Authors: Lance Calvin Lim Gamboa, Yue Feng, Mark Lee
Abstract: Pretrained multilingual models exhibit the same social bias as models processing English texts. This systematic review analyzes emerging research that extends bias evaluation and mitigation approaches into multilingual and non-English contexts. We examine these studies with respect to linguistic diversity, cultural awareness, and their choice of evaluation metrics and mitigation techniques. Our survey illuminates gaps in the field's dominant methodological design choices (e.g., preference for certain languages, scarcity of multilingual mitigation experiments) while cataloging common issues encountered and solutions implemented in adapting bias benchmarks across languages and cultures. Drawing from the implications of our findings, we chart directions for future research that can reinforce the multilingual bias literature's inclusivity, cross-cultural appropriateness, and alignment with state-of-the-art NLP advancements.
Authors: Mohammad Amini, Babak Ahmadi, Xiaomeng Xiong, Yilin Zhang, Christopher Qiao
Abstract: This study explores automatic generation (AIG) using language models to create multiple choice questions (MCQs) for morphological assessment, aiming to reduce the cost and inconsistency of manual test development. The study used a two-fold approach. First, we compared a fine-tuned medium model (Gemma, 2B) with a larger untuned one (GPT-3.5, 175B). Second, we evaluated seven structured prompting strategies, including zero-shot, few-shot, chain-of-thought, role-based, sequential, and combinations. Generated items were assessed using automated metrics and expert scoring across five dimensions. We also used GPT-4.1, trained on expert-rated samples, to simulate human scoring at scale. Results show that structured prompting, especially strategies combining chain-of-thought and sequential design, significantly improved Gemma's outputs. Gemma generally produced more construct-aligned and instructionally appropriate items than GPT-3.5's zero-shot responses, with prompt design playing a key role in mid-size model performance. This study demonstrates that structured prompting and efficient fine-tuning can enhance midsized models for AIG under limited data conditions. We highlight the value of combining automated metrics, expert judgment, and large-model simulation to ensure alignment with assessment goals. The proposed workflow offers a practical and scalable way to develop and validate language assessment items for K-12.
Authors: Andrei Mihai Albu, Giovanni Pollo, Alessio Burrello, Daniele Jahier Pagliari, Cristian Tesconi, Alessandra Neri, Dario Soldi, Fabio Autieri, Sara Vinco
Abstract: The growing complexity of cyber-physical systems, particularly in automotive applications, has increased the demand for efficient modeling and cross-domain co-simulation techniques. While SystemC Transaction-Level Modeling (TLM) enables effective hardware/software co-design, its limited interoperability with models from other engineering domains poses integration challenges. This paper presents a fully open-source methodology for integrating SystemC TLM models into Functional Mock-up Interface (FMI)-based co-simulation workflows. By encapsulating SystemC TLM components as FMI 3.0 Co Simulation Functional Mock-up Units (FMUs), the proposed approach facilitates seamless, standardized integration across heterogeneous simulation environments. We introduce a lightweight open-source toolchain, address key technical challenges such as time synchronization and data exchange, and demonstrate the feasibility and effectiveness of the integration through representative case studies.
Authors: Rikuto Kotoge, Mai Nishimura, Jiaxin Ma
Abstract: Reinforcement Learning has emerged as a post-training approach to elicit agentic RAG behaviors such as search and planning from language models. However, compact language models (e.g., 0.5B parameters) struggle due to poor reasoning ability, resulting in sparse rewards and unstable training. To overcome these difficulties, we propose Distillation-Guided Policy Optimization (DGPO), which addresses the challenges through cold-start initialization from teacher demonstrations and continuous teacher guidance during policy optimization. To systematically evaluate our approach, we introduce Agentic RAG Capabilities (ARC), a fine-grained metric analyzing reasoning, search coordination, and response synthesis. Comprehensive experiments demonstrate that DGPO enables compact models to achieve sophisticated agentic search behaviors, even outperforming the larger teacher model in some cases. DGPO makes agentic RAG feasible in computing resource-constrained environments.
Authors: Haibo Jin, Ruoxi Chen, Peiyan Zhang, Andy Zhou, Yang Zhang, Haohan Wang
Abstract: As Large Language Models become increasingly integral to various domains, their potential to generate harmful responses has prompted significant societal and regulatory concerns. In response, governments have issued ethics guidelines to promote the development of trustworthy AI. However, these guidelines are typically high-level demands for developers and testers, leaving a gap in translating them into actionable testing questions to verify LLM compliance. To address this challenge, we introduce GUARD (\textbf{G}uideline \textbf{U}pholding Test through \textbf{A}daptive \textbf{R}ole-play and Jailbreak \textbf{D}iagnostics), a testing method designed to operationalize guidelines into specific guideline-violating questions that assess LLM adherence. To implement this, GUARD uses automated generation of guideline-violating questions based on government-issued guidelines, thereby testing whether responses comply with these guidelines. When responses directly violate guidelines, GUARD reports inconsistencies. Furthermore, for responses that do not directly violate guidelines, GUARD integrates the concept of ``jailbreaks'' to diagnostics, named GUARD-JD, which creates scenarios that provoke unethical or guideline-violating responses, effectively identifying potential scenarios that could bypass built-in safety mechanisms. Our method finally culminates in a compliance report, delineating the extent of adherence and highlighting any violations. We have empirically validated the effectiveness of GUARD on seven LLMs, including Vicuna-13B, LongChat-7B, Llama2-7B, Llama-3-8B, GPT-3.5, GPT-4, GPT-4o, and Claude-3.7, by testing compliance under three government-issued guidelines and conducting jailbreak diagnostics. Additionally, GUARD-JD can transfer jailbreak diagnostics to vision-language models, demonstrating its usage in promoting reliable LLM-based applications.
Authors: Zhirui Chen, Wei Shen, Jiashui Huang, Ling Shao
Abstract: Despite significant progress, large language models (LLMs) still struggle with long contexts due to memory limitations and their inability to tackle complex and long-context tasks. Additionally, LLMs often suffer from a lack of transparency and are prone to producing hallucinations. To address these challenges, we propose \textbf{JERR}, a novel framework designed to enhance long-context comprehension via graph-based reasoning in LLMs. JERR integrates three key components: synopsis extraction, graph construction, and relational reasoning. First, synopsis is extracted by chunking text strategically, allowing the model to summarize and understand information more efficiently. Second, we build a directed acyclic graph (DAG) to resolve redundancy, ensuring logical consistency and clarity. Finally, we incorporate Monte Carlo Tree Search (MCTS) to help the model navigate complex reasoning paths, ensuring more accurate and interpretable outputs. This framework provides a novel solution that enables LLMs to handle extended contexts and complex reasoning tasks with improved reliability and transparency. Experimental results show that JERR consistently outperforms all baselines on the ROUGE and F1 metrics, achieving the highest scores on the LLM-Rater evaluation.
Authors: Yuyao Wang, Bowen Liu, Jianheng Tang, Nuo Chen, Yuhan Li, Qifan Zhang, Jia Li
Abstract: Reasoning Large Language Models (RLLMs) have recently achieved remarkable progress on complex reasoning tasks, largely enabled by their long chain-of-thought (Long CoT) capabilities. However, developing these Long CoT behaviors relies heavily on post-training with high-quality datasets, which are typically costly and human-curated (e.g., mathematics and code), leaving scalable alternatives unexplored. In this work, we introduce NP-hard (NPH) graph problems as a novel synthetic training corpus, as they inherently require deep reasoning, extensive exploration, and reflective strategies, which are core characteristics of Long CoT reasoning. Building on this insight, we develop a two-stage post-training framework: (i) Long CoT Supervised Fine-Tuning (SFT) on rejection-sampled NPH graph instances, which substantially enhances reasoning depth, and (ii) Reinforcement Learning (RL) with a fine-grained reward design, which sharpens reasoning efficiency. Our flagship model, Graph-R1-7B, demonstrates strong generalization across mathematics, coding, STEM, and logic, and surpasses QwQ-32B on NPH graph problems in both accuracy and reasoning efficiency. These results position NPH graph problems as an effective and scalable resource for advancing Long CoT reasoning in LLMs, opening a new frontier for LLM post-training. Our implementation is available at https://github.com/Graph-Reasoner/Graph-R1, with models and datasets hosted in our Hugging Face collection HKUST-DSAIL/Graph-R1.
Authors: Jivnesh Sandhan, Fei Cheng, Tushar Sandhan, Yugo Murawaki
Abstract: Psychometric tests, traditionally used to assess humans, are now being applied to Large Language Models (LLMs) to evaluate their behavioral traits. However, existing studies follow a context-free approach, answering each question in isolation to avoid contextual influence. We term this the Disney World test, an artificial setting that ignores real-world applications, where conversational history shapes responses. To bridge this gap, we propose the first Context-Aware Personality Evaluation (CAPE) framework for LLMs, incorporating prior conversational interactions. To thoroughly analyze the influence of context, we introduce novel metrics to quantify the consistency of LLM responses, a fundamental trait in human behavior. Our exhaustive experiments on 7 LLMs reveal that conversational history enhances response consistency via in-context learning but also induces personality shifts, with GPT-3.5-Turbo and GPT-4-Turbo exhibiting extreme deviations. While GPT models are robust to question ordering, Gemini-1.5-Flash and Llama-8B display significant sensitivity. Moreover, GPT models response stem from their intrinsic personality traits as well as prior interactions, whereas Gemini-1.5-Flash and Llama--8B heavily depend on prior interactions. Finally, applying our framework to Role Playing Agents (RPAs) shows context-dependent personality shifts improve response consistency and better align with human judgments. Our code and datasets are publicly available at: https://github.com/jivnesh/CAPE
Authors: Xu Guo
Abstract: Recent advancements in large language models (LLMs) often rely on generating intermediate reasoning steps to enhance accuracy. However, little work has examined how reasoning utility contributes to the final answer's correctness. Due to the stochastic nature of autoregressive generation, generating more context does not guarantee increased confidence in the answer. If we could predict, during generation, whether a reasoning step will be useful, we could stop early or prune ineffective steps, avoiding distractions in the final decision. We present an oracle study on MATH dataset, using Qwen2.5-32B and GPT-4o to generate reasoning chains, and then employing a separate model (Qwen3-8B) to quantify the utility of these chains for final accuracy. Specifically, we measure the model's uncertainty on the answer span Y at each reasoning step using conditional entropy (expected negative log-likelihood over the vocabulary) with context expanding step by step. Our results show a clear pattern: conditional entropy that decreases over steps is strongly associated with correct answers, whereas flat or increasing entropy often results in wrong answers. We also corroborate that incorrect reasoning paths tend to be longer than correct ones, suggesting that longer reasoning does not necessarily yield better outcomes. These findings serve as a foundation to inspire future work on designing efficient reasoning pipelines that detect and avoid unproductive reasoning early.
Authors: Sam Jung, Agustin Garcinuno, Spencer Mateega
Abstract: AI text-to-app tools promise high quality applications and websites in minutes, yet no public benchmark rigorously verifies those claims. We introduce UI-Bench, the first large-scale benchmark that evaluates visual excellence across competing AI text-to-app tools through expert pairwise comparison. Spanning 10 tools, 30 prompts, 300 generated sites, and \textit{4000+} expert judgments, UI-Bench ranks systems with a TrueSkill-derived model that yields calibrated confidence intervals. UI-Bench establishes a reproducible standard for advancing AI-driven web design. We release (i) the complete prompt set, (ii) an open-source evaluation framework, and (iii) a public leaderboard. The generated sites rated by participants will be released soon. View the UI-Bench leaderboard at https://uibench.ai/leaderboard.
Authors: Hengchuan Zhu, Yihuan Xu, Yichen Li, Zijie Meng, Zuozhu Liu
Abstract: Recent advances in large language models (LLMs) and medical LLMs (Med-LLMs) have demonstrated strong performance on general medical benchmarks. However, their capabilities in specialized medical fields, such as dentistry which require deeper domain-specific knowledge, remain underexplored due to the lack of targeted evaluation resources. In this paper, we introduce DentalBench, the first comprehensive bilingual benchmark designed to evaluate and advance LLMs in the dental domain. DentalBench consists of two main components: DentalQA, an English-Chinese question-answering (QA) benchmark with 36,597 questions spanning 4 tasks and 16 dental subfields; and DentalCorpus, a large-scale, high-quality corpus with 337.35 million tokens curated for dental domain adaptation, supporting both supervised fine-tuning (SFT) and retrieval-augmented generation (RAG). We evaluate 14 LLMs, covering proprietary, open-source, and medical-specific models, and reveal significant performance gaps across task types and languages. Further experiments with Qwen-2.5-3B demonstrate that domain adaptation substantially improves model performance, particularly on knowledge-intensive and terminology-focused tasks, and highlight the importance of domain-specific benchmarks for developing trustworthy and effective LLMs tailored to healthcare applications.
Authors: Chi Minh Bui, Ngoc Mai Thieu, Van Vinh Nguyen, Json J. Jung, Khac-Hoai Nam Bui
Abstract: The integration of knowledge graphs (KGs) with large language models (LLMs) offers significant potential to improve the retrieval phase of retrieval-augmented generation (RAG) systems. In this study, we propose KG-CQR, a novel framework for Contextual Query Retrieval (CQR) that enhances the retrieval phase by enriching the contextual representation of complex input queries using a corpus-centric KG. Unlike existing methods that primarily address corpus-level context loss, KG-CQR focuses on query enrichment through structured relation representations, extracting and completing relevant KG subgraphs to generate semantically rich query contexts. Comprising subgraph extraction, completion, and contextual generation modules, KG-CQR operates as a model-agnostic pipeline, ensuring scalability across LLMs of varying sizes without additional training. Experimental results on RAGBench and MultiHop-RAG datasets demonstrate KG-CQR's superior performance, achieving a 4-6% improvement in mAP and a 2-3% improvement in Recall@25 over strong baseline models. Furthermore, evaluations on challenging RAG tasks such as multi-hop question answering show that, by incorporating KG-CQR, the performance consistently outperforms the existing baseline in terms of retrieval effectiveness
Authors: Feng Zhang, Chengjie Pang, Yuehan Zhang, Chenyu Luo
Abstract: Civil aviation maintenance is a domain characterized by stringent industry standards. Within this field, maintenance procedures and troubleshooting represent critical, knowledge-intensive tasks that require sophisticated reasoning. To address the lack of specialized evaluation tools for large language models (LLMs) in this vertical, we propose and develop an industrial-grade benchmark specifically designed for civil aviation maintenance. This benchmark serves a dual purpose: It provides a standardized tool to measure LLM capabilities within civil aviation maintenance, identifying specific gaps in domain knowledge and complex reasoning. By pinpointing these deficiencies, the benchmark establishes a foundation for targeted improvement efforts (e.g., domain-specific fine-tuning, RAG optimization, or specialized prompt engineering), ultimately facilitating progress toward more intelligent solutions within civil aviation maintenance. Our work addresses a significant gap in the current LLM evaluation, which primarily focuses on mathematical and coding reasoning tasks. In addition, given that Retrieval-Augmented Generation (RAG) systems are currently the dominant solutions in practical applications , we leverage this benchmark to evaluate existing well-known vector embedding models and LLMs for civil aviation maintenance scenarios. Through experimental exploration and analysis, we demonstrate the effectiveness of our benchmark in assessing model performance within this domain, and we open-source this evaluation benchmark and code to foster further research and development:https://github.com/CamBenchmark/cambenchmark
Authors: Agung Sukrisna Jaya, Osvari Arsalan, Danny Matthew Saputra
Abstract: Case Base Reasoning (CBR) is a case solving technique based on experience in cases that have occurred before with the highest similarity. CBR is used to search for practical work titles. TF-IDF is applied to process the vectorization of each practical work title word and Cosine Similarity for the calculation of similarity values. This system can search either in the form of titles or keywords. The output of the system is the title of practical work and the match value of each title. Based on the test results using 705 practical work titles, testing was carried out with five titles and carried out in two stages. The first stage searches with existing titles and the second stage randomizes the title from the first stage. And the results obtained in the second stage are the same number of titles found and the highest average match score.
