving absolute improvements of 13% in mean average precision (mAP@5) and 11% in normalized discounted cumulative gain. These gains correspond to relative improvements of 32% in mAP@5 and 20% in nDCG@5, providing stronger evidence of their practical impact. We additionally find that direct multimodal retrieval produces more accurate and factually consistent answers as measured by LLM-as-a-judge pairwise comparisons. We demonstrate that LLM summarization introduces information loss during preprocessing, whereas direct multimodal embeddings preserve visual context for retrieval and inference.
Authors: Yuetian Zou, Hanlei Zhang, Hua Xu, Songze Li, Long Xiao
Abstract: Textual open intent classification is crucial for real-world dialogue systems, enabling robust detection of unknown user intents without prior knowledge and contributing to the robustness of the system. While adaptive decision boundary methods have shown great potential by eliminating manual threshold tuning, existing approaches assume isotropic distributions of known classes, restricting boundaries to balls and overlooking distributional variance along different directions. To address this limitation, we propose EliDecide, a novel method that learns ellipsoid decision boundaries with varying scales along different feature directions. First, we employ supervised contrastive learning to obtain a discriminative feature space for known samples. Second, we apply learnable matrices to parameterize ellipsoids as the boundaries of each known class, offering greater flexibility than spherical boundaries defined solely by centers and radii. Third, we optimize the boundaries via a novelly designed dual loss function that balances empirical and open-space risks: expanding boundaries to cover known samples while contracting them against synthesized pseudo-open samples. Our method achieves state-of-the-art performance on multiple text intent benchmarks and further on a question classification dataset. The flexibility of the ellipsoids demonstrates superior open intent detection capability and strong potential for generalization to more text classification tasks in diverse complex open-world scenarios.
Authors: Yesheng Liu, Hao Li, Haiyu Xu, Baoqi Pei, Jiahao Wang, Mingxuan Zhao, Jingshu Zheng, Zheqi He, JG Yao, Bowen Qin, Xi Yang, Jiajun Zhang
Abstract: Multiple-choice question answering (MCQA) has been a popular format for evaluating and reinforcement fine-tuning (RFT) of modern multimodal language models. Its constrained output format allows for simplified, deterministic automatic verification. However, we find that the options may leak exploitable signals, which makes the accuracy metrics unreliable for indicating real capabilities and encourages explicit or implicit answer guessing behaviors during RFT. We propose ReVeL (Rewrite and Verify by LLM), a framework that rewrites multiple-choice questions into open-form questions while keeping answers verifiable whenever possible. The framework categorizes questions according to different answer types, apply different rewriting and verification schemes, respectively. When applied for RFT, we converted 20k MCQA examples and use GRPO to finetune Qwen2.5-VL models. Models trained on ReVeL-OpenQA match MCQA accuracy on multiple-choice benchmarks and improve OpenQA accuracy by about six percentage points, indicating better data efficiency and more robust reward signals than MCQA-based training. When used for evaluation, ReVeL also reveals up to 20 percentage points of score inflation in MCQA benchmarks (relative to OpenQA), improves judging accuracy, and reduces both cost and latency. We will release code and data publicly.
Authors: Sam Earle, Samyak Parajuli, Andrzej Banburski-Fahey
Abstract: Coding assistants are increasingly leveraged in game design, both generating code and making high-level plans. To what degree can these tools align with developer workflows, and what new modes of human-computer interaction can emerge from their use? We present DreamGarden, an AI system capable of assisting with the development of diverse game environments in Unreal Engine. At the core of our method is an LLM-driven planner, capable of breaking down a single, high-level prompt -- a dream, memory, or imagined scenario provided by a human user -- into a hierarchical action plan, which is then distributed across specialized submodules facilitating concrete implementation. This system is presented to the user as a garden of plans and actions, both growing independently and responding to user intervention via seed prompts, pruning, and feedback. Through a user study, we explore design implications of this system, charting courses for future work in semi-autonomous assistants and open-ended simulation design.
