Authors: Michael Y. Hu, Apurva Gandhi, Kyunghyun Cho, Tal Linzen, Pratyusha Sharma
Abstract: Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.
Authors: Sarah Griebel, Ted Underwood
Abstract: Literary translation requires balancing target-language fluency with faithfulness to the source. Recent large language models (LLMs) often produce fluent translations, but it remains unclear whether fluency corresponds to semantic preservation in literary text. We examine this relationship using 130,486 translated paragraphs from 106 novels in 16 source languages, including human, Google Translate, and TranslateGemma translations. Fluency is measured as original-likeness with a translationese classifier trained on paragraph part-of-speech n-grams, and faithfulness with the automatic translation evaluation metric COMET-KIWI. We control for paragraph length and find a consistent negative correlation between fluency and faithfulness. The pattern appears for both human and Google Translate, but is weaker and often non-significant for TranslateGemma. These results show that segment length matters for automatic evaluation and suggest a tradeoff between fluency and faithfulness in literary translation.
Authors: Amir Zeldes
Abstract: The relations connecting propositions in discourse such as cause (A because B) or concession (A although B) are a subject of intense interest in Computational Linguistics and Pragmatics, but challenging to study and compare across languages. Recent progress in standardizing discourse relation inventories across datasets offers the potential to facilitate such studies, but is hindered by the complexity of relevant data and the lack of easily accessible interfaces to analyze it. In this paper we present DiscoExplorer, a new open source web interface, capable of running on local computers, which we use to make datasets from the DISRPT Shared Task on discourse relation classification publicly available, covering 16 different languages. We present the query language, search and visualization facilities for relations and signaling devices such as connectives, as well as some example studies.
Authors: Volodymyr Ovcharov
Abstract: Half a billion citation edges extracted from 100.7 million Ukrainian court decisions reveal that judicial citation structure encodes legal domain boundaries without supervision and predicts future legislative importance with near-perfect accuracy. We construct the first large-scale citation graph from the complete EDRSR registry (99.5 million full texts, 1.1 TB), extracting 502 million citation links across six types via regex on commodity hardware in approximately 5 hours, with precision of 1.00 on a 200-decision validation sample (95% Wilson CI: [0.982, 1.000]). Three principal findings emerge. (1) The degree distribution follows a power law (alpha = 1.57 +/- 0.008), placing the Ukrainian court network near the EU Court of Justice and below the US Supreme Court, with hub articles cited by millions of decisions. (2) Louvain community detection on the co-citation projection recovers legal domain boundaries (civil, criminal, administrative, commercial) with modularity Q = 0.44-0.55 and temporal stability (NMI = 0.83-0.86 across periods), constituting an automatically constructed legal ontology grounded in judicial practice. (3) Citation features predict top-1000 articles with AUC = 0.9984, substantially outperforming a naive frequency baseline (P@1000 = 0.655); temporal dynamics detect legislative regime changes as phase transitions and the 2022 invasion as a citation entropy spike (H: 11.02 -> 13.49) with emergent wartime legislation nodes. The citation-derived ontology is operationalized as the domain layer of a workflow memory system for LLM-assisted legal analysis, connecting to the ontology-controlled paradigm. The extraction pipeline, analysis code, and aggregated statistics are released as open data.
Authors: Thomas Hikaru Clark, Sihan Chen, Laura Nicolae
Abstract: Communicating using only a limited vocabulary is a common but challenging cognitive phenomenon, requiring an ideal communicator to plan carefully to optimize for intelligibility while circumventing a constrained lexicon. In this work, we investigate how humans respond to a broad array of questions under variable vocabulary limitations, consisting of only 250 highly frequent words at the most restrictive. We provide theoretically motivated comparisons to greedy and globally optimal sampling algorithms using Sequential Monte Carlo inference with large language models. Humans generally resemble greedy sampling more than globally optimal sampling, though more skilled humans are more likely to backtrack and revise -- a non-greedy behavior. An observed human pattern of leaning on semantically light words in high-constraint settings falls out of both greedy and globally optimal sampling. We discuss the results and their broader implications for resource-rational cognition, psycholinguistics, L2 communication, and language impairments.
Authors: George Boateng, Evans Atompoya, Philemon Badu, Samuel John, Samuel Ansah, Patrick Agyeman-Budu, Victor Wumbor-Apin Kumbol
Abstract: Sub-Saharan Africa faces persistently high student-teacher ratios and shortages of qualified teachers, limiting students' access to personalized learning support and formative assessment. To address this challenge, we present Adesua, a WhatsApp-based AI Teaching Assistant for science education that extends the Kwame for Science platform. Adesua leverages WhatsApp's widespread adoption in Africa to provide accessible, curriculum-aligned learning support for Junior High School (JHS) and Senior High School (SHS) students across West Africa. The system integrates curated textbooks and 33 years of national examination questions with generative AI to enable conversational question answering and automated assessment with feedback via a WhatsApp bot. Students can ask science questions, take timed or untimed multiple-choice tests by topic or exam year, and receive instant grading and detailed explanations of correct and incorrect responses. A 6-month feasibility deployment in 2025 had 56 active users in Ghana, including students and parents. Quantitative evaluation showed a high perceived usefulness, with a helpfulness score of 93.75\% for AI-generated answers, albeit with a small number of ratings (n=16). These preliminary results provide a basis for more extensive future evaluation of a WhatsApp-based AI assistant to assess its potential to offer scalable, low-cost personalized learning support and formative assessment in resource-constrained educational contexts.
Authors: George Boateng, Philemon Badu, Patrick Agyeman-Budu, Samuel Ansah, Evans Atompoya, Evan Igwilo, Lord Baah, Frederick Abu-Bonsrah, Victor Wumbor-Apin Kumbol
Abstract: Recent advances in generative AI have shown their potential to be leveraged for legal education. Yet, work on the development and deployment of such systems for legal education in the Global South is limited. In this work, we developed Eskwai for Students, a generative AI assistant to help law students with their legal education. Eskwai for Students is a retrieval augmented generation (RAG) system that provides answers to a wide range of legal questions for law students grounded in a curated database of over 12K case laws and 1.4K legislation in Ghana. We deployed Eskwai for Students in a longitudinal study of 30 months (2.5 years) used by 3.1K law students in Ghana who made 32K queries. We evaluated the helpfulness of our AI, and provided insight into the kinds of queries law students submit to this generative AI tool, which raises some ethical concerns. This work contributes to an understanding of how law students in the Global South are using generative AI for their studies and the ways it could be leveraged responsibly to advance legal education.
Authors: Sen Yang, Yinglei Ma
Abstract: Large language model personalization typically adapts outputs to user preferences and style but does not account for differences in user evaluation capacity across domains of expertise. This limitation can encourage Professional Domain Drift, where users rely on AI generated reasoning in domains they cannot reliably evaluate. We introduce Capability Conditioned Scaffolding, a typed framework that partitions expertise into strong, mixed, and weak domains and conditions intervention behavior on structured capability profiles. A pilot evaluation across multiple MMLU subsets and four LLM substrates shows consistent profile conditioned intervention behavior, including categorical inversion under profile swapping and selective activation in mixed domain risk zones. These findings suggest that capability aware scaffolding can support more reliable professional human AI collaboration beyond stylistic personalization.
Authors: Mahdi Naser-Moghadasi, Faezeh Ghaderi
Abstract: This paper presents a comprehensive analysis of neural activation patterns across six distinct large language model (LLM) architectures, examining their performance on twelve cognitive task categories. Through systematic measurement of final activation values, attention entropy, and sparsity patterns, we reveal fundamental differences in how encoder and decoder architectures process diverse cognitive tasks. Our analysis of 144 task-model combinations demonstrates that mathematical reasoning consistently produces the highest attention entropy across all architectures, while decoder models exhibit significantly higher sparsity patterns compared to encoder models. The findings provide critical insights into the computational characteristics of modern language models and their task-specific neural behaviors, with implications for model selection and optimization in big data applications.
Authors: William Timkey, Brian Dillon, Tal Linzen
Abstract: Surprisal theory posits that the processing difficulty of a word is determined by its predictability in context, offering a potential link between human sentence processing and next-word predictions from language models. While language model (LM) surprisals successfully predict reading times in naturalistic text, they systematically underpredict the magnitude of difficulty observed in controlled studies of syntactic ambiguity, particularly in garden path sentences. This mismatch might arise from differences in the computational constraints between humans and LMs. Here we test one such hypothesis, specifically, that LMs may be able to simultaneously consider a greater number of distinct sentence interpretations at once, compared to humans. Using Recurrent Neural Network Grammars (RNNGs) with word-synchronous beam search, we systematically vary the number of simultaneous parses used to compute word surprisal, and then use these surprisals to predict human reading times. Reducing the number of simultaneous active parses indeed increases the magnitude of predicted garden path effects, but not nearly enough to capture the full magnitude of the effects in humans. This suggests that differences in the number of simultaneous parses available to LMs and humans cannot reconcile LM-based surprisal with human sentence processing.
Authors: Anders Gj{\o}lbye, Lars Kai Hansen, Sanmi Koyejo
Abstract: Reasoning-trained language models often spend more tokens on harder problems, but longer chains of thought do not show whether a model is merely computing for more steps or following a different internal trajectory. We study this distinction through hidden-state trajectories during chain-of-thought generation across competitive programming, mathematics, and Boolean satisfiability. Raw trajectory geometry is strongly shaped by generation length: longer generations mechanically alter path statistics, so difficulty-dependent comparisons are misleading without adjustment. After residualizing trajectory statistics on length, difficulty remains systematically coupled to corrected trajectory geometry across all domains studied. The clearest reasoning-specific separation appears in the code domain, where harder problems show more direct corrected trajectories and less heterogeneous local curvature in reasoning-trained models than in matched instruction-tuned baselines. Corrected difficulty-geometry coupling is weaker, but still present, in mathematics and Boolean satisfiability. Prompt-stage linear probes do not mirror the code-domain separation, and behavioral annotations show that stronger corrected coupling co-occurs with strategy shifts and uncertainty monitoring. Together, these findings establish length correction as a prerequisite for generation-time trajectory analysis and show that reasoning training can be associated with distinct corrected trajectory geometry, with the strength of the effect depending on the domain.