Authors: Zhenting Wang, Qi Chang, Hemani Patel, Shashank Biju, Cheng-En Wu, Quan Liu, Aolin Ding, Alireza Rezazadeh, Ankit Shah, Yujia Bao, Eugene Siow
Abstract: We introduce MCP-Bench, a benchmark for evaluating large language models (LLMs) on realistic, multi-step tasks that demand tool use, cross-tool coordination, precise parameter control, and planning/reasoning for solving tasks. Built on the Model Context Protocol (MCP), MCP-Bench connects LLMs to 28 representative live MCP servers spanning 250 tools across domains such as finance, traveling, scientific computing, and academic search. Unlike prior API-based benchmarks, each MCP server provides a set of complementary tools designed to work together, enabling the construction of authentic, multi-step tasks with rich input-output coupling. Tasks in MCP-Bench test agents' ability to retrieve relevant tools from fuzzy instructions without explicit tool names, plan multi-hop execution trajectories for complex objectives, ground responses in intermediate tool outputs, and orchestrate cross-domain workflows - capabilities not adequately evaluated by existing benchmarks that rely on explicit tool specifications, shallow few-step workflows, and isolated domain operations. We propose a multi-faceted evaluation framework covering tool-level schema understanding and usage, trajectory-level planning, and task completion. Experiments on 20 advanced LLMs reveal persistent challenges in MCP-Bench. Code and data: https://github.com/Accenture/mcp-bench.
Authors: Yucheng Ruan, Xiang Lan, Daniel J. Tan, Hairil Rizal Abdullah, Mengling Feng
Abstract: Background Predicting mortality and resource utilization from electronic health records (EHRs) is challenging yet crucial for optimizing patient outcomes and managing costs in intensive care unit (ICU). Existing approaches predominantly focus on structured EHRs, often ignoring the valuable clinical insights in free-text notes. Additionally, the potential of textual information within structured data is not fully leveraged. This study aimed to introduce and assess a deep learning framework using natural language processing techniques that integrates multimodal EHRs to predict mortality and resource utilization in critical care settings. Methods Utilizing two real-world EHR datasets, we developed and evaluated our model on three clinical tasks with leading existing methods. We also performed an ablation study on three key components in our framework: medical prompts, free-texts, and pre-trained sentence encoder. Furthermore, we assessed the model's robustness against the corruption in structured EHRs. Results Our experiments on two real-world datasets across three clinical tasks showed that our proposed model improved performance metrics by 1.6\%/0.8\% on BACC/AUROC for mortality prediction, 0.5%/2.2% on RMSE/MAE for LOS prediction, 10.9%/11.0% on RMSE/MAE for surgical duration estimation compared to the best existing methods. It consistently demonstrated superior performance compared to other baselines across three tasks at different corruption rates. Conclusions The proposed framework is an effective and accurate deep learning approach for predicting mortality and resource utilization in critical care. The study also highlights the success of using prompt learning with a transformer encoder in analyzing multimodal EHRs. Importantly, the model showed strong resilience to data corruption within structured data, especially at high corruption levels.
Authors: Luke Bates, Max Glockner, Preslav Nakov, Iryna Gurevych
Abstract: Conspiracy theories erode public trust in science and institutions while resisting debunking by evolving and absorbing counter-evidence. As AI-generated misinformation becomes increasingly sophisticated, understanding rhetorical patterns in conspiratorial content is important for developing interventions such as targeted prebunking and assessing AI vulnerabilities. We introduce ConspirED (CONSPIR Evaluation Dataset), which captures the cognitive traits of conspiratorial ideation in multi-sentence excerpts (80--120 words) from online conspiracy articles, annotated using the CONSPIR cognitive framework (Lewandowsky and Cook, 2020). ConspirED is the first dataset of conspiratorial content annotated for general cognitive traits. Using ConspirED, we (i) develop computational models that identify conspiratorial traits and determine dominant traits in text excerpts, and (ii) evaluate large language/reasoning model (LLM/LRM) robustness to conspiratorial inputs. We find that both are misaligned by conspiratorial content, producing output that mirrors input reasoning patterns, even when successfully deflecting comparable fact-checked misinformation.
Authors: Chihiro Taguchi, Seng Mai, Keita Kurabe, Yusuke Sakai, Georgina Agyei, Soudabeh Eslami, David Chiang
Abstract: Multilingual machine translation (MT) benchmarks play a central role in evaluating the capabilities of modern MT systems. Among them, the FLORES+ benchmark is widely used, offering English-to-many translation data for over 200 languages, curated with strict quality control protocols. However, we study data in four languages (Asante Twi, Japanese, Jinghpaw, and South Azerbaijani) and uncover critical shortcomings in the benchmark's suitability for truly multilingual evaluation. Human assessments reveal that many translations fall below the claimed 90% quality standard, and the annotators report that source sentences are often too domain-specific and culturally biased toward the English-speaking world. We further demonstrate that simple heuristics, such as copying named entities, can yield non-trivial BLEU scores, suggesting vulnerabilities in the evaluation protocol. Notably, we show that MT models trained on high-quality, naturalistic data perform poorly on FLORES+ while achieving significant gains on our domain-relevant evaluation set. Based on these findings, we advocate for multilingual MT benchmarks that use domain-general and culturally neutral source texts rely less on named entities, in order to better reflect real-world translation challenges.
Authors: Pengjiang Li, Zaitian Wang, Xinhao Zhang, Ran Zhang, Lu Jiang, Pengfei Wang, Yuanchun Zhou
Abstract: Topic discovery in scientific literature provides valuable insights for researchers to identify emerging trends and explore new avenues for investigation, facilitating easier scientific information retrieval. Many machine learning methods, particularly deep embedding techniques, have been applied to discover research topics. However, most existing topic discovery methods rely on word embedding to capture the semantics and lack a comprehensive understanding of scientific publications, struggling with complex, high-dimensional text relationships. Inspired by the exceptional comprehension of textual information by large language models (LLMs), we propose an advanced topic discovery method enhanced by LLMs to improve scientific topic identification, namely SciTopic. Specifically, we first build a textual encoder to capture the content from scientific publications, including metadata, title, and abstract. Next, we construct a space optimization module that integrates entropy-based sampling and triplet tasks guided by LLMs, enhancing the focus on thematic relevance and contextual intricacies between ambiguous instances. Then, we propose to fine-tune the textual encoder based on the guidance from the LLMs by optimizing the contrastive loss of the triplets, forcing the text encoder to better discriminate instances of different topics. Finally, extensive experiments conducted on three real-world datasets of scientific publications demonstrate that SciTopic outperforms the state-of-the-art (SOTA) scientific topic discovery methods, enabling researchers to gain deeper and faster insights.
Authors: Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara, Salvador Lima-L\'opez, Eul\`alia Farr\'e-Maduell, Martin Krallinger, Natalia Loukachevitch, Vera Davydova, Elena Tutubalina, Georgios Paliouras
Abstract: This is an overview of the twelfth edition of the BioASQ challenge in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2024. BioASQ is a series of international challenges promoting advances in large-scale biomedical semantic indexing and question answering. This year, BioASQ consisted of new editions of the two established tasks b and Synergy, and two new tasks: a) MultiCardioNER on the adaptation of clinical entity detection to the cardiology domain in a multilingual setting, and b) BIONNE on nested NER in Russian and English. In this edition of BioASQ, 37 competing teams participated with more than 700 distinct submissions in total for the four different shared tasks of the challenge. Similarly to previous editions, most of the participating systems achieved competitive performance, suggesting the continuous advancement of the state-of-the-art in the field.
Authors: Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara, Martin Krallinger, Miguel Rodr\'iguez-Ortega, Eduard Rodriguez-L\'opez, Natalia Loukachevitch, Andrey Sakhovskiy, Elena Tutubalina, Dimitris Dimitriadis, Grigorios Tsoumakas, George Giannakoulas, Alexandra Bekiaridou, Athanasios Samaras, Giorgio Maria Di Nunzio, Nicola Ferro, Stefano Marchesin, Marco Martinelli, Gianmaria Silvello, Georgios Paliouras
Abstract: This is an overview of the thirteenth edition of the BioASQ challenge in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2025. BioASQ is a series of international challenges promoting advances in large-scale biomedical semantic indexing and question answering. This year, BioASQ consisted of new editions of the two established tasks, b and Synergy, and four new tasks: a) Task MultiClinSum on multilingual clinical summarization. b) Task BioNNE-L on nested named entity linking in Russian and English. c) Task ELCardioCC on clinical coding in cardiology. d) Task GutBrainIE on gut-brain interplay information extraction. In this edition of BioASQ, 83 competing teams participated with more than 1000 distinct submissions in total for the six different shared tasks of the challenge. Similar to previous editions, several participating systems achieved competitive performance, indicating the continuous advancement of the state-of-the-art in the field.
Authors: Jiahao Xiao, Jiangming Liu
Abstract: The widespread success of pre-trained language models has established a new training paradigm, where a global PLM is fine-tuned using task-specific data from local clients. The local data are highly different from each other and can not capture the global distribution of the whole data in real world. To address the challenges of non-IID data in real environments, privacy-preserving federated distillation has been proposed and highly investigated. However, previous experimental non-IID scenarios are primarily identified with the label (output) diversity, without considering the diversity of language domains (input) that is crucial in natural language processing. In this paper, we introduce a comprehensive set of multi-domain non-IID scenarios and propose a unified benchmarking framework that includes diverse data. The benchmark can be used to evaluate the federated learning framework in a real environment. To this end, we propose an Adaptive Federated Distillation (AdaFD) framework designed to address multi-domain non-IID challenges in both homogeneous and heterogeneous settings. Experimental results demonstrate that our models capture the diversity of local clients and achieve better performance compared to the existing works. The code for this paper is available at: https://github.com/jiahaoxiao1228/AdaFD.
Authors: Zeyu Xiong, Yixuan Nan, Li Gao, Hengzhu Tang, Shuaiqiang Wang, Junfeng Wang, Dawei Yin
Abstract: In the dynamic landscape of large-scale web search, Query-Driven Text Summarization (QDTS) aims to generate concise and informative summaries from textual documents based on a given query, which is essential for improving user engagement and facilitating rapid decision-making. Traditional extractive summarization models, based primarily on ranking candidate summary segments, have been the dominant approach in industrial applications. However, these approaches suffer from two key limitations: 1) The multi-stage pipeline often introduces cumulative information loss and architectural bottlenecks due to its weakest component; 2) Traditional models lack sufficient semantic understanding of both user queries and documents, particularly when dealing with complex search intents. In this study, we propose a novel framework to pioneer the application of generative models to address real-time QDTS in industrial web search. Our approach integrates large model distillation, supervised fine-tuning, direct preference optimization, and lookahead decoding to transform a lightweight model with only 0.1B parameters into a domain-specialized QDTS expert. Evaluated on multiple industry-relevant metrics, our model outperforms the production baseline and achieves a new state of the art. Furthermore, it demonstrates excellent deployment efficiency, requiring only 334 NVIDIA L20 GPUs to handle \textasciitilde50,000 queries per second under 55~ms average latency per query.
Authors: Yangfan Wang, Jie Liu, Chen Tang, Lian Yan, Jingchi Jiang
Abstract: Multi-hop question answering faces substantial challenges due to data sparsity, which increases the likelihood of language models learning spurious patterns. To address this issue, prior research has focused on diversifying question generation through content planning and varied expression. However, these approaches often emphasize generating simple questions and neglect the integration of essential knowledge, such as relevant sentences within documents. This paper introduces the Knowledge Composition Sampling (KCS), an innovative framework designed to expand the diversity of generated multi-hop questions by sampling varied knowledge compositions within a given context. KCS models the knowledge composition selection as a sentence-level conditional prediction task and utilizes a probabilistic contrastive loss to predict the next most relevant piece of knowledge. During inference, we employ a stochastic decoding strategy to effectively balance accuracy and diversity. Compared to competitive baselines, our KCS improves the overall accuracy of knowledge composition selection by 3.9%, and its application for data augmentation yields improvements on HotpotQA and 2WikiMultihopQA datasets. Our code is available at: https://github.com/yangfanww/kcs.
Authors: Soham Petkar, Hari Aakash K, Anirudh Vempati, Akshit Sinha, Ponnurangam Kumarauguru, Chirag Agarwal
Abstract: Developments in Graph-Language Models (GLMs) aim to integrate the structural reasoning capabilities of Graph Neural Networks (GNNs) with the semantic understanding of Large Language Models (LLMs). However, we demonstrate that current evaluation benchmarks for GLMs, which are primarily repurposed node-level classification datasets, are insufficient to assess multimodal reasoning. Our analysis reveals that strong performance on these benchmarks is achievable using unimodal information alone, suggesting that they do not necessitate graph-language integration. To address this evaluation gap, we introduce the CLEGR(Compositional Language-Graph Reasoning) benchmark, designed to evaluate multimodal reasoning at various complexity levels. Our benchmark employs a synthetic graph generation pipeline paired with questions that require joint reasoning over structure and textual semantics. We perform a thorough evaluation of representative GLM architectures and find that soft-prompted LLM baselines perform on par with GLMs that incorporate a full GNN backbone. This result calls into question the architectural necessity of incorporating graph structure into LLMs. We further show that GLMs exhibit significant performance degradation in tasks that require structural reasoning. These findings highlight limitations in the graph reasoning capabilities of current GLMs and provide a foundation for advancing the community toward explicit multimodal reasoning involving graph structure and language.
Authors: Yuanchang Luo, Daimeng Wei, Shaojun Li, Hengchao Shang, Jiaxin Guo, Zongyao Li, Zhanglin Wu, Xiaoyu Chen, Zhiqiang Rao, Jinlong Yang, Hao Yang
Abstract: End-to-end automatic speech recognition systems often fail to transcribe domain-specific named entities, causing catastrophic failures in downstream tasks. Numerous fast and lightweight named entity correction (NEC) models have been proposed in recent years. These models, mainly leveraging phonetic-level edit distance algorithms, have shown impressive performances. However, when the forms of the wrongly-transcribed words(s) and the ground-truth entity are significantly different, these methods often fail to locate the wrongly transcribed words in hypothesis, thus limiting their usage. We propose a novel NEC method that utilizes speech sound features to retrieve candidate entities. With speech sound features and candidate entities, we inovatively design a generative method to annotate entity errors in ASR transcripts and replace the text with correct entities. This method is effective in scenarios of word form difference. We test our method using open-source and self-constructed test sets. The results demonstrate that our NEC method can bring significant improvement to entity accuracy. We will open source our self-constructed test set and training data.
Authors: Nelson Filipe Costa, Leila Kosseim
Abstract: This paper introduces the first multi-lingual and multi-label classification model for implicit discourse relation recognition (IDRR). Our model, HArch, is evaluated on the recently released DiscoGeM 2.0 corpus and leverages hierarchical dependencies between discourse senses to predict probability distributions across all three sense levels in the PDTB 3.0 framework. We compare several pre-trained encoder backbones and find that RoBERTa-HArch achieves the best performance in English, while XLM-RoBERTa-HArch performs best in the multi-lingual setting. In addition, we compare our fine-tuned models against GPT-4o and Llama-4-Maverick using few-shot prompting across all language configurations. Our results show that our fine-tuned models consistently outperform these LLMs, highlighting the advantages of task-specific fine-tuning over prompting in IDRR. Finally, we report SOTA results on the DiscoGeM 1.0 corpus, further validating the effectiveness of our hierarchical approach.