Authors: Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, Haoyu Wang
Abstract: Fine-tuning large pre-trained foundation models often yields excellent downstream performance but is prohibitively expensive when updating all parameters. Parameter-efficient fine-tuning (PEFT) methods such as LoRA alleviate this by introducing lightweight update modules, yet they commonly rely on weight-agnostic linear approximations, limiting their expressiveness. In this work, we propose PEANuT, a novel PEFT framework that introduces weight-aware neural tweakers, compact neural modules that generate task-adaptive updates conditioned on frozen pre-trained weights. PEANuT provides a flexible yet efficient way to capture complex update patterns without full model tuning. We theoretically show that PEANuT achieves equivalent or greater expressivity than existing linear PEFT methods with comparable or fewer parameters. Extensive experiments across four benchmarks with over twenty datasets demonstrate that PEANuT consistently outperforms strong baselines in both NLP and vision tasks, while maintaining low computational overhead.
Authors: Addison Kristanto Julistiono, Davoud Ataee Tarzanagh, Navid Azizan
Abstract: Attention mechanisms have revolutionized several domains of artificial intelligence, such as natural language processing and computer vision, by enabling models to selectively focus on relevant parts of the input data. While recent work has characterized the optimization dynamics of gradient descent (GD) in attention-based models and the structural properties of its preferred solutions, less is known about more general optimization algorithms such as mirror descent (MD). In this paper, we investigate the convergence properties and implicit biases of a family of MD algorithms tailored for softmax attention mechanisms, with the potential function chosen as the $p$-th power of the $\ell_p$-norm. Specifically, we show that these algorithms converge in direction to a generalized hard-margin SVM with an $\ell_p$-norm objective when applied to a classification problem using a softmax attention model. Notably, our theoretical results reveal that the convergence rate is comparable to that of traditional GD in simpler models, despite the highly nonlinear and nonconvex nature of the present problem. Additionally, we delve into the joint optimization dynamics of the key-query matrix and the decoder, establishing conditions under which this complex joint optimization converges to their respective hard-margin SVM solutions. Lastly, our numerical experiments on real data demonstrate that MD algorithms improve generalization over standard GD and excel in optimal token selection.
Authors: Yueqian Wang, Xiaojun Meng, Yuxuan Wang, Jianxin Liang, Jiansheng Wei, Huishuai Zhang, Dongyan Zhao
Abstract: Recent researches on video large language models (VideoLLM) predominantly focus on model architectures and training datasets, leaving the interaction format between the user and the model under-explored. In existing works, users often interact with VideoLLMs by using the entire video and a query as input, after which the model generates a response. This interaction format constrains the application of VideoLLMs in scenarios such as live-streaming comprehension where videos do not end and responses are required in a real-time manner, and also results in unsatisfactory performance on time-sensitive tasks that requires localizing video segments. In this paper, we focus on a video-text duet interaction format. This interaction format is characterized by the continuous playback of the video, and both the user and the model can insert their text messages at any position during the video playback. When a text message ends, the video continues to play, akin to the alternative of two performers in a duet. We construct MMDuetIT, a video-text training dataset designed to adapt VideoLLMs to video-text duet interaction format. We also introduce the Multi-Answer Grounded Video Question Answering (MAGQA) task to benchmark the real-time response ability of VideoLLMs. Trained on MMDuetIT, MMDuet demonstrates that adopting the video-text duet interaction format enables the model to achieve significant improvements in various time-sensitive tasks (76% CIDEr on YouCook2 dense video captioning, 90\% mAP on QVHighlights highlight detection and 25% R@0.5 on Charades-STA temporal video grounding) with minimal training efforts, and also enable VideoLLMs to reply in a real-time manner as the video plays.