Authors: A H M Rezaul Karim, Ozlem Uzuner
Abstract: Conversational nurse-patient transcripts contain actionable observations, but converting these transcripts into structured representations at scale remains challenging. Documentation burden is substantial, with prior studies showing clinicians spend large portions of their workday on documentation and related desk work rather than direct patient care. MEDIQA-SYNUR focuses on observation extraction from conversational nurse-patient transcripts, requiring systems to normalize these narratives into a predefined schema with value-type constraints. We propose a modular retrieval-augmented generation (RAG) pipeline that uses the training set as an exemplar corpus, combines schema-constrained prompting (full schema vs. pruned candidate schema), deterministic schema-based postprocessing, and a second-pass audit, with two LLM backbones: Llama-4-Scout-17B-16E-Instruct and GPT-5.2 with corresponding embedding models for RAG. Our best configuration uses GPT-5.2 with full schema, RAG, and a second-pass auditing, achieving 80.36% F1 score. Overall, our results show that RAG consistently improves performance, while the optimal degree of schema constraint depends on the model, and second-pass auditing yields modest additional gains by correcting residual schema-adherence errors.
Authors: Dmitry Stanishevskii, Nini Kamkia, Alexey Khoroshilov, Dmitry Zmitrovich, Denis Kokosinskii, Zhirayr Hayrapetyan, Andrei Kalmykov
Abstract: Large language models (LLMs) are increasingly being applied to financial analysis, reporting, investment decision support, risk management, compliance, and professional training. However, robust evaluation of their domain competence in finance remains incomplete. Widely used open benchmarks such as FinQA, ConvFinQA, and TAT-QA have played an important role in advancing financial question answering and numerical reasoning, but they focus primarily on question answering over financial reports and do not provide an explicit hierarchy of professional difficulty. Broader resources, including FinanceBench, PIXIU, FinBen, and FLaME, expand the coverage of financial tasks, yet the problem of evaluating the transition from foundational knowledge to expert-level financial reasoning remains open. In this work, we present FINESSE-Bench, a suite of eight specialized benchmarks comprising 3,993 questions for hierarchical evaluation of financial competencies in LLMs. FINESSE-Bench combines exam-oriented datasets inspired by professional certifications (CFA-like Levels 1-3, CMT-like Level 2, and CFTe-like Level 1), applied trading task collections, and a Russian-language olympiad benchmark. This design enables evaluation of domain breadth, performance degradation as difficulty increases, the ability to solve computational tasks, and model behavior in specialized financial domains. We also describe a unified evaluation protocol covering multiple-choice questions, numerical answers, and short open-ended responses, together with an automated scoring scheme for freeform answers based on the LLM-as-judge paradigm. FINESSE-Bench is intended both as a complement to existing open financial benchmarks and as a tool for more substantive evaluation of professionally relevant financial competencies in large language models.
Authors: Yufeng Du, Phillip Harris, Minyang Tian, Eliu A Huerta, Srikanth Ronanki, Subendhu Rongali, Aram Galstyan, Hao Peng
Abstract: We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.
Authors: Junchao Wu, Yefeng Liu, Chenyu Zhu, Hao Zhang, Zeyu Wu, Tianqi Shi, Yichao Du, Longyue Wang, Weihua Luo, Jinsong Su, Derek F. Wong
Abstract: The effective detection and governance of Large Language Model (LLM) generated content has become increasingly critical due to the growing risk of misuse. Despite the impressive performance of existing detectors, their reliability and potential in multilingual, real-world scenarios remain largely underexplored. In this study, we introduce DetectRL-X, a comprehensive multilingual benchmark designed to evaluate advanced detectors across 8 dimensions. The benchmark encompasses 8 languages commonly used in commercial contexts and collects human-written texts from 6 domains highly susceptible to LLM misuse. To better aligned with real-world applications, We create LLM-generated texts using 4 popular commercial LLMs, and include typical AI-assisted writing operations such as polishing, expanding, and condensing to capture authentic usage patterns. Furthermore, we develop a multilingual framework for paraphrasing and perturbation attacks to simulate diverse human modifications and writing noise, enabling stress testing of detectors across languages. Experimental results on DetectRL-X reveal the strengths and limitations of current state-of-the-art detectors when applied to diverse linguistic resources. We further analyze how domains, generators, attack strategies, text length, and refinement operations influence performance in different languages, underscoring DetectRL-X as an effective benchmark for strengthening multilingual and language-specific detectors.
Authors: Jinyuan Li, Langlin Huang, Chengsong Huang, Shaoyang Xu, Donghong Cai, Yuyi Yang, Wenxuan Zhang, Jiaxin Huang
Abstract: Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.
Authors: De Shuai Zhang
Abstract: Continuous diffusion and flow models are attractive for non-autoregressive text generation because they can update all positions in parallel. A major difficulty is the interface between continuous latent states and discrete tokens. This report studies a draft-conditioned latent refinement model built from a frozen BERT encoder, a parallel decoder, a denoising DraftPrior, a local FlowNet, and a learned diagonal MetricNet. Early Gaussian-start experiments showed that good latent-space metrics, such as scale matching or cosine similarity, do not guarantee good decoding. Generated latents can be close to real encoder latents but still produce high-entropy, biased, or repetitive token distributions. We therefore frame the task as controlled local refinement rather than full generation from noise. On ROCStories, using the first two sentences as prompt and the last three as target, full 768-dimensional BERT latents recover tokens much better than compressed 256-dimensional latents. With 768-dimensional latents, DraftPrior target-token probability is 0.938 for clean drafts, 0.613 for 3% token dropout, 0.483 for 5% dropout, and 0.272 for 10% dropout. Local flow refinement and fused decoder-aware readout give modest additional gains, while metric learning and OT-style alignment improve geometry but do not close the decoder gap. The main result is a diagnostic one: latent geometry alone is not enough. Continuous latent text generation should be evaluated by decoder recoverability, the quality of the start distribution, and whether refinement preserves decoder-readable structure.
Authors: Tianyu Huang, Yida Zhao, Chuyan Zhou, Kewei Tu
Abstract: Augmenting Transformers with linguistic structures effectively enhances the syntactic generalization performance of language models. Previous work in this direction focuses on syntactic tree structures of languages, in particular constituency tree structures. We propose Graph-Infused Layers Transformer Language Model (GiLT) which leverages dependency graphs for augmenting Transformer language models. Unlike most previous work, GiLT does not insert extra structural tokens in language modeling; instead, it injects structural information into language modeling by modulating attention weights in the Transformer with features extracted from the dependency graph that is incrementally constructed along with token prediction. In our experiments, GiLT with semantic dependency graphs achieves better syntactic generalization while maintaining competitive perplexity in comparison with Transformer language model baselines. In addition, GiLT can be finetuned from a pretrained language model to achieve improved downstream task performance. Our code is released at https://github.com/cookie-pie-oops/GiLT-LM.
Authors: Luxuan Chen, Han Tian, Xinran Chen, Rui Kong, Fang Wang, Jiamin Chen, Yuchen Li, Jiashu Zhao, Shuaiqiang Wang, Haoyi Xiong, Dawei Yin
Abstract: The dynamic range of activations is a first-order constraint for low-bit quantization, activation scaling, and stable LLM inference. Prior work characterized outlier features and massive activations on pre-2024 LLaMA-style models, and the downstream activation-quantization stack inherits that picture without revisiting it for the post-LLaMA open-model boom. We ask the deployment-oriented question: how large can activations get in modern open LLMs, and how does this magnitude vary across families, generations, and training stages? Under a unified pipeline (5,000-sample multi-domain corpus, family-specific tokenization, identical hooks across embeddings, hidden states, attention, MLP/MoE, SwiGLU gates, and final norm), we measure global and layerwise maxima on 27 checkpoints from 8 open families spanning dense, MoE, vision-language, intermediate-training, and instruction-tuned variants. We find that (i) global maxima span over nearly four orders of magnitude at comparable parameter counts, with Qwen3.5 and MoE checkpoints in the 10^2 to 10^3 range and Gemma3-27B-it reaching ~7 x 10^5; (ii) cross-family and cross-generation comparisons break simple monotonic scaling; and (iii) MoE checkpoints exhibit 14.0-23.4x lower peaks than matched-scale dense counterparts, while the residual stream carries the global maximum in 22/24 checkpoints. A lightweight INT-8 sanity check shows that measured maxima co-vary with low-bit reconstruction error via activation-scale selection. We conclude that maximum activation magnitude is a model property tied to family, architecture, and training stage - not a simple byproduct of size - and should be measured and reported alongside any open-weight release before low-bit deployment. The code is publicly available at https://github.com/clx1415926/Max_act_llm.
Authors: Nurbek Tastan, Alex Iacob, Lorenzo Sani, Meghdad Kurmanji, Nicholas D. Lane, Samuel Horvath, Karthik Nandakumar
Abstract: Multi-agent systems can solve complex tasks through collaboration between multiple Large Language Model agents. Existing collaboration frameworks typically operate in either a parallel or a sequential mode. In the parallel mode, agents respond independently to queries followed by aggregation of responses. In contrast, sequential systems allow agents to communicate via a directed topology and refine one another step by step. However, both modes are inadequate for achieving the desired objectives of minimizing communication and latency while simultaneously maximizing the accuracy of the final response. In this work, we introduce a hybrid paradigm called Nexa, a trainable response-conditioned policy that bridges the gap between the two modes. Nexa begins with a parallel execution stage, embeds the resulting responses into a shared semantic space, and then predicts a sparse directed acyclic communication graph. If the graph is empty, the system remains purely parallel; if it is non-empty, the system performs one sequential message propagation. The policy is a lightweight transformer model, and the method avoids the need for external LLM judges or reward models, as well as hand-crafted test-time topology search. We formalize this hybrid execution problem, show that the resulting graph is acyclic by construction, and that the framework strictly subsumes pure parallel execution, and present a training procedure based on policy-gradient optimization. Results demonstrate that the response-conditioned policy learned by Nexa under one setting can be reused when the number of agents, the task, or the underlying agent changes, thus emphasizing the generalizability of the learned communication policy.
Authors: Fengfei Yu, Ruijia Niu, Dongxia Wu, Yian Ma, Rose Yu
Abstract: As large language models (LLMs) are deployed in consequential settings such as medical question answering and legal reasoning, the ability to estimate when their outputs are likely to be correct is essential for safe and reliable use, requiring well-calibrated uncertainty. Standard reinforcement learning with verifiable rewards (RLVR) trains models with a binary correctness reward that is indifferent to confidence, providing no penalty for confident but wrong predictions and thereby degrading calibration. Recent work addresses this by training models to produce verbalized confidence scores alongside answers and rewarding agreement with correctness. However, verbalized confidence is calibrated at the token level and thus exhibits inconsistency across textual variations with same semantic meaning. We propose \textbf{Calibration with Semantic Reward (CSR)}, a framework that calibrates language models directly in semantic space without a verbalized confidence interface. CSR combines the correctness reward with a novel semantic calibration reward that encourages exploitation among correct rollouts by promoting semantic agreement, and exploration among incorrect ones by discouraging spurious consistency. Experiments across three model families on HotpotQA (in-distribution) and TriviaQA, MSMARCO, and NQ-Open (out-of-distribution) show that CSR consistently achieves lower ECE and higher AUROC than verbalized-confidence baselines across nearly all settings, reducing ECE by up to $40\%$ and improving AUROC by up to $31\%$ over verbalized-confidence baselines, with calibration behavior generalizing robustly across all four evaluation settings.