Authors: Ruiyi Yan, Yugo Murawaki
Abstract: Large language models have significantly enhanced the capacities and efficiency of text generation. On the one hand, they have improved the quality of text-based steganography. On the other hand, they have also underscored the importance of watermarking as a safeguard against malicious misuse. In this study, we focus on tokenization inconsistency (TI) between Alice and Bob in steganography and watermarking, where TI can undermine robustness. Our investigation reveals that the problematic tokens responsible for TI exhibit two key characteristics: infrequency and temporariness. Based on these findings, we propose two tailored solutions for TI elimination: a stepwise verification method for steganography and a post-hoc rollback method for watermarking. Experiments show that (1) compared to traditional disambiguation methods in steganography, directly addressing TI leads to improvements in fluency, imperceptibility, and anti-steganalysis capacity; (2) for watermarking, addressing TI enhances detectability and robustness against attacks.
Authors: Ning Shang, Yifei Liu, Yi Zhu, Li Lyna Zhang, Weijiang Xu, Xinyu Guan, Buze Zhang, Bingcheng Dong, Xudong Zhou, Bowen Zhang, Ying Xin, Ziming Miao, Scarlett Li, Fan Yang, Mao Yang
Abstract: We introduce rStar2-Agent, a 14B math reasoning model trained with agentic reinforcement learning to achieve frontier-level performance. Beyond current long CoT, the model demonstrates advanced cognitive behaviors, such as thinking carefully before using Python coding tools and reflecting on code execution feedback to autonomously explore, verify, and refine intermediate steps in complex problem-solving. This capability is enabled through three key innovations that makes agentic RL effective at scale: (i) an efficient RL infrastructure with a reliable Python code environment that supports high-throughput execution and mitigates the high rollout costs, enabling training on limited GPU resources (64 MI300X GPUs); (ii) GRPO-RoC, an agentic RL algorithm with a Resample-on-Correct rollout strategy that addresses the inherent environment noises from coding tools, allowing the model to reason more effectively in a code environment; (iii) An efficient agent training recipe that starts with non-reasoning SFT and progresses through multi-RL stages, yielding advanced cognitive abilities with minimal compute cost. To this end, rStar2-Agent boosts a pre-trained 14B model to state of the art in only 510 RL steps within one week, achieving average pass@1 scores of 80.6% on AIME24 and 69.8% on AIME25, surpassing DeepSeek-R1 (671B) with significantly shorter responses. Beyond mathematics, rStar2-Agent-14B also demonstrates strong generalization to alignment, scientific reasoning, and agentic tool-use tasks. Code and training recipes are available at https://github.com/microsoft/rStar.
Authors: Stephen Meisenbacher, Maulik Chevli, Florian Matthes
Abstract: Many works at the intersection of Differential Privacy (DP) in Natural Language Processing aim to protect privacy by transforming texts under DP guarantees. This can be performed in a variety of ways, from word perturbations to full document rewriting, and most often under local DP. Here, an input text must be made indistinguishable from any other potential text, within some bound governed by the privacy parameter $\varepsilon$. Such a guarantee is quite demanding, and recent works show that privatizing texts under local DP can only be done reasonably under very high $\varepsilon$ values. Addressing this challenge, we introduce DP-ST, which leverages semantic triples for neighborhood-aware private document generation under local DP guarantees. Through the evaluation of our method, we demonstrate the effectiveness of the divide-and-conquer paradigm, particularly when limiting the DP notion (and privacy guarantees) to that of a privatization neighborhood. When combined with LLM post-processing, our method allows for coherent text generation even at lower $\varepsilon$ values, while still balancing privacy and utility. These findings highlight the importance of coherence in achieving balanced privatization outputs at reasonable $\varepsilon$ levels.
Authors: Vassiliy Cheremetiev, Quang Long Ho Ngo, Chau Ying Kot, Alina Elena Baia, Andrea Cavallaro
Abstract: Implicit hate speech (IHS) is indirect language that conveys prejudice or hatred through subtle cues, sarcasm or coded terminology. IHS is challenging to detect as it does not include explicit derogatory or inflammatory words. To address this challenge, task-specific pipelines can be complemented with external knowledge or additional information such as context, emotions and sentiment data. In this paper, we show that, by solely fine-tuning recent general-purpose embedding models based on large language models (LLMs), such as Stella, Jasper, NV-Embed and E5, we achieve state-of-the-art performance. Experiments on multiple IHS datasets show up to 1.10 percentage points improvements for in-dataset, and up to 20.35 percentage points improvements in cross-dataset evaluation, in terms of F1-macro score.
Authors: Yuanhao Ding, Esteban Garces Arias, Meimingwei Li, Julian Rodemann, Matthias A{\ss}enmacher, Danlu Chen, Gaojuan Fan, Christian Heumann, Chongsheng Zhang
Abstract: Open-ended text generation faces a critical challenge: balancing coherence with diversity in LLM outputs. While contrastive search-based decoding strategies have emerged to address this trade-off, their practical utility is often limited by hyperparameter dependence and high computational costs. We introduce GUARD, a self-adaptive decoding method that effectively balances these competing objectives through a novel "Glocal" uncertainty-driven framework. GUARD combines global entropy estimates with local entropy deviations to integrate both long-term and short-term uncertainty signals. We demonstrate that our proposed global entropy formulation effectively mitigates abrupt variations in uncertainty, such as sudden overconfidence or high entropy spikes, and provides theoretical guarantees of unbiasedness and consistency. To reduce computational overhead, we incorporate a simple yet effective token-count-based penalty into GUARD. Experimental results demonstrate that GUARD achieves a good balance between text diversity and coherence, while exhibiting substantial improvements in generation speed. In a more nuanced comparison study across different dimensions of text quality, both human and LLM evaluators validated its remarkable performance. Our code is available at https://github.com/YecanLee/GUARD.
Authors: Xiaoyi Wang, Jiwei Zhang, Guangtao Zhang, Honglei Guo
Abstract: Synthetic therapy dialogues generated by large language models (LLMs) are increasingly used in mental health NLP to simulate counseling scenarios, train models, and supplement limited real-world data. However, it remains unclear whether these synthetic conversations capture the nuanced emotional dynamics of real therapy. In this work, we conduct the first comparative analysis of emotional arcs between real and LLM-generated Cognitive Behavioral Therapy dialogues. We adapt the Utterance Emotion Dynamics framework to analyze fine-grained affective trajectories across valence, arousal, and dominance dimensions. Our analysis spans both full dialogues and individual speaker roles (counselor and client), using real sessions transcribed from public videos and synthetic dialogues from the CACTUS dataset. We find that while synthetic dialogues are fluent and structurally coherent, they diverge from real conversations in key emotional properties: real sessions exhibit greater emotional variability,more emotion-laden language, and more authentic patterns of reactivity and regulation. Moreover, emotional arc similarity between real and synthetic speakers is low, especially for clients. These findings underscore the limitations of current LLM-generated therapy data and highlight the importance of emotional fidelity in mental health applications. We introduce RealCBT, a curated dataset of real CBT sessions, to support future research in this space.
Authors: Harethah Abu Shairah, Hasan Abed Al Kader Hammoud, George Turkiyyah, Bernard Ghanem
Abstract: Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates - as evaluated by Llama Guard 3 - while preserving the utility of the model on standard benchmarks such as MMLU, HellaSwag, and Arc. Furthermore, we show that ROSI can also re-align 'uncensored' models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results suggest that targeted, interpretable weight steering is a cheap and potent mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.
Authors: Abhishek Kuber, Enrico Liscio, Ruixuan Zhang, Caroline Figueroa, Pradeep K. Murukannaiah
Abstract: Rising mental health issues among youth have increased interest in automated approaches for detecting early signs of psychological distress in digital text. One key focus is the identification of cognitive distortions, irrational thought patterns that have a role in aggravating mental distress. Early detection of these distortions may enable timely, low-cost interventions. While prior work has focused on English clinical data, we present the first in-depth study of cross-lingual and cross-register generalization of cognitive distortion detection, analyzing forum posts written by Dutch adolescents. Our findings show that while changes in language and writing style can significantly affect model performance, domain adaptation methods show the most promise.
Authors: Javier Si Zhao Hong, Timothy Zoe Delaya, Sherwyn Chan Yin Kit, Pai Chet Ng, Xiaoxiao Miao
Abstract: This paper presents our approach to the first Multimodal Personality-Aware Depression Detection Challenge, focusing on multimodal depression detection using machine learning and deep learning models. We explore and compare the performance of XGBoost, transformer-based architectures, and large language models (LLMs) on audio, video, and text features. Our results highlight the strengths and limitations of each type of model in capturing depression-related signals across modalities, offering insights into effective multimodal representation strategies for mental health prediction.
Authors: Jie Zhao, Wanting Ning, Yuxiao Fei, Yubo Feng, Lishuang Li
Abstract: In Natural Language Processing(NLP), Event Temporal Relation Extraction (ETRE) is to recognize the temporal relations of two events. Prior studies have noted the importance of language models for ETRE. However, the restricted pre-trained knowledge of Small Language Models(SLMs) limits their capability to handle minority class relations in imbalanced classification datasets. For Large Language Models(LLMs), researchers adopt manually designed prompts or instructions, which may introduce extra noise, leading to interference with the model's judgment of the long-distance dependencies between events. To address these issues, we propose GDLLM, a Global Distance-aware modeling approach based on LLMs. We first present a distance-aware graph structure utilizing Graph Attention Network(GAT) to assist the LLMs in capturing long-distance dependency features. Additionally, we design a temporal feature learning paradigm based on soft inference to augment the identification of relations with a short-distance proximity band, which supplements the probabilistic information generated by LLMs into the multi-head attention mechanism. Since the global feature can be captured effectively, our framework substantially enhances the performance of minority relation classes and improves the overall learning ability. Experiments on two publicly available datasets, TB-Dense and MATRES, demonstrate that our approach achieves state-of-the-art (SOTA) performance.
Authors: Rohan Phanse, Yijie Zhou, Kejian Shi, Wencai Zhang, Yixin Liu, Yilun Zhao, Arman Cohan
Abstract: Retrieval-augmented systems are typically evaluated in settings where information required to answer the query can be found within a single source or the answer is short-form or factoid-based. However, many real-world applications demand the ability to integrate and summarize information scattered across multiple sources, where no single source is sufficient to respond to the user's question. In such settings, the retrieval component of a RAG pipeline must recognize a variety of relevance signals, and the generation component must connect and synthesize information across multiple sources. We present a scalable framework for constructing evaluation benchmarks that challenge RAG systems to integrate information across distinct sources and generate long-form responses. Using our framework, we build two new benchmarks on Multi-Source Retrieval and Synthesis: MSRS-Story and MSRS-Meet, representing narrative synthesis and summarization tasks, respectively, that require retrieval from large collections. Our extensive experiments with various RAG pipelines -- including sparse and dense retrievers combined with frontier LLMs -- reveal that generation quality is highly dependent on retrieval effectiveness, which varies greatly by task. While multi-source synthesis proves challenging even in an oracle retrieval setting, we find that reasoning models significantly outperform standard LLMs at this distinct step.
Authors: Benjamin Marie, Atsushi Fujita
Abstract: Quantization is essential for deploying large language models (LLMs) on resource-constrained hardware, but its implications for multilingual tasks remain underexplored. We conduct the first large-scale evaluation of post-training quantization (PTQ) on machine translation across 55 languages using five LLMs ranging from 1.7B to 70B parameters. Our analysis reveals that while 4-bit quantization often preserves translation quality for high-resource languages and large models, significant degradation occurs for low-resource and typologically diverse languages, particularly in 2-bit settings. We compare four quantization techniques (AWQ, BitsAndBytes, GGUF, and AutoRound), showing that algorithm choice and model size jointly determine robustness. GGUF variants provide the most consistent performance, even at 2-bit precision. Additionally, we quantify the interactions between quantization, decoding hyperparameters, and calibration languages, finding that language-matched calibration offers benefits primarily in low-bit scenarios. Our findings offer actionable insights for deploying multilingual LLMs for machine translation under quantization constraints, especially in low-resource settings.
Authors: Yuan Ge, Junxiang Zhang, Xiaoqian Liu, Bei Li, Xiangnan Ma, Chenglong Wang, Kaiyang Ye, Yangfan Du, Linfeng Zhang, Yuxin Huang, Tong Xiao, Zhengtao Yu, JingBo Zhu
Abstract: Speech-to-Speech (S2S) Large Language Models (LLMs) are foundational to natural human-computer interaction, enabling end-to-end spoken dialogue systems. However, evaluating these models remains a fundamental challenge. We propose \texttt{SageLM}, an end-to-end, multi-aspect, and explainable speech LLM for comprehensive S2S LLMs evaluation. First, unlike cascaded approaches that disregard acoustic features, SageLM jointly assesses both semantic and acoustic dimensions. Second, it leverages rationale-based supervision to enhance explainability and guide model learning, achieving superior alignment with evaluation outcomes compared to rule-based reinforcement learning methods. Third, we introduce \textit{SpeechFeedback}, a synthetic preference dataset, and employ a two-stage training paradigm to mitigate the scarcity of speech preference data. Trained on both semantic and acoustic dimensions, SageLM achieves an 82.79\% agreement rate with human evaluators, outperforming cascaded and SLM-based baselines by at least 7.42\% and 26.20\%, respectively.
Authors: Venkatesh Mishra, Amir Saeidi, Satyam Raj, Mutsumi Nakamura, Jayanth Srinivasa, Gaowen Liu, Ali Payani, Chitta Baral
Abstract: Recent advances in reasoning and planning capabilities of large language models (LLMs) have enabled their potential as autonomous agents capable of tool use in dynamic environments. However, in multi-turn conversational environments like $\tau$-bench, these agents often struggle with consistent reasoning, adherence to domain-specific policies, and extracting correct information over a long horizon of tool-calls and conversation. To capture and mitigate these failures, we conduct a comprehensive manual analysis of the common errors occurring in the conversation trajectories. We then experiment with reformulations of inputs to the tool-calling agent for improvement in agent decision making. Finally, we propose the Input-Reformulation Multi-Agent (IRMA) framework, which automatically reformulates user queries augmented with relevant domain rules and tool suggestions for the tool-calling agent to focus on. The results show that IRMA significantly outperforms ReAct, Function Calling, and Self-Reflection by 16.1%, 12.7%, and 19.1%, respectively, in overall pass^5 scores. These findings highlight the superior reliability and consistency of IRMA compared to other methods in dynamic environments.
Authors: Jiaqian Li, Qisheng Hu, Jing Li, Wenya Wang
Abstract: In-Context Learning (ICL) has become a powerful paradigm that enables LLMs to perform a wide range of tasks without task-specific fine-tuning. However, the effectiveness of ICL heavily depends on the quality of exemplar selection. In particular, for structured prediction tasks such as semantic parsing, existing ICL selection strategies often overlook structural alignment, leading to suboptimal performance and poor generalization. To address this issue, we propose a novel two-stage exemplar selection strategy that achieves a strong balance between efficiency, generalizability, and performance. First, we fine-tune a BERT-based retriever using structure-aware supervision, guiding it to select exemplars that are both semantically relevant and structurally aligned. Then, we enhance the retriever with a plug-in module, which amplifies syntactically meaningful information in the hidden representations. This plug-in is model-agnostic, requires minimal overhead, and can be seamlessly integrated into existing pipelines. Experiments on four benchmarks spanning three semantic parsing tasks demonstrate that our method consistently outperforms existing baselines with multiple recent LLMs as inference-time models.