Authors: Xiaoning Dong, Wenbo Hu, Wei Xu, Tianxing He
Abstract: Large language models (LLMs) have made significant advancements across various tasks, but their safety alignment remain a major concern. Exploring jailbreak prompts can expose LLMs' vulnerabilities and guide efforts to secure them. Existing methods primarily design sophisticated instructions for the LLM to follow, or rely on multiple iterations, which could hinder the performance and efficiency of jailbreaks. In this work, we propose a novel jailbreak paradigm, Simple Assistive Task Linkage (SATA), which can effectively circumvent LLM safeguards and elicit harmful responses. Specifically, SATA first masks harmful keywords within a malicious query to generate a relatively benign query containing one or multiple [MASK] special tokens. It then employs a simple assistive task such as a masked language model task or an element lookup by position task to encode the semantics of the masked keywords. Finally, SATA links the assistive task with the masked query to jointly perform the jailbreak. Extensive experiments show that SATA achieves state-of-the-art performance and outperforms baselines by a large margin. Specifically, on AdvBench dataset, with mask language model (MLM) assistive task, SATA achieves an overall attack success rate (ASR) of 85% and harmful score (HS) of 4.57, and with element lookup by position (ELP) assistive task, SATA attains an overall ASR of 76% and HS of 4.43.
Authors: Ting Bai, Yue Yu, Le Huang, Zenan Xu, Chuan Shi
Abstract: The sparse Mixture-of-Experts (MoE) architecture of large language models (LLMs) confronts an inherent issue of load imbalance arising from the simplistic linear router strategy, which ultimately causes the instability and inefficient learning of LLMs. To address this challenge, we introduce a novel MoE graph-based framework $\textbf{GMoE}$, aimed at enhancing the collaboration among multiple experts. In GMoE, a graph router function is designed to capture the collaboration signals among experts. This enables all experts to dynamically allocate information derived from input data by sharing information with their neighboring experts. Moreover, we put forward two coordination strategies in GMoE: the $\textit{Poisson distribution-based distinction strategy}$ and the $\textit{Normal distribution-based balance strategy}$, to further release the capacity of each expert and increase the model stability in the fine-tuning of LLMs. Specifically, we leverage a parameter-efficient fine-tuning technique, i.e., Low-Rank Adaptation (LoRA), to implement the graph MoE architecture. Extensive experiments on four real-world benchmark datasets demonstrate the effectiveness of GMoE, showing the benefits of facilitating collaborations of multiple experts in LLM fine-tuning. The code of experimental implementation is available at https://github.com/BAI-LAB/GMoE
Authors: Hadi Hosseini, Samarth Khanna
Abstract: The growing interest in employing large language models (LLMs) for decision-making in social and economic contexts has raised questions about their potential to function as agents in these domains. A significant number of societal problems involve the distribution of resources, where fairness, along with economic efficiency, play a critical role in the desirability of outcomes. In this paper, we examine whether LLM responses adhere to fundamental fairness concepts such as equitability, envy-freeness, and Rawlsian maximin, and investigate their alignment with human preferences. We evaluate the performance of several LLMs, providing a comparative benchmark of their ability to reflect these measures. Our results demonstrate a lack of alignment between current LLM responses and human distributional preferences. Moreover, LLMs are unable to utilize money as a transferable resource to mitigate inequality. Nonetheless, we demonstrate a stark contrast when (some) LLMs are tasked with selecting from a predefined menu of options rather than generating one. In addition, we analyze the robustness of LLM responses to variations in semantic factors (e.g., intentions or personas) or non-semantic prompting changes (e.g., templates or orderings). Finally, we highlight potential strategies aimed at enhancing the alignment of LLM behavior with well-established fairness concepts.
Authors: Shane Bergsma, Nolan Dey, Gurpreet Gosal, Gavia Gray, Daria Soboleva, Joel Hestness
Abstract: LLMs are commonly trained with a learning rate (LR) warmup, followed by cosine decay to 10% of the maximum (10x decay). In a large-scale empirical study, we show that under an optimal peak LR, a simple linear decay-to-zero (D2Z) schedule consistently outperforms other schedules when training at compute-optimal dataset sizes. D2Z is superior across a range of model sizes, batch sizes, datasets, and vocabularies. Benefits increase as dataset size increases. Leveraging a novel interpretation of AdamW as an exponential moving average of weight updates, we show how linear D2Z optimally balances the demands of early training (moving away from initial conditions) and late training (averaging over more updates in order to mitigate gradient noise). In experiments, a 610M-parameter model trained for 80 tokens-per-parameter (TPP) using D2Z achieves lower loss than when trained for 200 TPP using 10x decay, corresponding to an astonishing 60% compute savings. Models such as Llama2-7B, trained for 286 TPP with 10x decay, could likely have saved a majority of compute by training with D2Z.