Authors: Weixin Liu, Congning Ni, Shelagh A. Mulvaney, Susannah L. Rose, Murat Kantarcioglu, Bradley A. Malin, Zhijun Yin
Abstract: Large language models (LLMs) are increasingly used in the mental health domain, yet it remains unclear how well they capture related biomedical knowledge and how reliably they apply it to clinically salient structured judgments. Here, we present a knowledge-graph (KG)-grounded benchmark for assessing LLMs on mental-health entity recognition, relation judgment, and two-hop reasoning. The benchmark is derived from PrimeKG and comprises nine task families with KG-supported answers and controlled negative options. Experiments across 15 closed- and open-source LLMs reveal a persistent recognition-to-judgment gap: leading models achieve near-ceiling performance on entity typing and on the small relation-typing subset, yet they still struggle with relation prediction and two-hop reasoning. Additionally, short KG-derived snippets benefit some models but degrade performance for others. Moreover, output-format reliability can substantially influence measured performance under constrained multiple-choice settings, highlighting the critical role of response validity in benchmark-based evaluation. MHGraphBench should therefore be interpreted as evaluating agreement with a curated mental-health slice of PrimeKG under a constrained multiple-choice interface, rather than as a direct assessment of real-world clinical safety.
Authors: Vinayshekhar Bannihatti Kumar, Disha Makhija, Manoj Ghuhan Arivazhagan, Rashmi Gangadharaiah
Abstract: Large language models (LLMs) achieve high pass rates on code generation benchmarks, yet whether they can transfer this ability to languages absent from pretraining remains poorly understood. We introduce PyLang, a minimal imperative language absent from all pretraining corpora, and evaluate frontier models zero-shot and fine-tuned Qwen3 (4B, 8B, 32B) on 352 problems. We find that fine-tuning quickly teaches syntax but fails to transfer semantic competence: Python outperforms PyLang by up to 19% across all configurations, and no intervention (multi-task learning, preference tuning, code infilling, or latent-space objectives) closes the gap. An LLM judge reveals that frontier models select an identical algorithm to Python 80% of the time, yet cannot translate it into a working PyLang implementation., and CKA analysis confirms that fine-tuned models converge to nearly identical internal representations across languages (CKA > 0.97) while diverging at the output stage. We term this the implementation fidelity gap: models possess language-agnostic algorithmic understanding but cannot express it in an unfamiliar language. Our findings highlight the need for training methods that decouple reasoning from language-specific realization.
Authors: Shengyin Sun, Yiming Li, Renxi Liu, Xinqi Li, Hui-Ling Zhen, Weizhe Lin, Chen Chen, Xianzhi Yu, Mingxuan Yuan, Chen Ma
Abstract: Diffusion large language models (dLLMs) generate text by iteratively denoising masked token sequences. Although dLLMs can predict all masked positions in parallel within each step, the large number of denoising iterations still makes inference expensive. This cost can be reduced spatially by unmasking multiple tokens per step, or temporally by collapsing multiple denoising steps into one verification call. We propose Parallel Speculative Decoding (PSD), a training-free framework that jointly improves inference along both axes. Using the confidence scores from a single forward pass, PSD selects positions to unmask via a configurable, adaptive unmasking policy and constructs multi-depth speculative drafts without extra model calls. A final batched verification pass then applies hierarchical acceptance, keeping the deepest draft that remains consistent with the updated predictions. Experiments on three dLLMs across reasoning and code generation tasks show that PSD achieves favorable trade-offs between inference efficiency and generation quality, reaching up to $5.5\times$ tokens per forward pass with accuracy comparable to greedy decoding.
Authors: Sho Takase, Ukyo Honda
Abstract: Through an analysis of sequences generated by open-weight large language models (LLMs), we demonstrate that LLMs are heavily biased toward English. While continual pre-training is commonly used to adapt LLMs to a target language, we show that it does not offer a cost advantage over training from scratch, even for improving cultural understanding in the target language. These findings suggest that dedicated per-language investment may become increasingly important for future LLM development, rather than relying primarily on the expansion of English-centric resources.
Authors: Junwen Mo, Yuanzhi Lu, Yifang Xue, Ke Xu, Hideki Nakayama
Abstract: Linguistic ambiguity is critical to the robustness of Large Language Models (LLMs), yet existing research focuses mostly on English, with limited attention devoted to Chinese. Existing Chinese ambiguity datasets (e.g., CHAmbi) suffer from poor scalability. Guided by Potential Ambiguity (PA) Theory, we design a semi-automatic pipeline to construct CHA-Gen. It is the first PA Theory-grounded Chinese ambiguity dataset, which comprises 5,712 sentences (2,414 ambiguous, 3,298 unambiguous) across 18 potential ambiguous structures. Evaluating LLMs (e.g. Gemma 3, Qwen 2.5/3 series) via direct querying and machine translation, we find that LLMs struggle with ambiguity detection (improved by CoT prompting). Analysis of Qwen3-32B's CoT rationales reveals three common failure modes: ambiguity blindness, misattribution, and premature resolution. Uncertainty quantification with semantic entropy metric shows higher uncertainty for ambiguous sentences. Moreover, instruction tuning induces overconfidence, whereas Base models better capture semantic diversity. We further observe that models exhibit a bias toward dominant interpretations. Our work provides a scalable approach for Chinese ambiguity corpus and insights into LLMs' ambiguity handling, laying a foundation for enhancing Chinese ambiguity research in LLMs.
Authors: Yichen Zhu, Xiaoming Shi, Peng Zhao, Weiyu Chen, Debing Zhang, James Kwok
Abstract: Block discrete diffusion language models factorize a sequence autoregressively over fixed-size positional blocks, decoupling within-block parallel denoising from across-block conditioning. We argue that this rigid partition wastes structure already present in the sequence: blocks defined by position rather than by content separate semantically coherent tokens and group unrelated ones together. We introduce the \textbf{D}ynamic \textbf{C}hunking \textbf{D}iffusion \textbf{M}odel (DCDM), which replaces positional blocks with content-defined semantic chunks. At its core is Chunking Attention, a differentiable layer that routes tokens into $K$ clusters parameterized by learnable subspaces and shaped end-to-end by the diffusion objective. The resulting cluster assignments induce a chunk-causal attention mask under which a discrete diffusion denoiser factorizes the sequence likelihood autoregressively over semantic chunks, strictly generalizing block discrete diffusion. On downstream benchmarks at parameter scales up to 1.5B, DCDM consistently improves over both unstructured and positional-block diffusion baselines, with the advantage stable across scales and visible early in training.
Authors: Xiaoyan Su, Peijie Dong, Zhenheng Tang, Song Tang, Yuyao Zhai, Kaitao Lin, Liang Chen, Gai Yuhang, Yuyu Luo, Qiang Wang, Xiaowen Chu
Abstract: Despite the rapid advancements in Vision-Language Models (VLMs), a critical gap remains in their ability to handle structured, controllable diagrammatic tasks essential for professional workflows. Existing methods predominantly rely on pixel-based synthesis, which operates in probabilistic pixel spaces and is inherently limited in editability and fidelity. Instead, we propose a new Diagram-as-Code paradigm with symbolic logic that leverages mxGraph Extensible Markup Language (XML) for precise diagram generation and editing. We present VCG-Bench, a unified benchmark for visual-centric \texttt{mxGraph} tasks. VCG-Bench comprises: (1) a taxonomized dataset of 1,449 diverse diagrams spanning 6 domains and 15 sub-domains, (2) a paradigm definition that integrates Generation (Vision-to-Code) and Editability (Code-to-Code), (3) a Tailored Evaluation Protocol employing multi-dimensional metrics such as \texttt{mxGraph} Execution Success Rate, Style Consistency Score (SCS), etc. Experimental results highlight the challenges faced by current State-of-the-Art (SOTA) VLMs in structured fidelity and instruction compliance, reflecting their vision and reasoning capabilities.
Authors: Liqi Zhou, Jiafu Li
Abstract: Online patient inquiries are often informal, incomplete, and written before professional assessment, yet they must still be routed to an appropriate level of clinical follow-up. We study this as a four-class actionable triage task -- self-care, schedule-visit, urgent-clinician-review, or emergency-referral, and ask whether prompted large language models (LLMs) can support such routing under low-resource labeling conditions. Using the public HealthCareMagic-100K corpus, we construct a 300-example human calibrated gold evaluation set, a 700-example auto-labeled silver training set, and a 40-example few-shot pool. We compare Term Frequency-Inverse Document Frequency (TF-IDF) and Bidirectional Encoder Representations from Transformers for Biomedical Text Mining (BioBERT) baselines train on silver labels against six prompted LLMs under 0-shot, 4-shot, and 12-shot conditions respectively. Accordingly, we evaluate with macro-$F_1$ alongside safety-aware metrics, including emergency-recall, under-triage rate, and severe under-triage rate. The strongest LLM (Claude Haiku 4.5, 12-shot) reaches macro-$F_1$ 0.475, exceeding the best supervised baseline (BioBERT, 0.378) on point estimate, with overlapping confidence intervals. Few-shot prompting and two-model agreement help in label-dependent ways: self-care agreement is reliable, urgent-clinician-review is not. We conclude that LLMs can support triage prioritization and selective human review, but not autonomous deployment.
Authors: Jiahui Guang, Yingjie Zhu, Cuiyun Gao, Haiyan Wang, Jing Li, Di Shao, Zhaoquan Gu
Abstract: Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.
Authors: Jiawei Yu, Yixiang Fang, Xilin Liu, Yuchi Ma
Abstract: Memory data are ubiquitous in Large Language Model (LLM)-based agents (e.g., OpenClaw and Manus). A few recent works have attempted to exploit agents'memory for improving their performance on the question-answering (QA) task, but they lack a principled mechanism for effectively modeling how memory data evolves over time and retrieving memory data effectively, leading to poor performance in memory utilization. To fill this gap, we present H-Mem, a novel memory mechanism via a hybrid structure that can not only effectively model the evolution of agent memory over a long period of time, but also provide an efficient memory retrieval approach. Particularly, H-Mem builds a temporal and semantic tree structure that allows the short-term memory data to evolve progressively into long-term memory data, where the latter provides summarized information about the former, while simultaneously constructing a knowledge graph to capture the relationships between entities in memory. Moreover, it offers an effective memory retrieval approach by exploiting the hybrid structure of the tree and graph structures. Extensive experiments on three agent memory benchmarks show that H-Mem achieves state-of-the-art performance on the QA task.