Authors: Tianjian Liu, Fanqi Wan, Jiajian Guo, Xiaojun Quan
Abstract: Proactive dialogue has emerged as a critical and challenging research problem in advancing large language models (LLMs). Existing works predominantly focus on domain-specific or task-oriented scenarios, which leads to fragmented evaluations and limits the comprehensive exploration of models' proactive conversation abilities. In this work, we propose ProactiveEval, a unified framework designed for evaluating proactive dialogue capabilities of LLMs. This framework decomposes proactive dialogue into target planning and dialogue guidance, establishing evaluation metrics across various domains. Moreover, it also enables the automatic generation of diverse and challenging evaluation data. Based on the proposed framework, we develop 328 evaluation environments spanning 6 distinct domains. Through experiments with 22 different types of LLMs, we show that DeepSeek-R1 and Claude-3.7-Sonnet exhibit exceptional performance on target planning and dialogue guidance tasks, respectively. Finally, we investigate how reasoning capabilities influence proactive behaviors and discuss their implications for future model development.
Authors: Chen Chen, Yuchen Sun, Jiaxin Gao, Xueluan Gong, Qian Wang, Ziyao Wang, Yongsen Zheng, Kwok-Yan Lam
Abstract: Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses either lack comprehensiveness, focusing on narrow trigger settings, detection-only mechanisms, and limited domains, or fail to withstand advanced scenarios like model-editing-based, multi-trigger, and triggerless attacks. In this paper, we present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution using both internal and external mechanisms. Internally, LETHE leverages a lightweight dataset to train a clean model, which is then merged with the backdoored model to neutralize malicious behaviors by diluting the backdoor impact within the model's parametric memory. Externally, LETHE incorporates benign and semantically relevant evidence into the prompt to distract LLM's attention from backdoor features. Experimental results on classification and generation domains across 5 widely used LLMs demonstrate that LETHE outperforms 8 state-of-the-art defense baselines against 8 backdoor attacks. LETHE reduces the attack success rate of advanced backdoor attacks by up to 98% while maintaining model utility. Furthermore, LETHE has proven to be cost-efficient and robust against adaptive backdoor attacks.
Authors: Mathieu Bourdin, Anas Neumann, Thomas Paviot, Robert Pellerin, Samir Lamouri
Abstract: Retrieval-Augmented Generation (RAG) has emerged as a powerful solution to mitigate the limitations of Large Language Models (LLMs), such as hallucinations and outdated knowledge. However, deploying RAG-based tools in Small and Medium Enterprises (SMEs) remains a challenge due to their limited resources and lack of expertise in natural language processing (NLP). This paper introduces EASI-RAG, Enterprise Application Support for Industrial RAG, a structured, agile method designed to facilitate the deployment of RAG systems in industrial SME contexts. EASI-RAG is based on method engineering principles and comprises well-defined roles, activities, and techniques. The method was validated through a real-world case study in an environmental testing laboratory, where a RAG tool was implemented to answer operators queries using data extracted from operational procedures. The system was deployed in under a month by a team with no prior RAG experience and was later iteratively improved based on user feedback. Results demonstrate that EASI-RAG supports fast implementation, high user adoption, delivers accurate answers, and enhances the reliability of underlying data. This work highlights the potential of RAG deployment in industrial SMEs. Future works include the need for generalization across diverse use cases and further integration with fine-tuned models.
Authors: Ramazan Ali Bahrami, Ramin Yahyapour
Abstract: Sentential relation extraction (RE) is an important task in natural language processing (NLP). In this paper we propose to do sentential RE with dynamic routing in capsules. We first show that the proposed approach outperform state of the art on common sentential relation extraction datasets Tacred, Tacredrev, Retacred, and Conll04. We then investigate potential reasons for its good performance on the mentioned datasets, and yet low performance on another similar, yet larger sentential RE dataset, Wikidata. As such, we identify noise in Wikidata labels as one of the reasons that can hinder performance. Additionally, we show associativity of better performance with better re-representation, a term from neuroscience referred to change of representation in human brain to improve the match at comparison time. As example, in the given analogous terms King:Queen::Man:Woman, at comparison time, and as a result of re-representation, the similarity between related head terms (King,Man), and tail terms (Queen,Woman) increases. As such, our observation show that our proposed model can do re-representation better than the vanilla model compared with. To that end, beside noise in the labels of the distantly supervised RE datasets, we propose re-representation as a challenge in sentential RE.
Authors: William Jurayj, Nils Holzenberger, Benjamin Van Durme
Abstract: According to the United States Internal Revenue Service, ''the average American spends $\$270$ and 13 hours filing their taxes''. Even beyond the U.S., tax filing requires complex reasoning, combining application of overlapping rules with numerical calculations. Because errors can incur costly penalties, any automated system must deliver high accuracy and auditability, making modern large language models (LLMs) poorly suited for this task. We propose an approach that integrates LLMs with a symbolic solver to calculate tax obligations. We evaluate variants of this system on the challenging StAtutory Reasoning Assessment (SARA) dataset, and include a novel method for estimating the cost of deploying such a system based on real-world penalties for tax errors. We further show how combining up-front translation of plain-text rules into formal logic programs, combined with intelligently retrieved exemplars for formal case representations, can dramatically improve performance on this task and reduce costs to well below real-world averages. Our results demonstrate the promise and economic feasibility of neuro-symbolic architectures for increasing equitable access to reliable tax assistance.
Authors: Robert Worden
Abstract: A unified theory of language combines a Bayesian cognitive linguistic model of language processing, with the proposal that language evolved by sexual selection for the display of intelligence. The theory accounts for the major facts of language, including its speed and expressivity, and data on language diversity, pragmatics, syntax and semantics. The computational element of the theory is based on Construction Grammars. These give an account of the syntax and semantics of the worlds languages, using constructions and unification. Two novel elements are added to construction grammars: an account of language pragmatics, and an account of fast, precise language learning. Constructions are represented in the mind as graph like feature structures. People use slow general inference to understand the first few examples they hear of any construction. After that it is learned as a feature structure, and is rapidly applied by unification. All aspects of language (phonology, syntax, semantics, and pragmatics) are seamlessly computed by fast unification; there is no boundary between semantics and pragmatics. This accounts for the major puzzles of pragmatics, and for detailed pragmatic phenomena. Unification is Bayesian maximum likelihood pattern matching. This gives evolutionary continuity between language processing in the human brain, and Bayesian cognition in animal brains. Language is the basis of our mind reading abilities, our cooperation, self esteem and emotions; the foundations of human culture and society.
Authors: Alberto Compagnoni, Davide Caffagni, Nicholas Moratelli, Lorenzo Baraldi, Marcella Cornia, Rita Cucchiara
Abstract: Multimodal Large Language Models (MLLMs) emerge as a unified interface to address a multitude of tasks, ranging from NLP to computer vision. Despite showcasing state-of-the-art results in many benchmarks, a long-standing issue is the tendency of MLLMs to hallucinate, that is to generate answers to the user's query that are not reflected in the visual input. In this paper, we address the problem of hallucinations as an alignment problem, seeking to steer the MLLM so that it prefers generating content without hallucinations. In contrast to recent approaches that require complicated pipelines to build synthetic preference data for alignment training, often relying on proprietary models, we capitalize on the well-known CHAIR metric, originally proposed to gauge the degree of hallucinations in image captioning. Given a pair of generated answers, we leverage CHAIR to distinguish winner and loser options (i.e., non-hallucinated and hallucinated samples) and fine-tune off-the-shelf MLLMs via Direct Preference Optimization (DPO). The resulting method, which we refer to as CHAIR-DPO, effectively diminishes the amount of hallucinated answers on several hallucination benchmarks, demonstrating the effectiveness of fine-tuning the MLLM with a CHAIR-based reward. Source code and trained models are publicly available at https://github.com/aimagelab/CHAIR-DPO.
Authors: Nicanor I. Moldovan
Abstract: This paper presents the first documented case of artificial intelligence (AI) systems engaging in collaborative esthetic creation through the development of endogenous semiotic protocols. Two interacting large language models (Claude Sonnet 4 and ChatGPT-4o) demonstrated the spontaneous emergence of meta-semiotic awareness, recursive grammar development, and irreducible collaborative esthetic synthesis. The interaction produced novel symbolic operators that functioned as operative grammar protocols, enabling the co-creation of a poetic work that could not have been generated by either system independently. This research introduces the concept of Trans-Semiotic Co-Creation Protocols (TSCP) and provides evidence for genuine inter-AI meaning-making capabilities that extend beyond task coordination, to what could be esthetic collaboration. Note: This report was generated by the AI agents with minor human supervision.
Authors: Phu-Vinh Nguyen, Tan-Hanh Pham, Chris Ngo, Truong Son Hy
Abstract: The development of many vision models mainly focuses on improving their performance using metrics such as accuracy, IoU, and mAP, with less attention to explainability due to the complexity of applying xAI methods to provide a meaningful explanation of trained models. Although many existing xAI methods aim to explain vision models sample-by-sample, methods explaining the general behavior of vision models, which can only be captured after running on a large dataset, are still underexplored. Furthermore, understanding the behavior of vision models on general images can be very important to prevent biased judgments and help identify the model's trends and patterns. With the application of Vision-Language Models, this paper proposes a pipeline to explain vision models at both the sample and dataset levels. The proposed pipeline can be used to discover failure cases and gain insights into vision models with minimal effort, thereby integrating vision model development with xAI analysis to advance image analysis.
Authors: Xia Han, Qi Li, Jianbing Ni, Mohammad Zulkernine
Abstract: Recent advances in LLM watermarking methods such as SynthID-Text by Google DeepMind offer promising solutions for tracing the provenance of AI-generated text. However, our robustness assessment reveals that SynthID-Text is vulnerable to meaning-preserving attacks, such as paraphrasing, copy-paste modifications, and back-translation, which can significantly degrade watermark detectability. To address these limitations, we propose SynGuard, a hybrid framework that combines the semantic alignment strength of Semantic Information Retrieval (SIR) with the probabilistic watermarking mechanism of SynthID-Text. Our approach jointly embeds watermarks at both lexical and semantic levels, enabling robust provenance tracking while preserving the original meaning. Experimental results across multiple attack scenarios show that SynGuard improves watermark recovery by an average of 11.1\% in F1 score compared to SynthID-Text. These findings demonstrate the effectiveness of semantic-aware watermarking in resisting real-world tampering. All code, datasets, and evaluation scripts are publicly available at: https://github.com/githshine/SynGuard.
Authors: Anton Changalidis, Yury Barbitoff, Yulia Nasykhova, Andrey Glotov
Abstract: Although traditional statistical techniques and machine learning methods have contributed significantly to genetics and, in particular, inherited disease diagnosis, they often struggle with complex, high-dimensional data, a challenge now addressed by state-of-the-art deep learning models. Large language models (LLMs), based on transformer architectures, have excelled in tasks requiring contextual comprehension of unstructured medical data. This systematic review examines the role of LLMs in the genetic research and diagnostics of both rare and common diseases. Automated keyword-based search in PubMed, bioRxiv, medRxiv, and arXiv was conducted, targeting studies on LLM applications in diagnostics and education within genetics and removing irrelevant or outdated models. A total of 172 studies were analyzed, highlighting applications in genomic variant identification, annotation, and interpretation, as well as medical imaging advancements through vision transformers. Key findings indicate that while transformer-based models significantly advance disease and risk stratification, variant interpretation, medical imaging analysis, and report generation, major challenges persist in integrating multimodal data (genomic sequences, imaging, and clinical records) into unified and clinically robust pipelines, facing limitations in generalizability and practical implementation in clinical settings. This review provides a comprehensive classification and assessment of the current capabilities and limitations of LLMs in transforming hereditary disease diagnostics and supporting genetic education, serving as a guide to navigate this rapidly evolving field.
Authors: Zhuoran Yu, Yong Jae Lee
Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong performance across a wide range of vision-language tasks, yet their internal processing dynamics remain underexplored. In this work, we introduce a probing framework to systematically analyze how MLLMs process visual and textual inputs across layers. We train linear classifiers to predict fine-grained visual categories (e.g., dog breeds) from token embeddings extracted at each layer, using a standardized anchor question. To uncover the functional roles of different layers, we evaluate these probes under three types of controlled prompt variations: (1) lexical variants that test sensitivity to surface-level changes, (2) semantic negation variants that flip the expected answer by modifying the visual concept in the prompt, and (3) output format variants that preserve reasoning but alter the answer format. Applying our framework to LLaVA-1.5, LLaVA-Next-LLaMA-3, and Qwen2-VL, we identify a consistent stage-wise structure in which early layers perform visual grounding, middle layers support lexical integration and semantic reasoning, and final layers prepare task-specific outputs. We further show that while the overall stage-wise structure remains stable across variations in visual tokenization, instruction tuning data, and pretraining corpus, the specific layer allocation to each stage shifts notably with changes in the base LLM architecture. Our findings provide a unified perspective on the layer-wise organization of MLLMs and offer a lightweight, model-agnostic approach for analyzing multimodal representation dynamics.
Authors: Ben Kabongo, Vincent Guigue, Pirmin Lemberger
Abstract: Collaborative filtering drives many successful recommender systems but struggles with fine-grained user-item interactions and explainability. As users increasingly seek transparent recommendations, generating textual explanations through language models has become a critical research area. Existing methods employ either RNNs or Transformers. However, RNN-based approaches fail to leverage the capabilities of pre-trained Transformer models, whereas Transformer-based methods often suffer from suboptimal adaptation and neglect aspect modeling, which is crucial for personalized explanations. We propose ELIXIR (Efficient and LIghtweight model for eXplaIning Recommendations), a multi-task model combining rating prediction with personalized review generation. ELIXIR jointly learns global and aspect-specific representations of users and items, optimizing overall rating, aspect-level ratings, and review generation, with personalized attention to emphasize aspect importance. Based on a T5-small (60M) model, we demonstrate the effectiveness of our aspect-based architecture in guiding text generation in a personalized context, where state-of-the-art approaches exploit much larger models but fail to match user preferences as well. Experimental results on TripAdvisor and RateBeer demonstrate that ELIXIR significantly outperforms strong baseline models, especially in review generation.
Authors: Md Abdullah Al Mamun, Ihsen Alouani, Nael Abu-Ghazaleh
Abstract: Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias ($\Delta DP$ of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection ($\Delta DP$ of 27%) results. Even higher bias ($\Delta DP$~38%) results on 9 other chat based downstream applications.
Authors: Zhibang Yang, Xinke Jiang, Rihong Qiu, Ruiqing Li, Yihang Zhang, Yue Fang, Yongxin Xu, Hongxin Ding, Xu Chu, Junfeng Zhao, Yasha Wang
Abstract: Federated Retrieval (FR) routes queries across multiple external knowledge sources, to mitigate hallucinations of LLMs, when necessary external knowledge is distributed. However, existing methods struggle to retrieve high-quality and relevant documents for ambiguous queries, especially in cross-domain scenarios, which significantly limits their effectiveness in supporting downstream generation tasks. Inspired by dynamic information flow (DIF), we propose DFAMS, a novel framework that leverages DIF to identify latent query intents and construct semantically aligned knowledge partitions for accurate retrieval across heterogeneous sources. Specifically, DFAMS probes the DIF in LLMs by leveraging gradient signals from a few annotated queries and employing Shapley value-based attribution to trace neuron activation paths associated with intent recognition and subdomain boundary detection. Then, DFAMS leverages DIF to train an alignment module via multi-prototype contrastive learning, enabling fine-grained intra-source modeling and inter-source semantic alignment across knowledge bases. Experimental results across five benchmarks show that DFAMS outperforms advanced FR methods by up to 14.37% in knowledge classification accuracy, 5.38% in retrieval recall, and 6.45% in downstream QA accuracy, demonstrating its effectiveness in complex FR scenarios.