Authors: Jules Viennot, Guillaume Baudart, Emilio Jes\`us Gallego Arias, Marc Lelarge
Abstract: In this work, we conduct an experiment using state-of-the-art LLMs to translate MiniF2F into Rocq. The translation task focuses on generating a Rocq theorem based on three sources: a natural language description, the Lean formalization, and the Isabelle formalization. We conducted our experiment in 3 stages of increasing complexity, from basic one-shot prompting to multi-turn conversations that incorporate feedback from unsuccessful attempts. At each stage, we perform multiple rounds of translation using increasingly advanced models: GPT-4o mini, Claude 3.5 Sonnet, o1 mini, and o1. We successfully translated 478 out of 488 theorems. The dataset is opensource: https://github.com/LLM4Rocq/miniF2F-rocq.
Authors: Aske Plaat, Max van Duijn, Niki van Stein, Mike Preuss, Peter van der Putten, Kees Joost Batenburg
Abstract: Background: There is great interest in agentic LLMs, large language models that act as agents. Objectives: We review the growing body of work in this area and provide a research agenda. Methods: Agentic LLMs are LLMs that (1) reason, (2) act, and (3) interact. We organize the literature according to these three categories. Results: The research in the first category focuses on reasoning, reflection, and retrieval, aiming to improve decision making; the second category focuses on action models, robots, and tools, aiming for agents that act as useful assistants; the third category focuses on multi-agent systems, aiming for collaborative task solving and simulating interaction to study emergent social behavior. We find that works mutually benefit from results in other categories: retrieval enables tool use, reflection improves multi-agent collaboration, and reasoning benefits all categories. Conclusions: We discuss applications of agentic LLMs and provide an agenda for further research. Important applications are in medical diagnosis, logistics and financial market analysis. Meanwhile, self-reflective agents playing roles and interacting with one another augment the process of scientific research itself. Further, agentic LLMs provide a solution for the problem of LLMs running out of training data: inference-time behavior generates new training states, such that LLMs can keep learning without needing ever larger datasets. We note that there is risk associated with LLM assistants taking action in the real world-safety, liability and security are open problems-while agentic LLMs are also likely to benefit society.
Authors: Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, Gao Huang
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model's reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.
Authors: Zheng Wu, Pengzhou Cheng, Zongru Wu, Lingzhong Dong, Zhuosheng Zhang
Abstract: Graphical user interface (GUI) agents have recently emerged as an intriguing paradigm for human-computer interaction, capable of automatically executing user instructions to operate intelligent terminal devices. However, when encountering out-of-distribution (OOD) instructions that violate environmental constraints or exceed the current capabilities of agents, GUI agents may suffer task breakdowns or even pose security threats. Therefore, effective OOD detection for GUI agents is essential. Traditional OOD detection methods perform suboptimally in this domain due to the complex embedding space and evolving GUI environments. In this work, we observe that the in-distribution input semantic space of GUI agents exhibits a clustering pattern with respect to the distance from the centroid. Based on the finding, we propose GEM, a novel method based on fitting a Gaussian mixture model over input embedding distances extracted from the GUI agent that reflect its capability boundary. Evaluated on eight datasets spanning smartphones, computers, and web browsers, our method achieves an average accuracy improvement of 23.70\% over the best-performing baseline while only increasing training time by 4.9\% and testing time by 6.5\%. We also experimentally demonstrate that GEM can improve the step-wise success rate by 9.40\% by requesting assistance from the cloud model when encountering OOD samples. Analysis verifies the generalization ability of our method through experiments on nine different backbones. The codes are available at https://github.com/Wuzheng02/GEM-OODforGUIagents.