Authors: Huacan Chai, Yukai Wang, Yingxuan Yang, Dan Peng, Yuanyi Song, Zhihui Fu, Weiwen Liu, Jianghao Lin, Jun Wang, Weinan Zhang
Abstract: Existing benchmarks for multimodal memory reasoning largely evaluate systems within pre-assembled contexts, but under-evaluate whether agents can use evidence distributed across independently originated sources. We argue that source-distributed memory composition is an important and under-examined bottleneck in multimodal agent memory, especially when relevant evidence is fragmented across heterogeneous artifacts such as conversations, profiles, screenshots, tables, images, and documents. To address this gap, we introduce Source-distributed Multimodal Memory Benchmark(SMMBench), which measures whether agents can retrieve, align, and compose multimodal evidence scattered across multiple sources rather than reason within a single curated context. SMMBench evaluates four core capabilities: (1) cross-source multimodal reasoning; (2) conflict resolution; (3) preference reasoning; (4) memory-grounded action prediction. The benchmark contains 1877 samples grounded in 264 sources. Experiments on representative memory-style and retrieval-based baselines show that current systems still struggle on these capabilities, positioning source-distributed multimodal memory as an important and still under-evaluated challenge for multimodal agents. Our data are available at https://huggingface.co/datasets/HuacanChai/SMMBench.
Authors: Jiachen Zhu, Zhuoying Ou, Congmin Zheng, Yuxiang Chen, Zeyu Zheng, Rong Shan, Lingyu Yang, Lionel Z. Wang, Weiwen Liu, Yong Yu, Weinan Zhang, Jianghao Lin
Abstract: Large Language Models (LLMs) are highly sensitive to their input contexts, motivating the development of automated context engineering. However, existing methods predominantly treat this as a global search problem, seeking a single context strategy that maximizes average performance across a dataset. This restrictive assumption overlooks the fact that different inputs often require distinct guidance, leaving substantial instance-level performance gains untapped. In this paper, we propose a paradigm shift by formulating context engineering as a recommendation problem. We introduce \textbf{Neural Collaborative Context Engineering (NCCE)}, a framework that transitions optimization from a static global search to dynamic, instance-wise routing. NCCE first bootstraps a diverse catalog of anchor contexts and then employs a novel \textbf{Context-CF Co-Evolution} mechanism. This stage establishes a synergistic feedback loop: a lightweight Neural Collaborative Filtering (NCF) model learns instance-context preferences to guide the generation of specialized context variants, while the newly evaluated contexts continuously refine the NCF model's understanding of latent preferences. At inference time, the trained NCF model acts as a context router, dynamically assigning the most suitable context strategy to each unseen instance. Theoretical Proofs and comprehensive experiments demonstrate that by matching individual inputs with their optimal contexts, NCCE significantly improves task accuracy, highlighting the critical importance of personalization in LLM context engineering.
Authors: Wentao Qiu, Haotian Hu, Fanyi Wang, Jinwei Kong, Yu Zhang
Abstract: Large language model (LLM) agents require long-term memory to leverage information from past interactions. However, existing memory systems often face a fidelity--efficiency trade-off: raw dialogue histories are expensive, while flat facts or summaries may discard the structure needed for precise recall. We propose \textbf{DimMem}, a lightweight dimensional memory framework that represents each memory as an atomic, typed, and self-contained unit with explicit fields such as time, location, reason, purpose, and keywords. This representation exposes the structure needed for dimension-aware retrieval, memory update, and selective assistant-context recall without storing full histories in the model context. Across LoCoMo-10 and LongMemEval-S, DimMem achieves \textbf{81.43\%} and \textbf{78.20\%} overall accuracy, respectively, outperforming existing lightweight memory systems while reducing LoCoMo per-query token cost by \textbf{24\%}. We further show that dimensional memory extraction is learnable by compact models: after fine-tuning on the DimMem schema, a Qwen3-4B extractor surpasses LightMem with GPT-4.1-mini on both benchmarks and reaches performance comparable to, or better than, much larger extractors in key settings. These results suggest that explicit dimensional structuring is an effective and efficient foundation for long-term memory in LLM agents. Code is available at https://github.com/ChowRunFa/DimMem.
Authors: Kamil Guttmann, Zofia Fra\'s, Artur Nowakowski, Krzysztof Jassem
Abstract: Current state-of-the-art Quality Estimation (QE) in machine translation relies on massive, proprietary LLMs, raising data privacy concerns. We demonstrate that smaller, open-source LLMs (<30B parameters) are a viable, cost-effective and privacy-preserving alternative. Using a single-pass prompting strategy, our models simultaneously generate quality scores, MQM error annotations, suggested error corrections, and full post-editions. Our analysis shows these models achieve highly competitive system-level correlations with human judgments that outperform traditional neural metrics, fine-tuned models, and human inter-annotator agreement, effectively approximating the capabilities of much larger proprietary LLMs.
Authors: Micha{\l} Ciesi\'o{\l}ka, Dawid Wi\'sniewski, Adrian Charkiewicz, Kamil Guttmann
Abstract: We present ForMaT (Format-Preserving Multilingual Translation), a parallel corpus of 3,956 PDFs across 15 language pairs that preserves original layout metadata proposed for multimodal machine translation. To ensure structural diversity in the dataset, we employ K-Medoids sampling over 45 geometric features, capturing complex elements like nested tables and formulas to focus only on visually diverse PDF documents. Our evaluation reveals that current MT systems struggle with spatial grounding and geometric synchronization, often losing the link between text and its visual context. ForMaT provides a benchmark for developing layout-aware translation models that integrate visual and textual context for high-fidelity document reconstruction.
Authors: Daria Blinova, Gayathri Emuru, Rakesh Emuru, Kushagradheer Shridheer Srivastava, Mina Rulis, Sunita Chandrasekaran, Benjamin E. Bagozzi
Abstract: This paper introduces a dataset of interlinked multimodal political communications from the Russian government, addressing persistent deficiencies in the availability of social text- and image-based data for authoritarian politics contexts. The dataset comprises two large corpora of official speeches delivered by senior actors within the Kremlin and the Russian Ministry of Foreign Affairs over multiple decades. For each speech, we provide Russian- and English-language texts, associated images and captions where available, and harmonized metadata including (e.g.) dates, speakers, (geo)locations, and official government content tags. Unique identifiers link images to speeches and align Russian and English versions of the same communication texts. We further augment these linked datasets with validated topical annotations for both speech texts and speech images, which are generated via transformer-based multimodal topic modeling and refined by a Russian politics expert. The resulting data resources support multimodal, multilingual, temporal, and/or spatial analyses of (authoritarian) political communication and offer a valuable testbed for social science research and large language model (LLM) applications in political domains.
Authors: Shuaiyi Li, Zhisong Zhang, Yan Wang, Lei Zhu, Dongyang Ma, Chenlong Deng, Yang Deng, Wai Lam
Abstract: Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios such as Retrieval-Augmented Generation (RAG). However, its broader application is hindered by two key challenges: the difficulty of segmenting input text into meaningful, self-contained blocks, and the inefficiency of existing block fine-tuning methods that risk degrading performance. To address these, we first construct SemanticSeg, a large and diverse semantic segmentation dataset containing over 30k instances across 16 categories-including books, code, web text, and conversations with text lengths ranging from 2k to 32k. Using this dataset, we train a lightweight segmenter to automatically partition text into human-instinct-aligned blocks with controllable granularity. Second, we propose block distillation, a training framework that is more efficient than block fine-tuning, which uses a frozen full-attention teacher model to guide the block-attention student. This framework integrates three novel components: block sink tokens to mitigate information loss at block boundaries, block dropout to leverage training signals from all blocks, and token-level loss weighting to focus learning on block-attention-sensitive tokens. Experiments across multiple models and benchmarks demonstrate that our segmenter outperforms heuristic and statistical baselines, and block distillation achieves near-full-attention performance under block attention, establishing a practical and scalable pathway for deploying block attention.
Authors: Ernesto Garcia-Estrada, Carlos Escolano, Jos\'e A. R. Fonallosa
Abstract: Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at $\geq$7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.
Authors: Anita Srbinovska, Jansen Orfan, Adrian Martin, Ernest Fokou\'e
Abstract: Law enforcement reports contain structured fields and written narratives. However, many incident facts that are needed for review, police training, and investigations are in natural language and require manual reading. We propose a framework using symbolic methods for converting narratives into evidence-linked facts. Our objective is to measure the value of narratives to recover incident details only from the unstructured text and build temporal graphs with time cues and domain axioms. We achieve this by redacting personal identifiers, semantic parsing, predicate mapping to ontology, and reasoning. We evaluate the symbolic approach on 450 property crime reports and a short human review. Of the extracted events from the system, 54.1% had a confidence score of at least 0.80 and 93.7% were mapped through the PropBank--VerbNet--WordNet semantic path. 100% agreement was reached on incident initiation, stolen items, and temporal cues and lower agreement for forced entry interpretation.
Authors: Isar Nejadgholi, Masoud Kianpour, Krishnapriya Vishnubhotla, Maryam Molamohamadi
Abstract: Tremendous efforts have been put into evaluating the inclusivity and effectiveness of AI systems across cultures. However, the cultural capabilities considered in much of the literature remain vaguely defined, are referred to using interchangeable terminology, and are typically limited to recalling accurate information about various demographics, regions, and nationalities. To address this construct ambiguity, we draw from Intercultural Communication scholarship and propose a three-level taxonomy of AI-relevant cultural capabilities: Cultural Awareness answers "Does the model know?", Cultural Sensitivity answers "How does it frame its knowledge?", and Cultural Competence answers "Can it adapt as the interaction evolves?". Beyond conceptual clarification, we position this taxonomy as a practical tool for improving the validity and interpretability of AI evaluation in real-world, multicultural settings. Without such construct clarity, evaluation results risk overstating model capabilities and may lead to inappropriate deployment decisions in culturally sensitive contexts.
Authors: Jie Gao, Yongan Yu, Junzhu Su, Yiran Lin, Adam K. Dube, Jackie Chi Kit Cheung
Abstract: Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.