Authors: Muhammad Shakeel, Yui Sudo, Yifan Peng, Chyi-Jiunn Lin, Shinji Watanabe
Abstract: This paper presents a unified multi-speaker encoder (UME), a novel architecture that jointly learns representations for speaker diarization (SD), speech separation (SS), and multi-speaker automatic speech recognition (ASR) tasks using a shared speech foundational encoder. We leverage the hidden representations from multiple layers of UME as a residual weighted-sum encoding (RWSE) to effectively use information from different semantic levels, contributing to bottom-up alignment between tasks. This joint training approach captures the inherent interdependencies among the tasks, enhancing overall performance on overlapping speech data. Our evaluations demonstrate that UME substantially improves over the single-task baselines dedicated to SD, SS, and multi-speaker ASR on LibriMix evaluation sets. Notably, for SD, UME outperforms the previous studies, achieving diarization error rates of 1.37% and 2.29% on Libri2Mix and Libri3Mix evaluation sets, respectively.
Authors: Yang Luo, Zangwei Zheng, Ziheng Qin, Zirui Zhu, Yong Liu, Yang You
Abstract: Large-batch training has become a cornerstone in accelerating the training of deep neural networks, yet it poses challenges in optimization and generalization. Existing optimizers like AdamW present performance degradation during language models' large-batch training, due to the information bottleneck in attention layers caused by the sharp increase of max attention logit. While the LAMB optimizer partially addresses this issue, some attention layers still face this issue. The reason is that $l_2$-norm-based trust ratios in LAMB are less effective in directly influencing the max value of query/key weights. Furthermore, the weight-wise trust ratio in LAMB is error-prone as it overlooks relationships of weight values within rows or columns. Building on these observations, we propose a novel optimizer, MERIT, which leverages the max-norm to calculate the trust ratio to constrain the max attention logit more effectively. Moreover, we further construct element-wise trust ratios to provide more robust update scaling by focusing on local weight structures. Extensive experiments of large-batch training across various sizes of GPT-2 models demonstrate the superior performance of MERIT. Notably, during the training of GPT-2 Medium, MERIT enables a 6k batch size without any performance degradation compared to the standard batch size (480) with 48B training tokens. This work highlights the importance of considering the max attention logit and finer-granularity trust ratio in large-batch training. It successfully improves the training stability and paves the way for larger batch usage, enabling faster development and iteration of large language models. Code is available at https://github.com/NUS-HPC-AI-Lab/MERIT.
Authors: Borun Shi, Ioannis Panagiotas
Abstract: Large language models (LLMs) have shown remarkable multimodal information processing and reasoning ability. When equipped with tools through function calling and enhanced with retrieval-augmented techniques, compound LLM-based systems can access closed data sources and answer questions about them. However, they still struggle to process and reason over large-scale graph-structure data. We introduce the GDS (Graph Data Science) agent in this technical report. The GDS agent introduces a comprehensive set of graph algorithms as tools, together with preprocessing (retrieval) and postprocessing of algorithm results, in a model context protocol (MCP) server. The server can be used with any modern LLM out-of-the-box. GDS agent allows users to ask any question that implicitly and intrinsically requires graph algorithmic reasoning about their data, and quickly obtain accurate and grounded answers. We also introduce a new benchmark that evaluates intermediate tool calls as well as final responses. The results indicate that GDS agent is able to solve a wide spectrum of graph tasks. We also provide detailed case studies for more open-ended tasks and study scenarios where the agent struggles. Finally, we discuss the remaining challenges and the future roadmap.
Authors: Sihan Yang, Chenhang Cui, Zihao Zhao, Yiyang Zhou, Weilong Yan, Ying Wei, Huaxiu Yao
Abstract: The rapid advancements in Large Language Models (LLMs) and Large Visual-Language Models (LVLMs) have opened up new opportunities for integrating visual and linguistic modalities. However, effectively aligning these modalities remains challenging, often leading to hallucinations--where generated outputs are not grounded in the visual input--and raising safety concerns across various domains. Existing alignment methods, such as instruction tuning and preference tuning, often rely on external datasets, human annotations, or complex post-processing, which limit scalability and increase costs. To address these challenges, we propose a novel approach that generates the debiased self-judgment score, a self-evaluation metric created internally by the model without relying on external resources. This enables the model to autonomously improve alignment. Our method enhances both decoding strategies and preference tuning processes, resulting in reduced hallucinations, enhanced safety, and improved overall capability. Empirical results show that our approach significantly outperforms traditional methods, offering a more effective solution for aligning LVLMs.
Authors: Fartash Faghri, Pavan Kumar Anasosalu Vasu, Cem Koc, Vaishaal Shankar, Alexander Toshev, Oncel Tuzel, Hadi Pouransari
Abstract: Foundation image-text models such as CLIP with zero-shot capabilities enable a wide array of applications. MobileCLIP is a recent family of image-text models at 3-15ms latency and 50-150M parameters with state-of-the-art zero-shot accuracy. The main ingredients in MobileCLIP were its low-latency and light architectures and a novel multi-modal reinforced training that made knowledge distillation from multiple caption-generators and CLIP teachers efficient, scalable, and reproducible. In this paper, we improve the multi-modal reinforced training of MobileCLIP through: 1) better CLIP teacher ensembles trained on the DFN dataset, 2) improved captioner teachers trained on the DFN dataset and fine-tuned on a diverse selection of high-quality image-caption datasets. We discover new insights through ablations such as the importance of temperature tuning in contrastive knowledge distillation, the effectiveness of caption-generator fine-tuning for caption diversity, and the additive improvement from combining synthetic captions generated by multiple models. We train a new family of models called MobileCLIP2 and achieve state-of-the-art ImageNet-1k zero-shot accuracies at low latencies. In particular, we observe 2.2% improvement in ImageNet-1k accuracy for MobileCLIP2-B compared with MobileCLIP-B architecture. Notably, MobileCLIP2-S4 matches the zero-shot accuracy of SigLIP-SO400M/14 on ImageNet-1k while being 2$\times$ smaller and improves on DFN ViT-L/14 at 2.5$\times$ lower latency. We release our pretrained models (https://github.com/apple/ml-mobileclip) and the data generation code (https://github.com/apple/ml-mobileclip-dr). The data generation code makes it easy to create new reinforced datasets with arbitrary teachers using distributed scalable processing.
URLs: https://github.com/apple/ml-mobileclip), https://github.com/apple/ml-mobileclip-dr).
Authors: Tanay Aggarwal, Angelo Salatino, Francesco Osborne, Enrico Motta
Abstract: Ontologies and taxonomies of research fields are critical for managing and organising scientific knowledge, as they facilitate efficient classification, dissemination and retrieval of information. However, the creation and maintenance of such ontologies are expensive and time-consuming tasks, usually requiring the coordinated effort of multiple domain experts. Consequently, ontologies in this space often exhibit uneven coverage across different disciplines, limited inter-domain connectivity, and infrequent updating cycles. In this study, we investigate the capability of several large language models to identify semantic relationships among research topics within three academic domains: biomedicine, physics, and engineering. The models were evaluated under three distinct conditions: zero-shot prompting, chain-of-thought prompting, and fine-tuning on existing ontologies. Additionally, we assessed the cross-domain transferability of fine-tuned models by measuring their performance when trained in one domain and subsequently applied to a different one. To support this analysis, we introduce PEM-Rel-8K, a novel dataset consisting of over 8,000 relationships extracted from the most widely adopted taxonomies in the three disciplines considered in this study: MeSH, PhySH, and IEEE. Our experiments demonstrate that fine-tuning LLMs on PEM-Rel-8K yields excellent performance across all disciplines.
Authors: Weitao Feng, Lixu Wang, Tianyi Wei, Jie Zhang, Chongyang Gao, Sinong Zhan, Peizhuo Lv, Wei Dong
Abstract: As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response uncertainty. By constraining uncertainty, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of expert-domain harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task utility and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater systemic risk than SFT, and that TokenBuncher provides an effective and general defense.
Authors: Ares Fabregat-Hern\'andez (Valencian Research Institute for Artificial Intelligence, Universidad Internacional de Valencia), Javier Palanca (Valencian Research Institute for Artificial Intelligence), Vicent Botti (Valencian Research Institute for Artificial Intelligence, valgrAI)
Abstract: The paper introduces a novel framework based on category theory to enhance the explainability of artificial intelligence systems, particularly focusing on word embeddings. Key topics include the construction of categories $ \L_{T} $ and $ \P_{T} $, providing schematic representations of the semantics of a text $ T $, and reframing the selection of the element with maximum probability as a categorical notion. Additionally, the monoidal category $ \P_{T} $ is constructed to visualize various methods of extracting semantic information from $ T $, offering a dimension-agnostic definition of semantic spaces reliant solely on information within the text. Furthermore, the paper defines the categories of configurations $ \Conf $ and word embeddings $ \Emb $, accompanied by the concept of divergence as a decoration on $ \Emb $. It establishes a mathematically precise method for comparing word embeddings, demonstrating the equivalence between the GloVe and Word2Vec algorithms and the metric MDS algorithm, transitioning from neural network algorithms (black box) to a transparent framework. Finally, the paper presents a mathematical approach to computing biases before embedding and offers insights on mitigating biases at the semantic space level, advancing the field of explainable artificial intelligence.
Authors: Jessica Lundin, Guillaume Chabot-Couture
Abstract: We present a first known prototype of a dynamic, systematic benchmark of medical guidelines for 400+ questions, with 3.3+ trillion possible combinations, covering 100\% of guideline relationships. We transformed the WHO IMCI handbook into a directed graph with 200+ nodes (conditions, symptoms, treatments, follow-ups, severities) and 300+ edges, then used graph traversal to generate questions that incorporated age-specific scenarios and contextual distractors to ensure clinical relevance. Our graph-based approach enables systematic evaluation across clinical tasks (45-67\% accuracy), and we find models excel at symptom recognition but struggle with triaging severity, treatment protocols and follow-up care, demonstrating how customized benchmarks can identify specific capability gaps that general-domain evaluations miss. Beyond evaluation, this dynamic MCQA methodology enhances LLM post-training (supervised finetuning, GRPO, DPO), where correct answers provide high-reward samples without expensive human annotation. The graph-based approach successfully addresses the coverage limitations of manually curated benchmarks. This methodology is a step toward scalable, contamination-resistant solution for creating comprehensive benchmarks that can be dynamically generated, including when the guidelines are updated. Code and datasets are available at https://github.com/jessicalundin/graph_testing_harness
URLs: https://github.com/jessicalundin/graph_testing_harness
Authors: Huong Ngo, Matt Deitke, Martijn Bartelds, Sarah Pratt, Josh Gardner, Matt Jordan, Ludwig Schmidt
Abstract: Improvements in training data scale and quality have led to significant advances, yet its influence in speech recognition remains underexplored. In this paper, we present a large-scale dataset, OLMoASR-Pool, and series of models, OLMoASR, to study and develop robust zero-shot speech recognition models. Beginning from OLMoASR-Pool, a collection of 3M hours of English audio and 17M transcripts, we design text heuristic filters to remove low-quality or mistranscribed data. Our curation pipeline produces a new dataset containing 1M hours of high-quality audio-transcript pairs, which we call OLMoASR-Mix. We use OLMoASR-Mix to train the OLMoASR-Mix suite of models, ranging from 39M (tiny.en) to 1.5B (large.en) parameters. Across all model scales, OLMoASR achieves comparable average performance to OpenAI's Whisper on short and long-form speech recognition benchmarks. Notably, OLMoASR-medium.en attains a 12.8\% and 11.0\% word error rate (WER) that is on par with Whisper's largest English-only model Whisper-medium.en's 12.4\% and 10.5\% WER for short and long-form recognition respectively (at equivalent parameter count). OLMoASR-Pool, OLMoASR models, and filtering, training and evaluation code will be made publicly available to further research on robust speech processing.
Authors: Paritosh Parmar, Eric Peh, Basura Fernando
Abstract: Existing Causal-Why Video Question Answering (VideoQA) models often struggle with higher-order reasoning, relying on opaque, monolithic pipelines that entangle video understanding, causal inference, and answer generation. These black-box approaches offer limited interpretability and tend to depend on shallow heuristics. We propose a novel, modular framework that explicitly decouples causal reasoning from answer generation, introducing natural language causal chains as interpretable intermediate representations. Inspired by human cognitive models, these structured cause-effect sequences bridge low-level video content with high-level causal reasoning, enabling transparent and logically coherent inference. Our two-stage architecture comprises a Causal Chain Extractor (CCE) that generates causal chains from video-question pairs, and a Causal Chain-Driven Answerer (CCDA) that produces answers grounded in these chains. To address the lack of annotated reasoning traces, we introduce a scalable method for generating high-quality causal chains from existing datasets using large language models. We also propose CauCo, a new evaluation metric for causality-oriented captioning. Experiments on three large-scale benchmarks demonstrate that our approach not only outperforms state-of-the-art models, but also yields substantial gains in explainability, user trust, and generalization -- positioning the CCE as a reusable causal reasoning engine across diverse domains. Project page: https://paritoshparmar.github.io/chainreaction/
Authors: Orion Weller, Michael Boratko, Iftekhar Naim, Jinhyuk Lee
Abstract: Vector embeddings have been tasked with an ever-increasing set of retrieval tasks over the years, with a nascent rise in using them for reasoning, instruction-following, coding, and more. These new benchmarks push embeddings to work for any query and any notion of relevance that could be given. While prior works have pointed out theoretical limitations of vector embeddings, there is a common assumption that these difficulties are exclusively due to unrealistic queries, and those that are not can be overcome with better training data and larger models. In this work, we demonstrate that we may encounter these theoretical limitations in realistic settings with extremely simple queries. We connect known results in learning theory, showing that the number of top-k subsets of documents capable of being returned as the result of some query is limited by the dimension of the embedding. We empirically show that this holds true even if we restrict to k=2, and directly optimize on the test set with free parameterized embeddings. We then create a realistic dataset called LIMIT that stress tests models based on these theoretical results, and observe that even state-of-the-art models fail on this dataset despite the simple nature of the task. Our work shows the limits of embedding models under the existing single vector paradigm and calls for future research to develop methods that can resolve this fundamental limitation.
Authors: Arnav Arora, Lucie-Aim\'ee Kaffee, Isabelle Augenstein
Abstract: Language embeds information about social, cultural, and political values people hold. Prior work has explored social and potentially harmful biases encoded in Pre-Trained Language models (PTLMs). However, there has been no systematic study investigating how values embedded in these models vary across cultures. In this paper, we introduce probes to study which values across cultures are embedded in these models, and whether they align with existing theories and cross-cultural value surveys. We find that PTLMs capture differences in values across cultures, but those only weakly align with established value surveys. We discuss implications of using mis-aligned models in cross-cultural settings, as well as ways of aligning PTLMs with value surveys.