Authors: Shane Bergsma, Nolan Dey, Gurpreet Gosal, Gavia Gray, Daria Soboleva, Joel Hestness
Abstract: Efficient LLM pre-training requires well-tuned hyperparameters (HPs), including learning rate $\eta$ and weight decay $\lambda$. We study scaling laws for HPs: formulas for how to scale HPs as we scale model size N, dataset size D, and batch size B. Recent work suggests the AdamW timescale, $\tau = B/(\eta \lambda D)$, should remain constant across training settings, and we verify the implication that optimal $\lambda$ scales linearly with B, for a fixed N and D. However, as N and D scale, we show optimal $\tau$ obeys a precise power law in the tokens-per-parameter ratio, D/N. This law thus provides a method to accurately predict $\lambda$opt in advance of large-scale training. We also study scaling laws for optimal batch size Bopt (the B enabling lowest loss at a given N,D) and critical batch size Bcrit (the B beyond which further data parallelism becomes ineffective). In contrast to prior work, we find both Bopt and Bcrit scale as power laws in D, independent of model size, N. Finally, we analyze how these findings inform the real-world selection of Pareto-optimal N and D under dual training time and compute objectives. All experiments were run on Cerebras CS-3 systems.
Authors: Jintao Tong, Wenwei Jin, Pengda Qin, Anqi Li, Yixiong Zou, Yuhong Li, Yuhua Li, Ruixuan Li
Abstract: Large vision-language models (LVLMs) excel at multimodal understanding but suffer from high computational costs due to redundant vision tokens. Existing pruning methods typically rely on single-layer attention scores to rank and prune redundant visual tokens to solve this inefficiency. However, as the interaction between tokens and layers is complicated, this raises a basic question: Is such a simple single-layer criterion sufficient to identify redundancy? To answer this question, we rethink the emergence of redundant visual tokens from a fundamental perspective: information flow, which models the interaction between tokens and layers by capturing how information moves between tokens across layers. We find (1) the CLS token acts as an information relay, which can simplify the complicated flow analysis; (2) the redundancy emerges progressively and dynamically via layer-wise attention concentration; and (3) relying solely on attention scores from single layers can lead to contradictory redundancy identification. Based on this, we propose FlowCut, an information-flow-aware pruning framework, mitigating the insufficiency of the current criterion for identifying redundant tokens and better aligning with the model's inherent behaviors. Extensive experiments show that FlowCut achieves superior results, outperforming SoTA by 1.6% on LLaVA-1.5-7B with 88.9% token reduction, and by 4.3% on LLaVA-NeXT-7B with 94.4% reduction, delivering 3.2x speed-up in the prefilling stage. Our code is available at https://github.com/TungChintao/FlowCut
Authors: Shuai Wang, Zhenhua Liu, Jiaheng Wei, Xuanwu Yin, Dong Li, Emad Barsoum
Abstract: We present Athena-PRM, a multimodal process reward model (PRM) designed to evaluate the reward score for each step in solving complex reasoning problems. Developing high-performance PRMs typically demands significant time and financial investment, primarily due to the necessity for step-level annotations of reasoning steps. Conventional automated labeling methods, such as Monte Carlo estimation, often produce noisy labels and incur substantial computational costs. To efficiently generate high-quality process-labeled data, we propose leveraging prediction consistency between weak and strong completers as a criterion for identifying reliable process labels. Remarkably, Athena-PRM demonstrates outstanding effectiveness across various scenarios and benchmarks with just 5,000 samples. Furthermore, we also develop two effective strategies to improve the performance of PRMs: ORM initialization and up-sampling for negative data. We validate our approach in three specific scenarios: verification for test time scaling, direct evaluation of reasoning step correctness, and reward ranked fine-tuning. Our Athena-PRM consistently achieves superior performance across multiple benchmarks and scenarios. Notably, when using Qwen2.5-VL-7B as the policy model, Athena-PRM enhances performance by 10.2 points on WeMath and 7.1 points on MathVista for test time scaling. Furthermore, Athena-PRM sets the state-of-the-art (SoTA) results in VisualProcessBench and outperforms the previous SoTA by 3.9 F1-score, showcasing its robust capability to accurately assess the correctness of the reasoning step. Additionally, utilizing Athena-PRM as the reward model, we develop Athena-7B with reward ranked fine-tuning and outperforms baseline with a significant margin on five benchmarks.