Authors: Nils Feldhus, Tanja Baeumel, Elena Golimblevskaia, Qianli Wang, Van Bach Nguyen, Aaron Louis Eidt, Christopher Ebert, Wojciech Samek, Jing Yang, Vera Schmitt, Sebastian M\"oller, Simon Ostermann
Abstract: LLM-as-a-judge has become the dominant paradigm for grading model outputs at scale, yet the same model assigns systematically different scores when its output format changes (e.g., a 1-5 rating vs. a True/False label). Existing diagnoses of these format-induced inconsistencies stop at the input-output level. Using Position-aware Edge Attribution Patching (PEAP), we causally investigate the internal mechanism in Gemma-3, Qwen2.5, and Llama-3. We find that judgments across structured understanding and open-ended preference tasks share a sparse, generalized Latent Evaluator sub-graph in the mid-to-late multi-layer perceptrons (MLPs); zero-ablating it collapses judgment while preserving world knowledge in architecturally modular models. By structurally decoupling abstract judging from output formatting, we provide a mechanistic account of format-induced inconsistency on the open-weight models we study: a continuous judgment signal computed in the shared trunk is mapped through fragile, format-specific terminal branches, enabling format-independent preference to be isolated downstream of the requested output format. Our findings imply that benchmark-level reliability comparisons across formats are partially measuring formatter geometry rather than evaluation quality.
Authors: Yu Pan, Yang Hou, Xiongfei Wu, Liang Zhang, Yves Le Traon, Lei Ma, Jianjun Zhao
Abstract: Compositional speech-to-speech translation (S2ST) systems built upon speech large language models (SpeechLLMs) have recently shown promising performance. However, existing S2ST systems often either neglect source-language information or encode it through a language-as-label paradigm, representing each source language as an independent flat embedding. Such a design overlooks systematic linguistic structure shared across languages, which may limit data-efficient multilingual adaptation when supervised S2ST data are scarce. To address this issue, we propose S2ST-Omni 2, a many-to-one compositional S2ST framework that systematically reformulates multilingual language conditioning from flat language labels to structured typological priors. Specifically, S2ST-Omni 2 revisits language conditioning at three levels: typology-informed hierarchical language encoding for structured source-language representation, dynamically-gated language-aware Dual-CTC for content-adaptive acoustic modulation, and typology-aware LLM prompting for decoder-side linguistic guidance. Experiments on CVSS-C show that S2ST-Omni 2 achieves superior average performance among representative S2ST approaches across BLEU, COMET, ASR-BLEU, and BLASER 2.0 under the adopted evaluation protocol. Ablation studies indicate that the proposed representation-level, acoustic-level, and decoding-level strategies provide complementary benefits. Moreover, controlled data-budget analyses and a Japanese-to-English evaluation using only approximately 3 hours of supervised training data suggest that explicit typological priors provide useful inductive biases for data-efficient multilingual S2ST.
Authors: Zijie Dai, Shiyuan Deng, Sheng Guan, Yizhou Tian, Xin Yao, Xiao Yan, James Cheng
Abstract: Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.
Authors: Si-Belkacem Yamine Ketir, Lenard Paulo Tamayo, Shohei Hisada, Shaowen Peng, Shoko Wakamiya, Eiji Aramaki
Abstract: Accurate assessment of cognitive decline from spontaneous speech remains challenging due to limited dataset size and class imbalance. In this work, we propose a large language model (LLM)-driven data augmentation framework to improve the prediction of cognitive scores from speech. Experiments are conducted on a Japanese corpus in which each participant provides both a spontaneous oral narrative and a written response to the same clinical prompt. The written responses serve as semantic anchors to generate multiple oral-like monologues in different styles using GPT-5. We then predict Hasegawa Dementia Scale scores, a widely used cognitive screening tool in Japan, using a Partial Least Squares regression model trained on Sentence-BERT speech embeddings. We investigate two augmentation strategies: random class-balanced selection, which yields moderate but unstable improvements, and similarity-guided class-balanced selection. The latter prioritizes semantically close synthetic samples, leading to more consistent improvements and substantially reducing prediction error for minority low-score participants while maintaining performance for the majority group. Overall, our findings demonstrate the potential of semantically guided LLM-driven augmentation as a principled approach for addressing class imbalance and improving data efficiency in clinical speech analysis.
Authors: Chenwang Wu, Yiuming Cheung, Bo Han, Shuhai Zhang, Defu Lian
Abstract: Machine-generated texts (MGTs) pose risks such as disinformation and phishing, underscoring the need for reliable detection. Metric-based methods, which extract statistically distinguishable features of MGTs, are often more practical than complex model-based methods that are prone to overfitting. Given their diverse designs, we first place representative metric-based methods within a unified framework, enabling a clear assessment of their advantages and limitations. Our analysis identifies a core challenge across these methods: the token-level detection score is easily biased by the inherent randomness of the MGTs generation process. Then, we theoretically derive the multi-hop transitions of the token-level detection score and explore their local and global relations. Based on these findings, we propose a multi-level contextual token relation modeling framework for MGT detection. Specifically, for local relations, we model them through a lightweight Markov-informed calibration module that refines token-level evidence before aggregation. For global relations, we introduce a rule-support reasoning module that uses explicit logical rules derived from contextual score statistics. Finally, we combine the local calibrated score and the global rule-support reasoning signal in a joint multi-level inference framework. Extensive experiments show broad and substantial improvements across various real-world scenarios, including cross-LLM and cross-domain settings, with low computational overhead.
Authors: Rui Chu, Bingyin Zhao, Thanh Quoc Hung Le, Duy Cao Hoang, Huawei Lin, Ping Li, Weijie Zhao, Khoa D Doan, Yingjie Lao
Abstract: Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.
Authors: Xin Zhang, Yang Cao, Baoxing Wu, Kai Song, Siying Li
Abstract: Large Language Models (LLMs) have demonstrated strong capabilities across diverse NLP applications, such as translation, text generation, and question answering. Nevertheless, they remain limited in complex settings that demand deep reasoning and logical inference. Since these models are trained on large-scale text corpora, their generation process may still introduce irrelevant, noisy, or factually inconsistent content. To mitigate this problem, we introduce SGR, a stepwise framework that enhances LLM reasoning through external subgraph generation. SGR builds query-specific subgraphs from external knowledge bases and uses their semantic structure to support multi-step inference. By grounding intermediate reasoning steps in structured external knowledge, the framework helps the model concentrate on relevant entities, relations, and supporting evidence. In particular, SGR first constructs a subgraph tailored to the input question. It then guides the model to reason progressively over the generated structure and combines multiple reasoning trajectories to obtain the final prediction. Experimental results across several benchmark datasets show that SGR achieves consistent improvements over competitive baselines, highlighting its value for improving both reasoning accuracy and factual reliability.
Authors: Michael P. Brenner, Lizzie Dorfman, John C. Platt
Abstract: We present a case study for how AI coding systems can be used to generate novel scientific hypotheses. We combine a generic coding agent (Google's AntiGravity) with an LLM-driven tree search algorithm (Empirical Research Assistance / ERA) to autonomously generate high-efficiency three-dimensional photovoltaic (3DPV) structures that overcome losses limiting flat solar panels at mid-latitudes. These structures operate by presenting favorable angles to the sun throughout the day, and for illustrative purposes we focus on optimizing performance for a single solar day. Our workflow begins by using AntiGravity to reproduce calculations \cite{bernardi2012solar} showing that 3DPV can have energy densities much higher than stationary flat PV panels. We use these initial designs as the starting point for large scale tree search, where we seek improved solutions and score them for their diurnal yield. The initial tree search leads to nominally more efficient solutions, yet they are caused by algorithmic reward hacking, arising from non-physical design features such as structurally levitating disconnected tiers and exploitations of the discretizations in the optics solver. To counteract this, we develop a workflow where the coding agent iteratively patches the physics engine with constraints to eliminate reward hacking. With reward-hacking eliminated, ERA discovers a series of designs with various constraints and improved performance, including optimal designs with different fixed collector areas, optimizing zenith tracking and avoiding self shadowing. Combining coding agents with tree search (ERA) provides a powerful platform for scientific discovery, for problems whose solutions can be empirically evaluated with a score function.
Authors: Axel Abels, Elias Fernandez Domingos, Apurva Shah, Tom Lenaerts
Abstract: Large language models (LLMs) are increasingly used to simulate human opinions and survey responses, but their ability to reproduce population responses across cultures remains limited. Existing persona-based prompting methods typically rely on sociodemographic or personality traits, which are only indirect proxies for the values that shape human responses. We propose a value-based persona construction method that derives textual descriptors from survey responses capturing core cultural dimensions. By sampling value profiles from target populations and aggregating LLM responses across personas, we obtain population-level predictions grounded in observed value distributions. We further introduce a calibration procedure that improves response diversity while preserving estimated opinions. We show that our approach reduces prediction error across countries, with the largest improvements observed in underrepresented populations. This substantially narrows the performance gap between countries aligned with dominant LLM priors and those that are less represented in training data, while also yielding response distributions that closely match human diversity.
Authors: Zhen Zhang, Liangcai Su, Zhuo Chen, Xiang Lin, Haotian Xu, Simon Shaolei Du, Kaiyu Yang, Bo An, Lidong Bing, Xinyu Wang
Abstract: Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.
Authors: Nathan Roll, Jill Kries, Laura Gwilliams, Cory Shain
Abstract: Aphasias, selective language impairments which can arise from brain damage, reveal the functional organization of human language by providing causal links between affected brain regions and specific symptom profiles. Drawing on this literature, we introduce an aphasia-inspired technique to characterize the emergent functional organization of language models (LMs). We ``lesion'' (zero-out) model parameters and measure the effects of this intervention against clinical aphasia symptoms, as diagnosed by the Text Aphasia Battery (TAB). When applied to 112,426 outputs from five 1B-scale LMs, the full range of evaluated symptoms surface, but in distributions largely distinct from those of humans. Our method uncovers broad symptom-profile differences between attention components (query, key, value, output) and feed-forward components (up, gate, down), with weaker evidence for differences among components within the same mechanism. We also find an effect of depth, where lesions in early layers disproportionately cause syntactic and semantic symptoms while late-middle layers yield higher rates of phonological and fluency deficits. Although some LM lesions induce quantitatively more similar profiles to some human aphasia types than others, qualitative differences in symptom patterns between LMs and humans suggest that aphasia syndromes are heavily influenced by the details of learning and processing rather than being a domain-invariant consequence of disrupted language processing.