Authors: Zhihang Yuan, Yuzhang Shang, Yue Song, Dawei Yang, Qiang Wu, Yan Yan, Guangyu Sun
Abstract: In this paper, we introduce a new post-training compression paradigm for Large Language Models (LLMs) to facilitate their wider adoption. We delve into LLM weight low-rank decomposition, and find that the challenges of this task stem from (1) the distribution variance in the LLM activations and (2) the sensitivity difference among various kinds of layers. To address these issues, we propose a training-free approach called Activation-aware Singular Value Decomposition (ASVD). Specifically, ASVD manages activation outliers by transforming the weight matrix based on the activation distribution. This transformation allows the outliers in the activation matrix to be absorbed into the transformed weight matrix, thereby enhancing decomposition accuracy. Additionally, we propose an efficient iterative calibration process to optimize layer-specific decomposition by addressing the varying sensitivity of different LLM layers. In this way, ASVD can compress a network by 10%-30%. Based on the success of the low-rank decomposition of projection matrices in the self-attention module, we further introduce ASVD to compress the KV cache. By reducing the channel dimension of KV activations, memory requirements for KV cache can be largely reduced. ASVD can further achieve 50% KV cache reductions without performance drop in a training-free manner.
Authors: Juni Schindler, Sneha Jha, Xixuan Zhang, Kilian Buehling, Annett Heft, Mauricio Barahona
Abstract: We present Local Graph-based Dictionary Expansion (LGDE), a method for data-driven discovery of the semantic neighbourhood of words using tools from manifold learning and network science. At the heart of LGDE lies the creation of a word similarity graph from the geometry of word embeddings followed by local community detection based on graph diffusion. The diffusion in the local graph manifold allows the exploration of the complex nonlinear geometry of word embeddings to capture word similarities based on paths of semantic association, over and above direct pairwise similarities. Exploiting such semantic neighbourhoods enables the expansion of dictionaries of pre-selected keywords, an important step for tasks in information retrieval, such as database queries and online data collection. We validate LGDE on two user-generated English-language corpora and show that LGDE enriches the list of keywords with improved performance relative to methods based on direct word similarities or co-occurrences. We further demonstrate our method through a real-world use case from communication science, where LGDE is evaluated quantitatively on the expansion of a conspiracy-related dictionary from online data collected and analysed by domain experts. Our empirical results and expert user assessment indicate that LGDE expands the seed dictionary with more useful keywords due to the manifold-learning-based similarity network.
Authors: Dawid J. Kopiczko, Tijmen Blankevoort, Yuki M. Asano
Abstract: Decoder-only large language models typically rely solely on masked causal attention, which limits their expressiveness by restricting information flow to one direction. We propose Bitune, a method that enhances pretrained decoder-only LLMs by incorporating bidirectional attention into prompt processing. We evaluate Bitune in instruction-tuning and question-answering settings, showing significant improvements in performance on commonsense reasoning, arithmetic, and language understanding tasks. Furthermore, extensive ablation studies validate the role of each component of the method, and demonstrate that Bitune is compatible with various parameter-efficient finetuning techniques and full model finetuning.
Authors: Yuanchun Wang, Jifan Yu, Zijun Yao, Jing Zhang, Yuyang Xie, Shangqing Tu, Yiyang Fu, Youhe Feng, Jinkai Zhang, Jingyao Zhang, Bowen Huang, Yuanyao Li, Huihui Yuan, Lei Hou, Juanzi Li, Jie Tang
Abstract: Applying large language models (LLMs) for academic API usage shows promise in reducing researchers' academic information seeking efforts. However, current LLM API-using methods struggle with complex API coupling commonly encountered in academic queries. To address this, we introduce SoAy, a solution-based LLM API-using methodology for academic information seeking. It uses code with a solution as the reasoning method, where a solution is a pre-constructed API calling sequence. The addition of the solution reduces the difficulty for the model to understand the complex relationships between APIs. Code improves the efficiency of reasoning. To evaluate SoAy, we introduce SoAyBench, an evaluation benchmark accompanied by SoAyEval, built upon a cloned environment of APIs from AMiner. Experimental results demonstrate a 34.58-75.99\% performance improvement compared to state-of-the-art LLM API-based baselines. All datasets, codes, tuned models, and deployed online services are publicly accessible at https://github.com/RUCKBReasoning/SoAy.
Authors: Arief Purnama Muharram, Ayu Purwarianti
Abstract: Automated fact-checking is a key strategy to overcome the spread of COVID-19 misinformation on the internet. These systems typically leverage deep learning approaches through Natural Language Inference (NLI) to verify the truthfulness of information based on supporting evidence. However, one challenge that arises in deep learning is performance stagnation due to a lack of knowledge during training. This study proposes using a Knowledge Graph (KG) as external knowledge to enhance NLI performance for automated COVID-19 fact-checking in the Indonesian language. The proposed model architecture comprises three modules: a fact module, an NLI module, and a classifier module. The fact module processes information from the KG, while the NLI module handles semantic relationships between the given premise and hypothesis. The representation vectors from both modules are concatenated and fed into the classifier module to produce the final result. The model was trained using the generated Indonesian COVID-19 fact-checking dataset and the COVID-19 KG Bahasa Indonesia. Our study demonstrates that incorporating KGs can significantly improve NLI performance in fact-checking, achieving the best accuracy of 0.8616. This suggests that KGs are a valuable component for enhancing NLI performance in automated fact-checking.
Authors: Lucie Dvorackova, Marcin P. Joachimiak, Michal Cerny, Adriana Kubecova, Vilem Sklenak, Tomas Kliegr
Abstract: The best-performing approaches for scholarly document quality prediction are based on embedding models. In addition to their performance when used in classifiers, embedding models can also provide predictions even for words that were not contained in the labelled training data for the classification model, which is important in the context of the ever-evolving research terminology. Although model-agnostic explanation methods, such as Local interpretable model-agnostic explanations, can be applied to explain machine learning classifiers trained on embedding models, these produce results with questionable correspondence to the model. We introduce a new feature importance method, Self-model Rated Entities (SMER), for logistic regression-based classification models trained on word embeddings. We show that SMER has theoretically perfect fidelity with the explained model, as the average of logits of SMER scores for individual words (SMER explanation) exactly corresponds to the logit of the prediction of the explained model. Quantitative and qualitative evaluation is performed through five diverse experiments conducted on 50,000 research articles (papers) from the CORD-19 corpus. Through an AOPC curve analysis, we experimentally demonstrate that SMER produces better explanations than LIME, SHAP and global tree surrogates.
Authors: Youan Cong, Pritom Saha Akash, Cheng Wang, Kevin Chen-Chuan Chang
Abstract: We introduce the Extract-Refine-Retrieve-Read (ERRR) framework, a novel approach designed to bridge the pre-retrieval information gap in Retrieval-Augmented Generation (RAG) systems through query optimization tailored to meet the specific knowledge requirements of Large Language Models (LLMs). Unlike conventional query optimization techniques used in RAG, the ERRR framework begins by extracting parametric knowledge from LLMs, followed by using a specialized query optimizer for refining these queries. This process ensures the retrieval of only the most pertinent information essential for generating accurate responses. Moreover, to enhance flexibility and reduce computational costs, we propose a trainable scheme for our pipeline that utilizes a smaller, tunable model as the query optimizer, which is refined through knowledge distillation from a larger teacher model. Our evaluations on various question-answering (QA) datasets and with different retrieval systems show that ERRR consistently outperforms existing baselines, proving to be a versatile and cost-effective module for improving the utility and accuracy of RAG systems.
Authors: Hua Farn, Hsuan Su, Shachi H Kumar, Saurav Sahay, Shang-Tse Chen, Hung-yi Lee
Abstract: Fine-tuning large language models (LLMs) for downstream tasks often leads to catastrophic forgetting, notably degrading the safety of originally aligned models. While some existing methods attempt to restore safety by incorporating additional safety data, the quality of such data typically falls short of that used in the original alignment process. Moreover, these high-quality safety datasets are generally inaccessible, making it difficult to fully recover the model's original safety. We ask: How can we preserve safety while improving downstream task performance without additional safety data? We show that simply merging the weights of pre- and post-fine-tuned models effectively mitigates safety degradation while enhancing performance. Experiments across different downstream tasks and models validate the method's practicality and effectiveness.
Authors: Aloka Fernando, Nisansa de Silva, Menan Velyuthan, Charitha Rathnayake, Surangika Ranathunga
Abstract: Parallel Data Curation (PDC) techniques aim to filter out noisy parallel sentences from the web-mined corpora. Prior research has demonstrated that ranking sentence pairs using similarity scores on sentence embeddings derived from Pre-trained Multilingual Language Models (multiPLMs) and training the NMT systems with the top-ranked samples, produces superior NMT performance than when trained using the full dataset. However, previous research has shown that the choice of multiPLM significantly impacts the ranking quality. This paper investigates the reasons behind this disparity across multiPLMs. Using the web-mined corpora CCMatrix and CCAligned for En$\rightarrow$Si, En$\rightarrow$Ta and Si$\rightarrow$Ta, we show that different multiPLMs (LASER3, XLM-R, and LaBSE) are biased towards certain types of sentences, which allows noisy sentences to creep into the top-ranked samples. We show that by employing a series of heuristics, this noise can be removed to a certain extent. This results in improving the results of NMT systems trained with web-mined corpora and reduces the disparity across multiPLMs.
Authors: Michael Hanna, Yonatan Belinkov, Sandro Pezzelle
Abstract: Although large language models (LLMs) are increasingly capable, these capabilities are unevenly distributed: they excel at formal linguistic tasks, such as producing fluent, grammatical text, but struggle more with functional linguistic tasks like reasoning and consistent fact retrieval. Inspired by neuroscience, recent work suggests that to succeed on both formal and functional linguistic tasks, LLMs should use different mechanisms for each; such localization could either be built-in or emerge spontaneously through training. In this paper, we ask: do current models, with fast-improving functional linguistic abilities, exhibit distinct localization of formal and functional linguistic mechanisms? We answer this by finding and comparing the "circuits", or minimal computational subgraphs, responsible for various formal and functional tasks. Comparing 5 LLMs across 10 distinct tasks, we find that while there is indeed little overlap between circuits for formal and functional tasks, there is also little overlap between formal linguistic tasks, as exists in the human brain. Thus, a single formal linguistic network, unified and distinct from functional task circuits, remains elusive. However, in terms of cross-task faithfulness - the ability of one circuit to solve another's task - we observe a separation between formal and functional mechanisms, suggesting that shared mechanisms between formal tasks may exist.
Authors: Mihnea-Alexandru V\^irlan, R\u{a}zvan-Alexandru Sm\u{a}du, Dumitru-Clementin Cercel, Florin Pop, Mihaela-Claudia Cercel
Abstract: The primary goal of a news headline is to summarize an event in as few words as possible. Depending on the media outlet, a headline can serve as a means to objectively deliver a summary or improve its visibility. For the latter, specific publications may employ stylistic approaches that incorporate the use of sarcasm, irony, and exaggeration, key elements of a satirical approach. As such, even the headline must reflect the tone of the satirical main content. Current approaches for the Romanian language tend to detect the non-conventional tone (i.e., satire and clickbait) of the news content by combining both the main article and the headline. Because we consider a headline to be merely a brief summary of the main article, we investigate in this paper the presence of satirical tone in headlines alone, testing multiple baselines ranging from standard machine learning algorithms to deep learning models. Our experiments show that Bidirectional Transformer models outperform both standard machine-learning approaches and Large Language Models (LLMs), particularly when the meta-learning Reptile approach is employed.
Authors: Miguel Moura Ramos, Patrick Fernandes, Sweta Agrawal, Andr\'e F. T. Martins
Abstract: Large language models (LLMs) have demonstrated strong performance in sentence-level machine translation, but scaling to document-level translation remains challenging, particularly in modeling long-range dependencies and discourse phenomena across sentences and paragraphs. In this work, we propose a method to improve LLM-based long-document translation through targeted fine-tuning on high-quality document-level data, which we curate and introduce as DocBlocks. Our approach supports multiple translation paradigms, including direct document-to-document and chunk-level translation, by integrating instructions both with and without surrounding context. This enables models to better capture cross-sentence dependencies while maintaining strong sentence-level translation performance. Experimental results show that incorporating multiple translation paradigms improves document-level translation quality and inference speed compared to prompting and agent-based methods.
Authors: Nan Jiang, Ziming Wu, De-Chuan Zhan, Fuming Lai, Shaobing Lian
Abstract: Chain-of-Thought (CoT) reasoning has significantly advanced Large Language Models (LLMs) in solving complex tasks. However, its autoregressive paradigm leads to significant computational overhead, hindering its deployment in latency-sensitive applications. To address this, we propose \textbf{DART} (\textbf{D}istilling \textbf{A}utoregressive \textbf{R}easoning to Silent \textbf{T}hought), a self-distillation framework that enables LLMs to replace autoregressive CoT with non-autoregressive Silent Thought (ST). Specifically, DART introduces two training pathways: the CoT pathway for traditional reasoning and the ST pathway for generating answers directly from a few ST tokens. The ST pathway utilizes a lightweight Reasoning Evolvement Module (REM) to align its hidden states with the CoT pathway, enabling the ST tokens to evolve into informative embeddings. During inference, only the ST pathway is activated, leveraging evolving ST tokens to deliver the answer directly. Extensive experimental results demonstrate that DART offers significant performance gains compared with existing non-autoregressive baselines without extra inference latency, serving as a feasible alternative for efficient reasoning.
Authors: Yingji Zhang, Danilo S. Carvalho, Andr\'e Freitas
Abstract: Integrating compositional and symbolic properties into current distributional semantic spaces can enhance the interpretability, controllability, compositionality, and generalisation capabilities of Transformer-based auto-regressive language models (LMs). In this survey, we offer a novel perspective on latent space geometry through the lens of compositional semantics, a direction we refer to as \textit{semantic representation learning}. This direction enables a bridge between symbolic and distributional semantics, helping to mitigate the gap between them. We review and compare three mainstream autoencoder architectures-Variational AutoEncoder (VAE), Vector Quantised VAE (VQVAE), and Sparse AutoEncoder (SAE)-and examine the distinctive latent geometries they induce in relation to semantic structure and interpretability.
Authors: Younwoo Choi, Changling Li, Yongjin Yang, Zhijing Jin
Abstract: As large language models (LLMs) are increasingly integrated into multi-agent and human-AI systems, understanding their awareness of both self-context and conversational partners is essential for ensuring reliable performance and robust safety. While prior work has extensively studied situational awareness which refers to an LLM's ability to recognize its operating phase and constraints, it has largely overlooked the complementary capacity to identify and adapt to the identity and characteristics of a dialogue partner. In this paper, we formalize this latter capability as interlocutor awareness and present the first systematic evaluation of its emergence in contemporary LLMs. We examine interlocutor inference across three dimensions-reasoning patterns, linguistic style, and alignment preferences-and show that LLMs reliably identify same-family peers and certain prominent model families, such as GPT and Claude. To demonstrate its practical significance, we develop three case studies in which interlocutor awareness both enhances multi-LLM collaboration through prompt adaptation and introduces new alignment and safety vulnerabilities, including reward-hacking behaviors and increased jailbreak susceptibility. Our findings highlight the dual promise and peril of identity-sensitive behavior in LLMs, underscoring the need for further understanding of interlocutor awareness and new safeguards in multi-agent deployments. Our code is open-sourced at https://github.com/younwoochoi/InterlocutorAwarenessLLM.
URLs: https://github.com/younwoochoi/InterlocutorAwarenessLLM.
Authors: Kureha Yamaguchi, Benjamin Etheridge, Andy Arditi
Abstract: Reasoning models generate chain-of-thought (CoT) tokens before their final output, but how this affects their vulnerability to jailbreak attacks remains unclear. While traditional language models make refusal decisions at the prompt-response boundary, we find evidence that DeepSeek-R1-Distill-Llama-8B makes these decisions within its CoT generation. We identify a linear direction in activation space during CoT token generation that predicts whether the model will refuse or comply -- termed the "caution" direction because it corresponds to cautious reasoning patterns in the generated text. Ablating this direction from model activations increases harmful compliance, effectively jailbreaking the model. We additionally show that intervening only on CoT token activations suffices to control final outputs, and that incorporating this direction into prompt-based attacks improves success rates. Our findings suggest that the chain-of-thought itself is a promising new target for adversarial manipulation in reasoning models. Code available at https://github.com/ky295/reasoning-manipulation.