Authors: Longzhen Han, Awes Mubarak, Almas Baimagambetov, Nikolaos Polatidis, Thar Baker
Abstract: Multimodal Generative Models (MGMs) have rapidly evolved beyond text generation, now spanning diverse output modalities including images, music, video, human motion, and 3D objects, by integrating language with other sensory modalities under unified architectures. This survey categorises six primary generative modalities and examines how foundational techniques, namely Self-Supervised Learning (SSL), Mixture of Experts (MoE), Reinforcement Learning from Human Feedback (RLHF), and Chain-of-Thought (CoT) prompting, enable cross-modal capabilities. We analyze key models, architectural trends, and emergent cross-modal synergies, while highlighting transferable techniques and unresolved challenges. Building on a common taxonomy of models and training recipes, we propose a unified evaluation framework centred on faithfulness, compositionality, and robustness, and synthesise evidence from benchmarks and human studies across modalities. We further analyse trustworthiness, safety, and ethical risks, including multimodal bias, privacy leakage, and the misuse of high-fidelity media generation for deepfakes, disinformation, and copyright infringement in music and 3D assets, together with emerging mitigation strategies. Finally, we discuss how architectural trends, evaluation protocols, and governance mechanisms can be co-designed to close current capability and safety gaps, outlining critical paths toward more general-purpose, controllable, and accountable multimodal generative systems.
Authors: Junxian Li, Beining Xu, Simin Chen, Jiatong Li, Jingdi Lei, Haodong Zhao, Di Zhang
Abstract: Recent advances in vision-language models (VLMs) have significantly enhanced the visual grounding task, which involves locating objects in an image based on natural language queries. Despite these advancements, the security of VLM-based grounding systems has not been thoroughly investigated. This paper reveals a novel and realistic vulnerability: the first multi-target backdoor attack on VLM-based visual grounding. Unlike prior attacks that rely on static triggers or fixed targets, we propose IAG, a method that dynamically generates input-aware, text-guided triggers conditioned on any specified target object description to execute the attack. This is achieved through a text-conditioned UNet that embeds imperceptible target semantic cues into visual inputs while preserving normal grounding performance on benign samples. We further develop a joint training objective that balances language capability with perceptual reconstruction to ensure imperceptibility, effectiveness, and stealth. Extensive experiments on multiple VLMs (e.g., LLaVA, InternVL, Ferret) and benchmarks (RefCOCO, RefCOCO+, RefCOCOg, Flickr30k Entities, and ShowUI) demonstrate that IAG achieves the best ASRs compared with other baselines on almost all settings without compromising clean accuracy, maintaining robustness against existing defenses, and exhibiting transferability across datasets and models. These findings underscore critical security risks in grounding-capable VLMs and highlight the need for further research on trustworthy multimodal understanding.
Authors: Maxime Heuillet, Yufei Cui, Boxing Chen, Audrey Durand, Prasanna Parthasarathi
Abstract: Advanced reasoning in LLMs on challenging domains like mathematical reasoning can be tackled using verifiable rewards based reinforced fine-tuning (ReFT). In standard ReFT frameworks, a behavior model generates multiple completions with answers per problem, for the answer to be then scored by a reward function. While such RL post-training methods demonstrate significant performance improvements across challenging reasoning domains, the computational cost of generating completions during training with multiple inference steps makes the training cost non-trivial. To address this, we draw inspiration from off-policy RL, and speculative decoding to introduce a novel ReFT framework, dubbed Nested-ReFT, where a subset of layers of the target model acts as the behavior model to generate off-policy completions during training. The behavior model configured with dynamic layer skipping per batch during training decreases the inference cost compared to the standard ReFT frameworks. Our theoretical analysis shows that Nested-ReFT yields unbiased gradient estimates with controlled variance. Our empirical analysis demonstrates improved computational efficiency measured as tokens/sec across multiple math reasoning benchmarks and model sizes. Additionally, we explore three variants of bias mitigation to minimize the off-policyness in the gradient updates that allows for maintaining performance that matches the baseline ReFT performance.