Authors: Pavan Manjunath, Thomas pruefer
Abstract: The accelerating convergence of smart metering, generative artificial intelligence, and quantum-inspired combinatorial optimisation is reshaping how energy utilities manage physical infrastructure, customer engagement, and environmental accountability
Authors: Pavan Manjunath, Thomas Pruefer
Abstract: Distribution utilities are now expected to deliver bills that customers can actually read attach a defensible carbon number to every kWh sold and schedule load against grid stress and emissions constraints We propose an end-to-end framework that unifies four production-grade capabilities under one architectural roof a generative-AI agent that drafts each customers natural-language billing statement from structured numeric inputs under a constrained decoding policy a transformer-based forecaster that supplies the day-ahead consumption estimate with calibrated quantile bands
Authors: Jonas Belouadi, Tamy Boubekeur, Adrien Kaiser
Abstract: Material node graphs are programs that generate the 2D channels of procedural materials, including geometry such as roughness and displacement maps, and reflectance such as albedo and conductivity maps. They are essential in computer graphics for representing the appearance of virtual 3D objects parametrically and at arbitrary resolution. In particular, their directed acyclic graph structure and intermediate states enable a modular, interpretable workflow for interactive appearance modeling. However, creating such graphs remains challenging and typically requires professional training. While recent neural program synthesis approaches attempt to simplify this process, they solely represent graphs as textual programs, failing to capture the inherently visual-spatial nature of node graphs that makes them accessible to humans. To address this gap, we present MultiMat, a multimodal program synthesis framework that leverages large multimodal models to process both visual and textual graph representations for improved generation of procedural material graphs. We train our models on a new dataset of production-quality procedural materials and combine them with a constrained tree search inference algorithm that ensures static correctness while efficiently navigating the program space. Our experimental results show that our multimodal program synthesis method is more efficient in both unconditional and conditional graph synthesis with higher visual quality and fidelity than text-only baselines, establishing new state-of-the-art performance.
Authors: Ming Yang, Zhiwei Zhang, Jiahang Li, Haoseng Liu, Yuzheng Cai, Weiguo Zheng
Abstract: Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.
Authors: Yoichi Ishibashi, Taro Yano, Masafumi Oyamada
Abstract: AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.
Authors: Huihao Jing, Wenbin Hu, Haochen Shi, Hanyu Yang, Sirui Zhang, Shaojin Chen, Haoran Li, Yangqiu Song
Abstract: Large language models (LLMs) can often generate functionally correct code, but their ability to produce efficient implementations for performance-critical systems tasks remains limited. Existing code benchmarks mainly emphasize correctness or algorithmic problem solving, while realistic systems-level optimization is still underexplored. To address this gap, we introduce PerfCodeBench, an executable benchmark for evaluating LLMs on high-performance code optimization. The tasks require system-level implementation choices, hardware-aware optimization, and careful handling of performance bottlenecks. Each task includes executable correctness checks, a baseline implementation, and a reference optimized solution. This allows us to evaluate both correctness and runtime-oriented efficiency. Our evaluation on a broad set of state-of-the-art LLMs shows a clear gap between model-generated code and expert-optimized implementations. The gap is especially large on tasks involving parallelism and GPU operations. Current models also show weaknesses in cross-language robustness and in consistently reaching expert-level efficiency. These results suggest that performance-aware evaluation are still needed. LLMs should move beyond generating merely correct code toward producing efficient systems software. We submit the benchmark data, evaluation infrastructure, and complete logs of all LLMs-generated code at https://anonymous.4open.science/r/perfcodebench-7CDE.
Authors: Shijie Lian, Bin Yu, Xiaopeng Lin, Changti Wu, Hang Yuan, Xiaolin Hu, Zhaolong Shen, Yuzhuo Miao, Haishan Liu, Yuxuan Tian, Yukun Shi, Cong Huang, Kai Chen
Abstract: Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.
Authors: Jingjing Wang, Xiwen Chen, Wenhui Zhu, Huayu Li, Zhengxiao He, Feiyang Cai, Ana S. Carreon-Rascon, Xuanzhao Dong, Feng Luo
Abstract: LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.
Authors: Yihong Dong, Jiaru Qian, Haoran Zhang, Peixu Wang, Binhua Li, Zhi Jin, Yongbin Li, Ge Li, Xiaokang Yang, Xue Jiang
Abstract: The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.
Authors: Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Zixuan Xie, Chiming Duan, Minghua He, Philip S. Yu, Ying Li
Abstract: Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose \textsc{QuantEvolver}, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, \textsc{QuantEvolver} converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, \textsc{QuantEvolver} constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of \textsc{QuantEvolver}, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.
Authors: Shangjian Yin, Yu Fu, Yue Dong, Zhouxing Shi
Abstract: Post-training has become a crucial step for unlocking the capabilities of large language models, with reinforcement learning (RL) emerging as a critical paradigm. Recent RL-based post-training has increasingly split into two paradigms: reinforcement learning from human feedback (RLHF), which optimizes models using human preference signals in target domains, and reinforcement learning from verifiable rewards (RLVR), which operates in verifier-backed environments. The latter has dominated recent reasoning-oriented post-training because it delivers stronger gains and higher efficiency on domain-specific tasks (e.g., reasoning). However, although in-domain RL training achieves promising performance, it still requires a substantial amount of GPU compute, which remains a major barrier to broad adoption. In this work, we study the generalization ability of RLHF learned from scratch from a small set of interactions in open-ended environments, and investigate whether the conversational abilities it explicitly acquires can implicitly transfer to downstream tasks such as mathematical reasoning and code generation, namely GRLO. Specifically, on Qwen3-4B-Base backbone, GRLO improves the average performance across all domains from 24.1 to 63.1 with only 5K prompts and 22.7 GPU hours, requiring about $46\times$ less data and $68\times$ less compute than a strong in-domain RLVR baseline. The resulting model is even competitive with Qwen's released post-trained models which required a much larger training cost. Notably, a subsequent in-domain RLVR stage brings only selective gains, mainly on harder competition-math benchmarks. We hope GRLO offers a simple and efficient recipe for building broadly capable post-trained models. Our code and data will be available at: \href{https://github.com/SJY8460/GRLO}{https://github.com/SJY8460/GRLO}.
URLs: https://github.com/SJY8460/GRLO, https://github.com/SJY8460/GRLO
Authors: Ceyu Xu, Jiangnan Yu, Yongji Wu, Yuan Xie
Abstract: The quadratic complexity of attention imposes severe memory and computational bottlenecks on Large Language Model (LLM) inference. This challenge is particularly acute for emerging agentic applications that require processing multi-million token sequences. We propose STS, a sparse attention mechanism that requires no model retraining. STS leverages the key insight that tokens identified as important by a smaller draft model are highly predictive of important tokens for a larger target model. By integrating into speculative decoding frameworks, STS repurposes the draft model's attention scores to dynamically construct a token-and-head-wise sparsity mask. This mask effectively prunes the expensive attention computation in the target LLM. Our evaluation shows that STS achieves a 2.67x speedup operating at approximately 90% sparsity on representative benchmark NarrativeQA, maintaining negligible accuracy degradation compared to dense attention. STS establishes a new state-of-the-art on the sparsity-accuracy trade-off, outperforming prior techniques by enabling higher sparsity levels for a given accuracy budget.
Authors: Jaehun Jung, Hyunwoo Kim, Brandon Cui, Ximing Lu, David Acuna, Prithviraj Ammanabrolu, Yejin Choi
Abstract: Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($\Delta$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.
Authors: Xuechen Zhang, Zijian Huang, Kai Yang, Weijia Zhang, Jiasi Chen, Samet Oymak
Abstract: Modern language models often need to optimize a primary accuracy objective while also accommodating secondary behavioral preferences, such as verbosity, agreeableness, or the level of technical expertise in its response. In practice, a base model may exhibit a desired behavior very rarely or not at all. Thus, endowing the model with a target behavior creates a sparse behavioral reward bottleneck. To address such multi-objective problems, we introduce Vector-Steered Policy Optimization (VSPO) which employs a steering vector associated with the target behavior to control the behavior intensity of the generated rollouts. VSPO is obtained by modifying GRPO to sample rollouts with varying steering intensities. This process can be interpreted as an on-policy latent self-distillation procedure where the model internalizes its steering vector. By varying steering intensities, VSPO upsamples rare behaviors and enriches rollout diversity, which alleviates the sparse reward issue and provably accelerates the policy optimization. Through comprehensive theory and experiments, we establish that VSPO has favorable properties compared to vanilla reward shaping and other alternative approaches. Specifically, under a bandit abstraction, VSPO provably achieves better iteration complexity than reward-shaped GRPO when the steering-induced distributions are sufficiently aligned with the target behavior. We evaluate VSPO across multiple reasoning benchmarks, including MATH and MMLU-Pro, for four target behaviors: explanation expertise, confidence expression, robustness to misleading context, and response verbosity. Our results show that VSPO consistently improves the control along target behavior while maintaining or improving task accuracy compared with reward shaping, teacher-trace distillation, and guidance-based baselines.
Authors: Chanuk Lee, Sangwoo Park, Minki Kang, Sung Ju Hwang
Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a scalable paradigm for improving the reasoning capabilities of large language models. However, its effectiveness is fundamentally limited by exploration: the policy can only improve on trajectories it has already sampled. While increasing the number of rollouts alleviates this issue, such brute-force scaling is computationally expensive, and existing approaches that modify the optimization objective provide limited control over what is explored. In this work, we propose NudgeRL, a framework for structured and diversity-driven exploration in RLVR. Our approach introduces Strategy Nudging, which conditions each rollout on lightweight, strategy-level contexts to induce diverse reasoning trajectories without relying on expensive oracle supervision. To effectively learn from such structured exploration, we further propose a unified objective, which decomposes the reward signal into inter- and intra-context components and incorporates a distillation objective to transfer discovered behaviors back to the base policy. Empirically, NudgeRL outperforms standard GRPO with up to 8 times larger rollout budgets, while outperforming oracle-guided RL baseline on average across five challenging math benchmarks. These results demonstrate that structured, context-driven exploration can serve as an efficient and scalable alternative to both brute-force rollout scaling and feasibility-oriented methods based on privileged information. Our code is available at https://github.com/tally0818/NudgeRL.