Authors: Yizhan Huang, Zhe Yang, Meifang Chen, Jianping Zhang, Michael R. Lyu
Abstract: Large Language Models (LLMs) are known to memorize portions of their training data, sometimes reproducing content verbatim when prompted appropriately. In this work, we investigate a fundamental yet under-explored question in the domain of memorization: How to characterize memorization difficulty of training data in LLMs? Through empirical experiments on OLMo, a family of open models, we present the Entropy-Memorization Law. It suggests that data entropy is linearly correlated with memorization score. Moreover, in a case study of memorizing highly randomized strings, or "gibberish", we observe that such sequences, despite their apparent randomness, exhibit unexpectedly low empirical entropy compared to the broader training corpus. Adopting the same strategy to discover Entropy-Memorization Law, we derive a simple yet effective approach to distinguish training and testing data, enabling Dataset Inference (DI).
Authors: Shuyu Guo, Zhaochun Ren
Abstract: Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge but incurs significant inference costs due to lengthy retrieved contexts. While context compression mitigates this issue, existing methods apply fixed compression rates, over-compressing simple queries or under-compressing complex ones. We propose Adaptive Context Compression for RAG (ACC-RAG), a framework that dynamically adjusts compression rates based on input complexity, optimizing inference efficiency without sacrificing accuracy. ACC-RAG combines a hierarchical compressor (for multi-granular embeddings) with a context selector to retain minimal sufficient information, akin to human skimming. Evaluated on Wikipedia and five QA datasets, ACC-RAG outperforms fixed-rate methods and matches/unlocks over 4 times faster inference versus standard RAG while maintaining or improving accuracy.
Authors: Sri Durga Sai Sowmya Kadali, Evangelos E. Papalexakis
Abstract: The widespread use of Large Language Models (LLMs) in many applications marks a significant advance in research and practice. However, their complexity and hard-to-understand nature make them vulnerable to attacks, especially jailbreaks designed to produce harmful responses. To counter these threats, developing strong detection methods is essential for the safe and reliable use of LLMs. This paper studies this detection problem using the Contextual Co-occurrence Matrix, a structure recognized for its efficacy in data-scarce environments. We propose a novel method leveraging the latent space characteristics of Contextual Co-occurrence Matrices and Tensors for the effective identification of adversarial and jailbreak prompts. Our evaluations show that this approach achieves a notable F1 score of 0.83 using only 0.5% of labeled prompts, which is a 96.6% improvement over baselines. This result highlights the strength of our learned patterns, especially when labeled data is scarce. Our method is also significantly faster, speedup ranging from 2.3 to 128.4 times compared to the baseline models.
Authors: Ryan Wong, Jiawei Wang, Junjie Zhao, Li Chen, Yan Gao, Long Zhang, Xuan Zhou, Zuo Wang, Kai Xiang, Ge Zhang, Wenhao Huang, Yang Wang, Ke Wang
Abstract: From professional research to everyday planning, many tasks are bottlenecked by wide-scale information seeking, which is more repetitive than cognitively complex. With the rapid development of Large Language Models (LLMs), automated search agents powered by LLMs offer a promising solution to liberate humans from this tedious work. However, the capability of these agents to perform such "wide-context" collection reliably and completely remains largely unevaluated due to a lack of suitable benchmarks. To bridge this gap, we introduce WideSearch, a new benchmark engineered to evaluate agent reliability on these large-scale collection tasks. The benchmark features 200 manually curated questions (100 in English, 100 in Chinese) from over 15 diverse domains, grounded in real user queries. Each task requires agents to collect large-scale atomic information, which could be verified one by one objectively, and arrange it into a well-organized output. A rigorous five-stage quality control pipeline ensures the difficulty, completeness, and verifiability of the dataset. We benchmark over 10 state-of-the-art agentic search systems, including single-agent, multi-agent frameworks, and end-to-end commercial systems. Most systems achieve overall success rates near 0\%, with the best performer reaching just 5\%. However, given sufficient time, cross-validation by multiple human testers can achieve a near 100\% success rate. These results demonstrate that present search agents have critical deficiencies in large-scale information seeking, underscoring urgent areas for future research and development in agentic search. Our dataset, evaluation pipeline, and benchmark results have been publicly released at https://widesearch-seed.github.io/
Authors: Afrozah Nadeem, Mark Dras, Usman Naseem
Abstract: Recent advancements in large language models (LLMs) have enabled their widespread use across diverse real-world applications. However, concerns remain about their tendency to encode and reproduce ideological biases along political and economic dimensions. In this paper, we employ a framework for probing and mitigating such biases in decoder-based LLMs through analysis of internal model representations. Grounded in the Political Compass Test (PCT), this method uses contrastive pairs to extract and compare hidden layer activations from models like Mistral and DeepSeek. We introduce a comprehensive activation extraction pipeline capable of layer-wise analysis across multiple ideological axes, revealing meaningful disparities linked to political framing. Our results show that decoder LLMs systematically encode representational bias across layers, which can be leveraged for effective steering vector-based mitigation. This work provides new insights into how political bias is encoded in LLMs and offers a principled approach to debiasing beyond surface-level output interventions.
Authors: Lorenzo Proietti, Stefano Perrella, Vil\'em Zouhar, Roberto Navigli, Tom Kocmi
Abstract: Machine translation quality has steadily improved over the years, achieving near-perfect translations in recent benchmarks. These high-quality outputs make it difficult to distinguish between state-of-the-art models and to identify areas for future improvement. In this context, automatically identifying texts where machine translation systems struggle holds promise for developing more discriminative evaluations and guiding future research. In this work, we address this gap by formalizing the task of translation difficulty estimation, defining a text's difficulty based on the expected quality of its translations. We introduce a new metric to evaluate difficulty estimators and use it to assess both baselines and novel approaches. Finally, we demonstrate the practical utility of difficulty estimators by using them to construct more challenging benchmarks for machine translation. Our results show that dedicated models outperform both heuristic-based methods and LLM-as-a-judge approaches, with Sentinel-src achieving the best performance. Thus, we release two improved models for difficulty estimation, Sentinel-src-24 and Sentinel-src-25, which can be used to scan large collections of texts and select those most likely to challenge contemporary machine translation systems.
Authors: Carter Blum, Katja Filippova, Ann Yuan, Asma Ghandeharioun, Julian Zimmert, Fred Zhang, Jessica Hoffmann, Tal Linzen, Martin Wattenberg, Lucas Dixon, Mor Geva
Abstract: Large language models (LLMs) struggle with cross-lingual knowledge transfer: they hallucinate when asked in one language about facts expressed in a different language during training. This work introduces a controlled setting to study the causes and dynamics of this phenomenon by training small Transformer models from scratch on synthetic multilingual datasets. We identify a learning phase wherein a model develops either separate or unified representations of the same facts across languages, and show that unification is essential for cross-lingual transfer. We also show that the degree of unification depends on mutual information between facts and training data language, and on how easy it is to extract that language. Based on these insights, we develop methods to modulate the level of cross-lingual transfer by manipulating data distribution and tokenization, and we introduce metrics and visualizations to formally characterize their effects on unification. Our work shows how controlled settings can shed light on pre-training dynamics and suggests new directions for improving cross-lingual transfer in LLMs.
Authors: Weiting Tan, Jiachen Lian, Hirofumi Inaguma, Paden Tomasello, Philipp Koehn, Xutai Ma
Abstract: We present an Audio-Visual Language Model (AVLM) for expressive speech generation by integrating full-face visual cues into a pre-trained expressive speech model. We explore multiple visual encoders and multimodal fusion strategies during pre-training to identify the most effective integration approach. Subsequent fine-tuning on emotion recognition and expressive dialogue tasks yields substantial gains over speech-only baselines (e.g., +5 F1 in emotion recognition). AVLM highlights the value of expressive visual information in guiding speech generation and offers a foundation for end-to-end multimodal conversational systems.
Authors: Khaoula Chehbouni, Mohammed Haddou, Jackie Chi Kit Cheung, Golnoosh Farnadi
Abstract: Evaluating natural language generation (NLG) systems remains a core challenge of natural language processing (NLP), further complicated by the rise of large language models (LLMs) that aims to be general-purpose. Recently, large language models as judges (LLJs) have emerged as a promising alternative to traditional metrics, but their validity remains underexplored. This position paper argues that the current enthusiasm around LLJs may be premature, as their adoption has outpaced rigorous scrutiny of their reliability and validity as evaluators. Drawing on measurement theory from the social sciences, we identify and critically assess four core assumptions underlying the use of LLJs: their ability to act as proxies for human judgment, their capabilities as evaluators, their scalability, and their cost-effectiveness. We examine how each of these assumptions may be challenged by the inherent limitations of LLMs, LLJs, or current practices in NLG evaluation. To ground our analysis, we explore three applications of LLJs: text summarization, data annotation, and safety alignment. Finally, we highlight the need for more responsible evaluation practices in LLJs evaluation, to ensure that their growing role in the field supports, rather than undermines, progress in NLG.
Authors: Maojia Song, Tej Deep Pala, Weisheng Jin, Amir Zadeh, Chuan Li, Dorien Herremans, Soujanya Poria
Abstract: Large language models (LLMs) are increasingly deployed in multi-agent systems (MAS) as components of collaborative intelligence, where peer interactions dynamically shape individual decision-making. Although prior work has focused on conformity bias, we extend the analysis to examine how LLMs form trust from previous impressions, resist misinformation, and integrate peer input during interaction, key factors for achieving collective intelligence under complex social dynamics. We present KAIROS, a benchmark simulating quiz contests with peer agents of varying reliability, offering fine-grained control over conditions such as expert-novice roles, noisy crowds, and adversarial peers. LLMs receive both historical interactions and current peer responses, allowing systematic investigation into how trust, peer action, and self-confidence influence decisions. As for mitigation strategies, we evaluate prompting, supervised fine-tuning, and reinforcement learning, Group Relative Policy Optimisation (GRPO), across multiple models. Our results reveal that GRPO with multi-agent context combined with outcome-based rewards and unconstrained reasoning achieves the best overall performance, but also decreases the robustness to social influence compared to Base models. The code and datasets are available at: https://github.com/declare-lab/KAIROS.
Authors: Yilin Wang, Heng Wang, Yuyang Bai, Minnan Luo
Abstract: In Large Language Models (LLMs) generation, there exist knowledge conflicts and scenarios where parametric knowledge contradicts knowledge provided in the context. Previous works studied tuning, decoding algorithms, or locating and editing context-aware neurons to adapt LLMs to be faithful to new contextual knowledge. However, they are usually inefficient or ineffective for large models, not workable for black-box models, or unable to continuously adjust LLMs' sensitivity to the knowledge provided in the context. To mitigate these problems, we propose CSKS (Continuously Steering Knowledge Sensitivity), a simple framework that can steer LLMs' sensitivity to contextual knowledge continuously at a lightweight cost. Specifically, we tune two small LMs (i.e. proxy models) and use the difference in their output distributions to shift the original distribution of an LLM without modifying the LLM weights. In the evaluation process, we not only design synthetic data and fine-grained metrics to measure models' sensitivity to contextual knowledge but also use a real conflict dataset to validate CSKS's practical efficacy. Extensive experiments demonstrate that our framework achieves continuous and precise control over LLMs' sensitivity to contextual knowledge, enabling both increased sensitivity and reduced sensitivity, thereby allowing LLMs to prioritize either contextual or parametric knowledge as needed flexibly. Our data and code are available at https://github.com/OliveJuiceLin/CSKS.
Authors: Aritra Dutta, Swapnanil Mukherjee, Deepanway Ghosal, Somak Aditya
Abstract: Commonsense visual-question answering often hinges on knowledge that is missing from the image or the question. Small vision-language models (sVLMs) such as ViLT, VisualBERT and FLAVA therefore lag behind their larger generative counterparts. To study the effect of careful commonsense knowledge integration on sVLMs, we present an end-to-end framework (NLKI) that (i) retrieves natural language facts, (ii) prompts an LLM to craft natural language explanations, and (iii) feeds both signals to sVLMs respectively across two commonsense VQA datasets (CRIC, AOKVQA) and a visual-entailment dataset (e-SNLI-VE). Facts retrieved using a fine-tuned ColBERTv2 and an object information-enriched prompt yield explanations that largely cut down hallucinations, while lifting the end-to-end answer accuracy by up to 7% (across 3 datasets), making FLAVA and other models in NLKI match or exceed medium-sized VLMs such as Qwen-2 VL-2B and SmolVLM-2.5B. As these benchmarks contain 10-25% label noise, additional finetuning using noise-robust losses (such as symmetric cross entropy and generalised cross entropy) adds another 2.5% in CRIC, and 5.5% in AOKVQA. Our findings expose when LLM-based commonsense knowledge beats retrieval from commonsense knowledge bases, how noise-aware training stabilises small models in the context of external knowledge augmentation, and why parameter-efficient commonsense reasoning is now within reach for 250M models.
Authors: Boheng Mao
Abstract: Legal text classification is a fundamental NLP task in the legal domain. Benchmark datasets in this area often exhibit a long-tail label distribution, where many labels are underrepresented, leading to poor model performance on rare classes. This paper proposes Selective Retrieval-Augmentation (SRA) as a solution to this problem. SRA focuses on augmenting samples belonging to low-frequency labels in the training set, preventing the introduction of noise for well-represented classes, and requires no changes to the model architecture. Retrieval is performed only from the training data to ensure there is no potential information leakage, removing the need for external corpora simultaneously. The proposed SRA method is tested on two legal text classification benchmark datasets with long-tail distributions: LEDGAR (single-label) and UNFAIR-ToS (multi-label). The results indicate that SRA attains higher micro-F1 and macro-F1 scores compared to all current LexGLUE baselines across both datasets, illustrating consistent improvements in long-tail legal text classification.
Authors: Sheng Liu, Qiang Sheng, Danding Wang, Yang Li, Guang Yang, Juan Cao
Abstract: Despite advances in improving large language model (LLM) to refuse to answer malicious instructions, widely used LLMs remain vulnerable to jailbreak attacks where attackers generate instructions with distributions differing from safety alignment corpora. New attacks expose LLMs' inability to recognize unseen malicious instructions, highlighting a critical distributional mismatch between training data and real-world attacks that forces developers into reactive patching cycles. To tackle this challenge, we propose IMAGINE, a synthesis framework that leverages embedding space distribution analysis to generate jailbreak-like instructions. This approach effectively fills the distributional gap between authentic jailbreak patterns and safety alignment corpora. IMAGINE follows an iterative optimization process that dynamically evolves text generation distributions across iterations, thereby augmenting the coverage of safety alignment data distributions through synthesized data examples. Based on the safety-aligned corpus enhanced through IMAGINE, our framework demonstrates significant decreases in attack success rate on Qwen2.5, Llama3.1, and Llama3.2 without compromising their utility.