Authors: Mingxuan Cui, Yilan Jiang, Duo Zhou, Cheng Qian, Yuji Zhang, Qiong Wang
Abstract: Drug shortages pose critical risks to patient care and healthcare systems worldwide, yet the effectiveness of regulatory interventions remains poorly understood due to information asymmetries in pharmaceutical supply chains. We propose ShortageSim, which addresses this challenge by providing the first simulation framework that evaluates the impact of regulatory interventions on competition dynamics under information asymmetry. Using Large Language Model (LLM)-based agents, the framework models the strategic decisions of drug manufacturers and institutional buyers in response to shortage alerts given by the regulatory agency. Unlike traditional game theory models that assume perfect rationality and complete information, \name simulates heterogeneous interpretations on regulatory announcements and the resulting decisions. Experiments on a self-processed dataset of historical shortage events show that \name reduces the resolution lag for production disruption cases by up to 84\%, achieving closer alignment to real-world trajectories than the zero-shot baseline. Our framework confirms the effect of regulatory alert in addressing shortages and introduces a new method for understanding competition in multi-stage environments under uncertainty. We open-source \name and a dataset of 2,925 FDA shortage events in https://github.com/Lemutisme/ShortageSim, providing a novel framework for future research on policy design and testing in supply chains under information asymmetry.
Authors: Xiao Li, Jiaqi Zhang, Shuxiang Zhang, Tianshui Chen, Liang Lin, Guangrun Wang
Abstract: While diffusion models excel at generating continuous data such as images, adapting them to discrete tasks has relied on indirect approaches that either operate in continuous embedding spaces or use token masking mechanisms, both of which deviate from modeling the true discrete data distribution that can be theoretically guaranteed by Tweedie's formula. We propose in-situ Tweedie Discrete Diffusion (TDD), a framework that performs diffusion guaranteed by Tweedie's formula directly within the discrete one-hot space, hence "in-situ." Unlike prior methods that diffuse continuous embeddings or mask tokens, TDD directly corrupts one-hot vectors with Gaussian noise and performs iterative denoising through a timestep-conditioned cross-entropy objective rather than mean-squared-error reconstruction. At each denoising step, the model predicts class probabilities, applies argmax to obtain discrete predictions, converts them to one-hot vectors, and feeds them into the next iteration with progressively reduced noise. This process naturally unifies discriminative classification and generative modeling under a single framework. Experiments demonstrate that TDD achieves strong performance on both image classification and text generation tasks, with extensive ablation studies confirming the effectiveness of each design component. Our work establishes a principled approach to discrete diffusion that preserves the core characteristics of diffusion models while operating natively in discrete space.
Authors: Liang Ye, Shengqin Chen, Jiazhu Dai
Abstract: The rapid progress of graph generation has raised new security concerns, particularly regarding backdoor vulnerabilities. While prior work has explored backdoor attacks in image diffusion and unconditional graph generation, conditional, especially text-guided graph generation remains largely unexamined. This paper proposes BadGraph, a backdoor attack method against latent diffusion models for text-guided graph generation. BadGraph leverages textual triggers to poison training data, covertly implanting backdoors that induce attacker-specified subgraphs during inference when triggers appear, while preserving normal performance on clean inputs. Extensive experiments on four benchmark datasets (PubChem, ChEBI-20, PCDes, MoMu) demonstrate the effectiveness and stealth of the attack: less than 10% poisoning rate can achieves 50% attack success rate, while 24% suffices for over 80% success rate, with negligible performance degradation on benign samples. Ablation studies further reveal that the backdoor is implanted during VAE and diffusion training rather than pretraining. These findings reveal the security vulnerabilities in latent diffusion models of text-guided graph generation, highlight the serious risks in models' applications such as drug discovery and underscore the need for robust defenses against the backdoor attack in such diffusion models.