Authors: Sihan Fu, Oucheng Liu, Shiyuan Wang, Jin Shi, Chengkun Wei
Abstract: Code agents increasingly help developers work with unfamiliar repositories, but every such task depends on a costly prerequisite: bootstrapping the repository into a usable development state. This process requires substantial trial-and-error exploration, yet the resulting knowledge--resolved dependencies, repair strategies--stays trapped in a single conversation, unavailable to future agents. We therefore formulate repository bootstrapping as a reusable startup knowledge problem and introduce BootstrapAgent, a multi-agent framework that distills the heuristics discovered during bootstrap exploration into a persistent, verifiable, agent-consumable .bootstrap contract. Through evidence extraction, structured planning, deterministic Docker-based verification, and trace-driven repair, BootstrapAgent generates a contract covering environment setup, diagnostic checks, minimal verification, and accumulated repair knowledge. We further propose warm repair with clean replay to accelerate iterative debugging without sacrificing cold-start reproducibility, and a delta repair with sanity check to prevent reward hacking. Experiments on three benchmarks show that BootstrapAgent achieves a 92.9% success rate, outperforming the baseline by over 10% while reducing downstream agent token usage by 25.9% and build time by 22.3%. Our code is available at https://github.com/Vossera/BootstrapAgent.
Authors: Rifat Mehreen Amin, Alperen Adatepe, Daniela Fernandes, Daniel Buschek, Andreas Butz
Abstract: Conversational interfaces powered by large language models (LLMs) are widely used for ideation and analysis, yet their linear structure limits exploration of alternatives and management of long-running interactions. We present CanvasConvo, a conversational interface concept that transforms linear chat into a branching conversation tree embedded in a spatial canvas. CanvasConvo enables users to explore what-if scenarios by branching directly from conversational content, supporting parallel development of alternative directions. These branches are visualized on a canvas while remaining integrated with a familiar chat interface, allowing users to switch between linear and non-linear interaction. Features such as timeline-based navigation, automatic tagging and summarization, and context-aware controls (e.g., goals, reusable prompts) support structured interaction and continuity. We evaluated CanvasConvo in a 5-7 day field study with 24 participants. Our findings highlight how non-linear conversational structures support exploratory workflows and different interactions in LLM-based work.
Authors: Chufan Shi, Cheng Yang, Yaokang Wu, Linhao Jin, Bo Shui, Taylor Berg-Kirkpatrick, Xuezhe Ma
Abstract: Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning. Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns? We investigate this via VisualSwap, an image-swap probing framework: after a model reasons over an image, we replace it with a visually similar but semantically different one and test whether the model notices. We introduce VS-Bench, 800 image pairs curated from MathVista, MathVerse, MathVision, and MMMU-Pro. Experiments on Qwen3-VL, Kimi-VL, and ERNIE-VL reveal a striking failure: models overwhelmingly miss the swap, with accuracy dropping by up to 60%. Counterintuitively, thinking models are nearly 3x more vulnerable than their instructed counterparts, and scaling offers no mitigation. Multi-turn user instructions restore visual grounding, but self-generated reflective statements during continuous generation do not. Attention analysis explains why: user instructions substantially elevate attention to visual tokens, whereas self-reflection does not. Current VLMs tend to say rather than actually see when claiming to perform visual re-examination. Our code and dataset are available at the project page: https://visualswap.github.io
Authors: Minseo Kim
Abstract: AI emotional companions face a safety-rapport paradox: restrictive safeguards can damage supportive alliance, while permissive systems risk user harm. We present SLIP (Staged Layers of Intervention Protocol), a four-stage graduated methodology deriving interventions (none, soft, hard) from structured qualitative indicators -- affect intensity (a) and narrative dynamism (m) -- alongside ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals), a "signals not labels" taxonomy. An evaluation combining a small-scale production deployment (N=68 entries, 10 users, 10 weeks) with a synthetic persona battery (N=91, 5 behavioral-risk profiles) achieved 0% false positives for the flow persona and showed expected escalation patterns in crisis-oriented personas. However, initial results showed that 8 consecutive days of high-energy elevation produced zero interventions (0/8), exposing a boundary where the "do not pathologize" principle conflicts with safety. A subsequent three-model stress test demonstrated that increased model capability improves detection from 0/8 to 6/8 while preserving 0/10 flow false positives in the largest model. Read as preliminary, these findings position graduated intervention as a design direction for navigating -- not resolving -- the safety-rapport tension in affective computing.
Authors: Parisa Kordjamshidi, Samer Aslan, Madhavan Seshadri, Leslie Barrett, Enrico Santus
Abstract: Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.
Authors: Ziang Ye, Wentao Shi, Yuxin Liu, Yu Wang, Zhengzhou Cai, Yaorui Shi, Qi Gu, Xunliang Cai, Fuli Feng
Abstract: Large language model based agents often fail in unfamiliar environments due to premature exploitation: a tendency to act on prior knowledge before acquiring sufficient environment-specific information. We identify autonomous exploration as a critical yet underexplored capability for building adaptive agents. To formalize and quantify this capability, we introduce Exploration Checkpoint Coverage, a verifiable metric that measures how broadly an agent discovers key states, objects, and affordances. Our systematic evaluation reveals that agents trained with standard task-oriented reinforcement learning consistently exhibit narrow and repetitive behaviors that impede downstream performance. To address this limitation, we develop a training strategy that interleaves task-execution rollouts and exploration rollouts, with each type of rollout optimized by its corresponding verifiable reward. Building on this training strategy, we propose the Explore-then-Act paradigm, which decouples information-gathering from task execution: agents first utilize an interaction budget to acquire grounded environmental knowledge, then leverage it for task resolution. Our results demonstrate that learning to systematically explore is imperative for building generalizable and real-world-ready agents.
Authors: Igor Bogdanov, Chung-Horng Lung, Thomas Kunz, Jie Gao, Adrian Taylor, Marzia Zaman
Abstract: Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\times$ worse mean return while using 1.8-2.7$\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.
Authors: Tahreem Yasir, Wenbo Li, Sam Gilson, Sutapa Dey Tithi, Xiaoyi Tian, Tiffany Barnes
Abstract: Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.
Authors: Xavier Theimer-Lienhard, Mushtaha El-Amin, Fay Elhassan, Sahaj Vaidya, Victor Cartier-Negadi, David Sasu, Lars Klein, Mary-Anne Hartley
Abstract: Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most "open" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.
Authors: Igor Bogdanov, Chung-Horng Lung, Thomas Kunz, Jie Gao, Adrian Taylor, Marzia Zaman
Abstract: Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. FORGE wraps a Reflexion-style inner loop, where a dedicated reflection agent (using the same underlying LLM, no distillation from a stronger model) converts failed trajectories into reusable knowledge artifacts: textual heuristics (Rules), few-shot demonstrations (Examples), or both (Mixed), with an outer loop that propagates the best-performing instance's memory to the population between stages and freezes converged instances via a graduation criterion. We evaluate on CybORG CAGE-2, a stochastic network-defense POMDP at a 30-step horizon against the B-line attacker, where all four tested LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, Qwen3-235B) exhibit strongly negative, heavy-tailed zero-shot rewards. Compared against both a zero-shot baseline and a Reflexion baseline (isolated single-stream learning), FORGE improves average evaluation return by 1.7-7.7$\times$ over zero-shot and by 29-72% over Reflexion in all 12 model-representation conditions, reducing major-failure rates (below $-100$) to as low as $\sim$1%. We find that (1) population broadcast is critical mechanism, with a no-graduation ablation confirming that broadcast carries the performance gains while graduation primarily saves compute; (2) Examples achieves the strongest returns for three of four models, Rules offers the best cost-reliability profile with $\sim$40% fewer tokens; and (3) weaker baseline models benefit disproportionately, suggesting FORGE may mitigate capability gaps rather than amplify strong models. All evidence is confined to CAGE-2 B-line; cross-family findings are directional evidence.
Authors: Gabriel Garcia
Abstract: When researchers ask whether two transformer layers are "equivalent" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.
Authors: Stratis Tsirtsis, Kai Rawal, Chris Russell, Brent Mittelstadt, Sandra Wachter
Abstract: Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the "Explain this post" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.
Authors: Christopher Barrie, Elli Palaiologou, Petter T\"ornberg
Abstract: Researchers are increasingly using language models (LMs) for text annotation. These approaches rely only on a prompt telling the model to return a given output according to a set of instructions. The reproducibility of LM outputs may nonetheless be vulnerable to small changes in the prompt design. This calls into question the replicability of classification routines. To tackle this problem, researchers have typically tested a variety of semantically similar prompts to determine what we call ``prompt stability." These approaches remain ad-hoc and task specific. In this article, we propose a general framework for diagnosing prompt stability by adapting traditional approaches to intra- and inter-coder reliability scoring. We call the resulting metric the Prompt Stability Score (PSS) and provide a Python package \texttt{promptstability} for its estimation. Using six different datasets and twelve outcomes, we classify $\sim$3.1m rows of data and $\sim$300m input tokens to: a) diagnose when prompt stability is low; and b) demonstrate the functionality of the package. We conclude by providing best practice recommendations for applied researchers.
Authors: Preetika Verma, Kokil Jaidka
Abstract: We present MediaSpin, a large-scale language resource capturing how major news outlets modify headlines after publication, and MediaSpin-in-the-Wild, a complementary dataset linking these revised headlines to their downstream engagement on social media. The increasing editability of online news headlines offers new opportunities to study linguistic framing and bias through the lens of editorial revisions. The dataset contains 78,910 headline pairs annotated for 13 types of media bias, grounded in established media-bias taxonomies, covering both subjective (e.g., sensationalism, spin) and objective (e.g., omission, slant) forms, with annotation conducted through a human-supervised large-language-model pipeline with expert validation and quality control. We describe the annotation schema and demonstrate three downstream applications: (1) cross-national analysis of how country references are added or removed during editing, (2) transformer-based bias classification at both binary and fine-grained levels, and (3) behavioral analysis of biased headlines on X (Twitter) using 180,786 news-related tweets from 819 consenting users. The results reveal regional asymmetries in representational framing, measurable linguistic markers, and consistently higher engagement with biased content. MediaSpin and MediaSpin-in-the-Wild together provide a reproducible benchmark for bias detection and the study of editorial and behavioral dynamics in contemporary media ecosystems.