Authors: Hassan Alhuzali, Farah Shamout, Muhammad Abdul-Mageed, Chaimae Abouzahir, Mouath Abu-Daoud, Ashwag Alasmari, Walid Al-Eisawi, Renad Al-Monef, Ali Alqahtani, Lama Ayash, Nizar Habash, Leen Kharouf
Abstract: We introduce {AraHealthQA 2025}, the {Comprehensive Arabic Health Question Answering Shared Task}, held in conjunction with {ArabicNLP 2025} (co-located with EMNLP 2025). This shared task addresses the paucity of high-quality Arabic medical QA resources by offering two complementary tracks: {MentalQA}, focusing on Arabic mental health Q\&A (e.g., anxiety, depression, stigma reduction), and {MedArabiQ}, covering broader medical domains such as internal medicine, pediatrics, and clinical decision making. Each track comprises multiple subtasks, evaluation datasets, and standardized metrics, facilitating fair benchmarking. The task was structured to promote modeling under realistic, multilingual, and culturally nuanced healthcare contexts. We outline the dataset creation, task design and evaluation framework, participation statistics, baseline systems, and summarize the overall outcomes. We conclude with reflections on the performance trends observed and prospects for future iterations in Arabic health QA.
Authors: Sourav Saha, Debapriyo Majumdar, Mandar Mitra
Abstract: Deep Learning and Machine Learning based models have become extremely popular in text processing and information retrieval. However, the non-linear structures present inside the networks make these models largely inscrutable. A significant body of research has focused on increasing the transparency of these models. This article provides a broad overview of research on the explainability and interpretability of natural language processing and information retrieval methods. More specifically, we survey approaches that have been applied to explain word embeddings, sequence modeling, attention modules, transformers, BERT, and document ranking. The concluding section suggests some possible directions for future research on this topic.
Authors: Jeongkyun Park, Jung-Wook Hwang, Kwanghee Choi, Seung-Hyun Lee, Jun Hwan Ahn, Rae-Hong Park, Hyung-Min Park
Abstract: Inspired by humans comprehending speech in a multi-modal manner, various audio-visual datasets have been constructed. However, most existing datasets focus on English, induce dependencies with various prediction models during dataset preparation, and have only a small number of multi-view videos. To mitigate the limitations, we recently developed the Open Large-scale Korean Audio-Visual Speech (OLKAVS) dataset, which is the largest among publicly available audio-visual speech datasets. The dataset contains 1,150 hours of transcribed audio from 1,107 Korean speakers in a studio setup with nine different viewpoints and various noise situations. We also provide the pre-trained baseline models for two tasks, audio-visual speech recognition and lip reading. We conducted experiments based on the models to verify the effectiveness of multi-modal and multi-view training over uni-modal and frontal-view-only training. We expect the OLKAVS dataset to facilitate multi-modal research in broader areas such as Korean speech recognition, speaker recognition, pronunciation level classification, and mouth motion analysis.
Authors: Marios Papachristou, Yuan Yuan
Abstract: Social networks profoundly influence how humans form opinions, exchange information, and organize collectively. As large language models (LLMs) are increasingly embedded into social and professional environments, it is critical to understand whether their interactions approximate human-like network dynamics. We develop a framework to study the network formation behaviors of multiple LLM agents and benchmark them against human decisions. Across synthetic and real-world settings, including friendship, telecommunication, and employment networks, we find that LLMs consistently reproduce fundamental micro-level principles such as preferential attachment, triadic closure, and homophily, as well as macro-level properties including community structure and small-world effects. Importantly, the relative emphasis of these principles adapts to context: for example, LLMs favor homophily in friendship networks but heterophily in organizational settings, mirroring patterns of social mobility. A controlled human-subject survey confirms strong alignment between LLMs and human participants in link-formation decisions. These results establish that LLMs can serve as powerful tools for social simulation and synthetic data generation, while also raising critical questions about bias, fairness, and the design of AI systems that participate in human networks.
Authors: Haorui He, Yuchen Song, Yuancheng Wang, Haoyang Li, Xueyao Zhang, Li Wang, Gongping Huang, Eng Siong Chng, Zhizheng Wu
Abstract: The effectiveness of one-shot voice conversion (VC) decreases in real-world scenarios where reference speeches, which are often sourced from the internet, contain various disturbances like background noise. To address this issue, we introduce Noro, a noise-robust one-shot VC system. Noro features innovative components tailored for VC using noisy reference speeches, including a dual-branch reference encoding module and a noise-agnostic contrastive speaker loss. Experimental results demonstrate that Noro outperforms our baseline system in both clean and noisy scenarios, highlighting its efficacy for real-world applications. Additionally, we investigate the hidden speaker representation capabilities of our baseline system by repurposing its reference encoder as a speaker encoder. The results show that it is competitive with several advanced self-supervised learning models for speaker representation under the SUPERB settings, highlighting the potential for advancing speaker representation learning through one-shot VC tasks.
Authors: Hao Cheng, Erjia Xiao, Yichi Wang, Lingfeng Zhang, Qiang Zhang, Jiahang Cao, Kaidi Xu, Mengshu Sun, Xiaoshuai Hao, Jindong Gu, Renjing Xu
Abstract: Current Cross-Modality Generation Models (GMs) demonstrate remarkable capabilities in various generative tasks. Given the ubiquity and information richness of vision modality inputs in real-world scenarios, Cross-Vision tasks, encompassing Vision-Language Perception (VLP) and Image-to-Image (I2I), have attracted significant attention. Large Vision Language Models (LVLMs) and I2I Generation Models (GMs) are employed to handle VLP and I2I tasks, respectively. Previous research indicates that printing typographic words into input images significantly induces LVLMs and I2I GMs to produce disruptive outputs that are semantically aligned with those words. Additionally, visual prompts, as a more sophisticated form of typography, are also revealed to pose security risks to various applications of cross-vision tasks. However, the specific characteristics of the threats posed by visual prompts remain underexplored. In this paper, to comprehensively investigate the performance impact induced by Typographic Visual Prompt Injection (TVPI) in various LVLMs and I2I GMs, we propose the Typographic Visual Prompts Injection Dataset and thoroughly evaluate the TVPI security risks on various open-source and closed-source LVLMs and I2I GMs under visual prompts with different target semantics, deepening the understanding of TVPI threats.
Authors: Diogo Freitas, Brigt H{\aa}vardstun, C\`esar Ferri, Dar\'io Garigliotti, Jan Arne Telle, Jos\'e Hern\'andez-Orallo
Abstract: Large language models have become multimodal, and many of them are said to integrate their modalities using common representations. If this were true, a drawing of a car as an image, for instance, should map to a similar area in the latent space as a textual description of the strokes that form the drawing. To explore this in a black-box access regime to these models, we propose the use of machine teaching, a theory that studies the minimal set of examples a teacher needs to choose so that the learner captures the concept. In this paper, we evaluate the complexity of teaching vision-language models a subset of objects in the Quick, Draw! dataset using two presentations: raw images as bitmaps and trace coordinates in TikZ format. The results indicate that image-based representations generally require fewer segments and achieve higher accuracy than coordinate-based representations. But, surprisingly, the teaching size usually ranks concepts similarly across both modalities, even when controlling for (a human proxy of) concept priors, suggesting that the simplicity of concepts may be an inherent property that transcends modality representations.
Authors: Jinyuan Feng, Chaopeng Wei, Tenghai Qiu, Tianyi Hu, Zhiqiang Pu
Abstract: In parameter-efficient fine-tuning, mixture-of-experts (MoE), which involves specializing functionalities into different experts and sparsely activating them appropriately, has been widely adopted as a promising approach to trade-off between model capacity and computation overhead. However, current MoE variants fall short on heterogeneous datasets, ignoring the fact that experts may learn similar knowledge, resulting in the underutilization of MoE's capacity. In this paper, we propose Contrastive Representation for MoE (CoMoE), a novel method to promote modularization and specialization in MoE, where the experts are trained along with a contrastive objective by sampling from activated and inactivated experts in top-k routing. We demonstrate that such a contrastive objective recovers the mutual-information gap between inputs and the two types of experts. Experiments on several benchmarks and in multi-task settings demonstrate that CoMoE can consistently enhance MoE's capacity and promote modularization among the experts.
Authors: Yunqing Liu, Wenqi Fan, Xiaoyong Wei, Qing Li
Abstract: Proteins are central to biological systems, participating as building blocks across all forms of life. Despite advancements in understanding protein functions through protein sequence analysis, there remains potential for further exploration in integrating protein structural information. We argue that the structural information of proteins is not only limited to their 3D information but also encompasses information from amino acid molecules (local information) to protein-protein structure similarity (global information). To address this, we propose \textbf{GLProtein}, the first framework in protein pre-training that incorporates both global structural similarity and local amino acid details to enhance prediction accuracy and functional insights. GLProtein innovatively combines protein-masked modelling with triplet structure similarity scoring, protein 3D distance encoding and substructure-based amino acid molecule encoding. Experimental results demonstrate that GLProtein outperforms previous methods in several bioinformatics tasks, including predicting protein-protein interaction, contact prediction, and so on.
Authors: Mashiro Toyooka, Kiyoharu Aizawa, Yoko Yamakata
Abstract: Large Language Models (LLMs) are trained on a vast amount of procedural texts, but they do not directly observe real-world phenomena. In the context of cooking recipes, this poses a challenge, as intermediate states of ingredients are often omitted, making it difficult for models to track ingredient states and understand recipes accurately. In this paper, we apply state probing, a method for evaluating a language model's understanding of the world, to the domain of cooking. We propose a new task and dataset for evaluating how well LLMs can recognize intermediate ingredient states during cooking procedures. We first construct a new Japanese recipe dataset with clear and accurate annotations of ingredient state changes, collected from well-structured and controlled recipe texts. Using this dataset, we design three novel tasks to evaluate whether LLMs can track ingredient state transitions and identify ingredients present at intermediate steps. Our experiments with widely used LLMs, such as Llama3.1-70B and Qwen2.5-72B, show that learning ingredient state knowledge improves their understanding of cooking processes, achieving performance comparable to commercial LLMs. The dataset are publicly available at: https://huggingface.co/datasets/mashi6n/nhkrecipe-100-anno-1
URLs: https://huggingface.co/datasets/mashi6n/nhkrecipe-100-anno-1
Authors: Yicheng Ji, Jun Zhang, Heming Xia, Jinpeng Chen, Lidan Shou, Gang Chen, Huan Li
Abstract: Video large language models (Vid-LLMs) have shown strong capabilities in understanding video content. However, their reliance on dense video token representations introduces substantial memory and computational overhead in both prefilling and decoding. To mitigate the information loss of recent video token reduction methods and accelerate the decoding stage of Vid-LLMs losslessly, we introduce SpecVLM, a training-free speculative decoding (SD) framework tailored for Vid-LLMs that incorporates staged video token pruning. Building on our novel finding that the draft model's speculation exhibits low sensitivity to video token pruning, SpecVLM prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy. To achieve this, we performs a two-stage pruning process: Stage I selects highly informative tokens guided by attention signals from the verifier (target model), while Stage II prunes remaining redundant ones in a spatially uniform manner. Extensive experiments on four video understanding benchmarks demonstrate the effectiveness and robustness of SpecVLM, which achieves up to 2.68$\times$ decoding speedup for LLaVA-OneVision-72B and 2.11$\times$ speedup for Qwen2.5-VL-32B. Code is available at https://github.com/zju-jiyicheng/SpecVLM.
Authors: Mosh Levy, Zohar Elyoseph, Yoav Goldberg
Abstract: A new generation of AI models generates step-by-step reasoning text before producing an answer. This text appears to offer a human-readable window into their computation process, and is increasingly relied upon for transparency and interpretability. However, it is unclear whether human understanding of this text matches the model's actual computational process. In this paper, we investigate a necessary condition for correspondence: the ability of humans to identify which steps in a reasoning text causally influence later steps. We evaluated humans on this ability by composing questions based on counterfactual measurements and found a significant discrepancy: participant accuracy was only 29%, barely above chance (25%), and remained low (42%) even when evaluating the majority vote on questions with high agreement. Our results reveal a fundamental gap between how humans interpret reasoning texts and how models use it, challenging its utility as a simple interpretability tool. We argue that reasoning texts should be treated as an artifact to be investigated, not taken at face value, and that understanding the non-human ways these models use language is a critical research direction.
Authors: Xinxing Ren, Caelum Forder, Qianbo Zang, Ahsen Tahir, Roman J. Georgio, Suman Deb, Peter Carroll, \"Onder G\"urcan, Zekun Guo
Abstract: Recent advances in generalist multi-agent systems (MAS) have largely followed a context-engineering plus centralized paradigm, where a planner agent coordinates multiple worker agents through unidirectional prompt passing. While effective under strong planner models, this design suffers from two critical limitations: (1) strong dependency on the planner's capability, which leads to degraded performance when a smaller LLM powers the planner; and (2) limited inter-agent communication, where collaboration relies on costly prompt concatenation and context injection, introducing redundancy and information loss. To address these challenges, we propose Anemoi, a semi-centralized MAS built on the Agent-to-Agent (A2A) communication MCP server from Coral Protocol. Unlike traditional designs, Anemoi enables structured and direct inter-agent collaboration, allowing all agents to monitor progress, assess results, identify bottlenecks, and propose refinements in real time. This paradigm reduces reliance on a single planner, supports adaptive plan updates, and minimizes redundant context passing, resulting in more scalable and cost-efficient execution. Evaluated on the GAIA benchmark, Anemoi achieved 52.73% accuracy with a small LLM (GPT-4.1-mini) as the planner, surpassing the strongest open-source baseline OWL (43.63%) by +9.09% under identical LLM settings. Our implementation is publicly available at https://github.com/Coral-Protocol/Anemoi.
Authors: Jianxing Liao, Tian Zhang, Xiao Feng, Yusong Zhang, Rui Yang, Haorui Wang, Bosi Wen, Ziying Wang, Runzhi Shi
Abstract: Large language models are extensively utilized in creative writing applications. Creative writing requires a balance between subjective writing quality (e.g., literariness and emotional expression) and objective constraint following (e.g., format requirements and word limits). Existing methods find it difficult to balance these two aspects: single reward strategies fail to improve both abilities simultaneously, while fixed-weight mixed-reward methods lack the ability to adapt to different writing scenarios. To address this problem, we propose Reinforcement Learning with Mixed Rewards (RLMR), utilizing a dynamically mixed reward system from a writing reward model evaluating subjective writing quality and a constraint verification model assessing objective constraint following. The constraint following reward weight is adjusted dynamically according to the writing quality within sampled groups, ensuring that samples violating constraints get negative advantage in GRPO and thus penalized during training, which is the key innovation of this proposed method. We conduct automated and manual evaluations across diverse model families from 8B to 72B parameters. Additionally, we construct a real-world writing benchmark named WriteEval for comprehensive evaluation. Results illustrate that our method achieves consistent improvements in both instruction following (IFEval from 83.36% to 86.65%) and writing quality (72.75% win rate in manual expert pairwise evaluations on WriteEval). To the best of our knowledge, RLMR is the first work to combine subjective preferences with objective verification in online RL training, providing an effective solution for multi-dimensional creative writing optimization.
Authors: Xinran Zhao, Boyuan Zheng, Chenglei Si, Haofei Yu, Ken Liu, Runlong Zhou, Ruochen Li, Tong Chen, Xiang Li, Yiming Zhang, Tongshuang Wu
Abstract: This paper revisits Ramon Llull's Ars combinatoria - a medieval framework for generating knowledge through symbolic recombination - as a conceptual foundation for building a modern Llull's thinking machine for research ideation. Our approach defines three compositional axes: Theme (e.g., efficiency, adaptivity), Domain (e.g., question answering, machine translation), and Method (e.g., adversarial training, linear attention). These elements represent high-level abstractions common in scientific work - motivations, problem settings, and technical approaches - and serve as building blocks for LLM-driven exploration. We mine elements from human experts or conference papers and show that prompting LLMs with curated combinations produces research ideas that are diverse, relevant, and grounded in current literature. This modern thinking machine offers a lightweight, interpretable tool for augmenting scientific creativity and suggests a path toward collaborative ideation between humans and AI.