Authors: Yuping Yan, Yuhan Xie, Yuanshuai Li, Yingchao Yu, Lingjuan Lyu, Yaochu Jin
Abstract: Since Multimodal Large Language Models (MLLMs) are increasingly being integrated into everyday tools and intelligent agents, growing concerns have arisen regarding their possible output of unsafe contents, ranging from toxic language and biased imagery to privacy violations and harmful misinformation. Current safety benchmarks remain highly limited in both modality coverage and performance evaluations, often neglecting the extensive landscape of content safety. In this work, we introduce OutSafe-Bench, the first most comprehensive content safety evaluation test suite designed for the multimodal era. OutSafe-Bench includes a large-scale dataset that spans four modalities, featuring over 18,000 bilingual (Chinese and English) text prompts, 4,500 images, 450 audio clips and 450 videos, all systematically annotated across nine critical content risk categories. In addition to the dataset, we introduce a Multidimensional Cross Risk Score (MCRS), a novel metric designed to model and assess overlapping and correlated content risks across different categories. To ensure fair and robust evaluation, we propose FairScore, an explainable automated multi-reviewer weighted aggregation framework. FairScore selects top-performing models as adaptive juries, thereby mitigating biases from single-model judgments and enhancing overall evaluation reliability. Our evaluation of nine state-of-the-art MLLMs reveals persistent and substantial safety vulnerabilities, underscoring the pressing need for robust safeguards in MLLMs.
Authors: Chunqiu Steven Xia, Zhe Wang, Yan Yang, Yuxiang Wei, Lingming Zhang
Abstract: Large Language Models (LLMs) are reshaping almost all industries, including software engineering. In recent years, a number of LLM agents have been proposed to solve real-world software problems. Such software agents are typically equipped with a suite of coding tools and can autonomously decide the next actions to form complete trajectories to solve end-to-end software tasks. While promising, they typically require dedicated design and may still be suboptimal, since it can be extremely challenging and costly to exhaust the entire agent scaffold design space. Recognizing that software agents are inherently software themselves that can be further refined/modified, researchers have proposed a number of self-improving software agents recently, including the Darwin-G\"odel Machine (DGM). Meanwhile, such self-improving agents require costly offline training on specific benchmarks and may not generalize well across different LLMs or benchmarks. In this paper, we propose Live-SWE-agent, the first live software agent that can autonomously and continuously evolve itself on-the-fly during runtime when solving real-world software problems. More specifically, Live-SWE-agent starts with the most basic agent scaffold with only access to bash tools (e.g., mini-SWE-agent), and autonomously evolves its own scaffold implementation while solving real-world software problems. Our evaluation on the widely studied SWE-bench Verified benchmark shows that LIVE-SWE-AGENT can achieve an impressive solve rate of 77.4% without test-time scaling, outperforming all existing software agents, including the best proprietary solution. Moreover, Live-SWE-agent outperforms state-of-the-art manually crafted software agents on the recent SWE-Bench Pro benchmark, achieving the best-known solve rate of 45.8%.
Authors: Xiaochuan Liu, Yuanfeng Song, Xiaoming Yin, Xing Chen
Abstract: In today's data-driven era, fully automated end-to-end data analytics, particularly insight discovery, is critical for discovering actionable insights that assist organizations in making effective decisions. With the rapid advancement of large language models (LLMs), LLM-driven agents have emerged as a promising paradigm for automating data analysis and insight discovery. However, existing data insight agents remain limited in several key aspects, often failing to deliver satisfactory results due to: (1) insufficient utilization of domain knowledge, (2) shallow analytical depth, and (3) error-prone code generation during insight generation. To address these issues, we propose DataSage, a novel multi-agent framework that incorporates three innovative features including external knowledge retrieval to enrich the analytical context, a multi-role debating mechanism to simulate diverse analytical perspectives and deepen analytical depth, and multi-path reasoning to improve the accuracy of the generated code and insights. Extensive experiments on InsightBench demonstrate that DataSage consistently outperforms existing data insight agents across all difficulty levels, offering an effective solution for automated data insight discovery.