Authors: Atsushi Suzuki, Yulan He, Feng Tian, Zhongyuan Wang
Abstract: Hallucinations, a phenomenon where a language model (LM) generates nonfactual content, pose a significant challenge to the practical deployment of LMs. While many empirical methods have been proposed to mitigate hallucinations, recent studies established a computability-theoretic result showing that any LM will inevitably generate hallucinations on an infinite set of inputs, regardless of the quality and quantity of training datasets and the choice of the language model architecture and training and inference algorithms. Although the computability-theoretic result may seem pessimistic, its significance in practical viewpoints has remained unclear. This paper claims that those "innate" inevitability results from computability theory and diagonal argument, in principle, cannot explain practical issues of LLMs. We demonstrate this claim by presenting a positive theoretical result from a probabilistic perspective. Specifically, we prove that hallucinations can be made statistically negligible, provided that the quality and quantity of the training data are sufficient. Interestingly, our positive result coexists with the computability-theoretic result, implying that while hallucinations on an infinite set of inputs cannot be entirely eliminated, their probability can always be reduced by improving algorithms and training data. By evaluating the two seemingly contradictory results through the lens of information theory, we argue that our probability-theoretic positive result better reflects practical considerations than the computability-theoretic negative result.
Authors: Yash Akhauri, Ahmed F AbouElhamayed, Yifei Gao, Chi-Chih Chang, Sameh Gobriel, Nilesh Jain, Mohamed S. Abdelfattah
Abstract: Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck. However, there is an opportunity to alleviate this bottleneck, prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks of tokens and many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. TokenButler predicts low-dimensional importance queries at a fixed depth stride, and combines them with a learned projection of the real KV-cache keys to score tokens cheaply, enabling dynamic per-token selection under a fixed budget while preserving the full KV cache. We train TokenButler by distilling the model's masked causal attention distributions, optimizing a lightweight predictor with minimal parameter overhead. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy where existing methods fail. Furthermore, TokenButler achieves competitive or superior performance on long-context benchmarks (RULER, LongBench), up to $\approx1.6\times$ on-GPU speedup using our proposed *prediction interval with neighbor fetching* that amortizes predictor cost while maintaining accuracy within $\approx$1.1\%, and up to 7.6$\times$ reduction in latency compared to Dense Attention with CPU offloading. Code is available: https://github.com/abdelfattah-lab/TokenButler
Authors: Andrea W Wen-Yi, Unso Eun Seo Jo, David Mimno
Abstract: The release of top-performing open-weight LLMs has cemented China's role as a leading force in AI development. Do these models support languages spoken in China? Or do they support the same languages as models developed in the United States or in Europe? Comparing multilingual capabilities is important for two reasons. First, language ability provides insights into pre-training data curation, and thus into resource allocation and development priorities. Second, Chinese model developers need to navigate the tension between serving a linguistically diverse population domestically, and optimizing for globally visible benchmarks that are predominantly English. We investigate Chinese model developers' priorities through a comparative study of Chinese-developed and Western-developed open-weight LLMs, on 21 language variants including Asian regional, Chinese, and European languages. Our experiments on Information Parity and reading comprehension show Chinese models' performance across these languages correlates strongly (r=0.93) with their Western counterparts, with the sole exception being better Mandarin. Chinese-developed models are good at French and German, but they sometimes cannot identify languages spoken by Chinese minorities such as Kazakh and Uyghur. Overall, all open-weight LLMs we study have a similar multilingual performance profile, despite the diverse linguistic and cultural contexts the model developers operated within. We interpret the homogenization as consistent with the influence of global benchmarking practices and shared training resources. Rather than treating current language support as inevitable, our results highlight multilingual development as a space of prioritization and trade-offs, with implications for model developers, policymakers, and users.
Authors: Yuyang Xu, Renjun Hu, Haochao Ying, Jian Wu, Xing Shi, Wei Lin
Abstract: Benchmark-based evaluation, e.g., multiple-choice questions (MCQs) and open-ended questions (OEQs), is widely used for evaluating Large Language Models (LLMs), yet their reliability is undermined by benchmark contamination. When pre-exposed to the testing benchmark during training, less capable LLMs have been found to achieve inflated performance, thereby yielding erroneous results in LLM evaluation. In this study, we reframe contamination as an inherent aspect of learning and seek to disentangle and expose genuine capability acquisition from superficial memorization in LLM evaluation. Following this, firstly, by analyzing model performance under different memorization conditions of MCQs, we uncover a counterintuitive trend: LLMs perform worse on memorized benchmarks than on non-memorized ones, indicating the coexistence of two learning phenomena, i.e., rote memorization and genuine capability learning. To disentangle them, we propose TrinEval, a novel evaluation framework that reformulates MCQs into an alternative knowledge-centric trinity format, reducing memorization while preserving inherent knowledge, enabling the evaluation of genuine capability in the presence of memorization. Extensive experiments validate the effectiveness and robustness of TrinEval in reformulating benchmarks, and the evaluation results further reveal that mainstream LLMs rely on rote memorization for an average of 19.6% of knowledge points across the MMLU and the GSM8K dataset.
Authors: Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao
Abstract: Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO typically rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address this limitation, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.
Authors: Alexander Shabalin, Viacheslav Meshchaninov, Dmitry Vetrov
Abstract: Diffusion models have achieved state-of-the-art performance in generating images, audio, and video, but their adaptation to text remains challenging due to its discrete nature. Prior approaches either apply Gaussian diffusion in continuous latent spaces, which inherits semantic structure but struggles with token decoding, or operate in categorical simplex space, which respect discreteness but disregard semantic relation between tokens. In this paper, we propose Smoothing Diffusion on Token Embeddings (Smoothie), a novel diffusion method that combines the strengths of both approaches by progressively smoothing token embeddings based on semantic similarity. This technique enables gradual information removal while maintaining a natural decoding process. Experimental results on several sequence-to-sequence and unconditional generation tasks demonstrate that Smoothie outperforms existing diffusion-based models in generation quality. Furthermore, ablation studies show that our proposed diffusion space yields better performance than both the standard embedding space and the categorical simplex. The code is available at https://github.com/ashaba1in/smoothie.
Authors: Pierre-Carl Langlais, Pavel Chizhov, Catherine Arnett, Carlos Rosas Hinostroza, Mattia Nee, Eliot Krzystof Jones, Ir\`ene Girard, David Mach, Anastasia Stasenko, Ivan P. Yamshchikov
Abstract: Large Language Models (LLMs) are pre-trained on large amounts of data from different sources and domains. Such datasets often contain trillions of tokens, including large portions of copyrighted or proprietary content, which raises questions about the legal use of such models. This underscores the need for truly open pre-training data that complies with data security regulations. In this paper, we introduce Common Corpus, the largest open dataset for LLM pre-training. The data assembled in Common Corpus are either uncopyrighted or under open licenses, totaling about two trillion tokens. The dataset contains a wide variety of languages, ranging from the high-resource European languages to some low-resource languages rarely represented in pre-training datasets. In addition, it includes a large amount of code data. The diversity of data sources in terms of covered domains and time periods opens up the paths for both research and entrepreneurial needs across diverse areas of knowledge. In this paper, we present the detailed provenance of data assembling and the details of dataset filtering and curation. We train two small language models on Common Corpus and find that they perform comparably to other models of their size, indicating that our dataset is suitable for multilingual pretraining. Common Corpus represents a key contribution to the ecosystem for open science research on Large Language Models.
Authors: Shengji Tang, Jianjian Cao, Weihao Lin, Jiale Hong, Bo Zhang, Shuyue Hu, Lei Bai, Tao Chen, Wanli Ouyang, Peng Ye
Abstract: Existing multi-LLM collaboration systems often encounter scalability challenges when integrating new LLMs and tasks, leading to suboptimal performance. To address this, we propose SMCS, a Scalable Multi-LLM Collaboration System designed to effectively coordinate multiple open-source LLMs. The system consists of two core components: a Retrieval-based Prior Selection (RPS) module, which dynamically selects the most suitable LLMs for each input, and an Exploration-Exploitation-Driven Posterior Enhancement (EPE) module, which fosters response diversity and selects high-quality outputs through a hybrid scoring mechanism. Experiments on eight mainstream benchmarks validate the effectiveness of our system: by integrating fifteen open-source LLMs, SMCS outperforms prevailing closed-source LLMs, e.g., GPT-4.1(+5.36%) and GPT-o3-mini(+5.28%) across multiple tasks. Remarkably, it even exceeds the average of best results on different datasets with open-source LLMs (+2.86%), significantly advancing the empirical performance frontier of open-source collaboration. The code is released at https://github.com/magent4aci/SMCS.
Authors: Jiakang Wang, Runze Liu, Fuzheng Zhang, Xiu Li, Guorui Zhou, Ling Pan
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become an effective post-training method for improving the reasoning abilities of Large Language Models (LLMs). However, existing methods mainly apply uniform optimization constraints across all tokens, ignoring their heterogeneous roles. Prior work shows that high-entropy tokens are closely tied to reasoning, while low-entropy tokens primarily encode factual knowledge, and recent approaches attempt to exploit this distinction by isolating token updates via masking or asynchronous training. We argue that such isolation breaks the sequential dependency structure of autoregressive generation, leading to suboptimal learning. To address this, we propose \textbf{Archer}, an entropy-aware RLVR framework with \textbf{dual-token constraints} that preserves joint optimization while modulating update strength across token types. Our method introduces response-level entropy normalization for stable token classification and applies differentiated clipping ranges and KL regularization to encourage exploration on reasoning tokens while preserving knowledge tokens. Experiments on mathematical reasoning and code generation benchmarks show that Archer consistently outperforms strong baselines across multiple model scales, improving both \textit{pass@1} and \textit{pass@K} performance. These results highlight the importance of respecting sequence-level dependencies when designing fine-grained RL optimization strategies for LLMs.
Authors: Deep Anil Patel, Iain Melvin, Christopher Malon, Martin Renqiang Min
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating human-like text, yet they largely operate as reactive agents, responding only when directly prompted. This passivity creates an "awareness gap," limiting their potential as truly collaborative partners in dynamic human discussions. We introduce $\textit{DiscussLLM}$, a framework designed to bridge this gap by training models to proactively decide not just $\textit{what}$ to say, but critically, $\textit{when}$ to speak. Our primary contribution is a scalable two-stage data generation pipeline that synthesizes a large-scale dataset of realistic multi-turn human discussions. Each discussion is annotated with one of five intervention types (e.g., Factual Correction, Concept Definition) and contains an explicit conversational trigger where an AI intervention adds value. By training models to predict a special silent token when no intervention is needed, they learn to remain quiet until a helpful contribution can be made. We explore two architectural baselines: an integrated end-to-end model and a decoupled classifier-generator system optimized for low-latency inference. We evaluate these models on their ability to accurately time interventions and generate helpful responses, paving the way for more situationally aware and proactive conversational AI.
Authors: Sasha Cui, Zhongren Chen
Abstract: Language models (LMs) are typically post-trained for desired capabilities and behaviors via weight-based or prompt-based steering, but the former is time-consuming and expensive, and the latter is not precisely controllable and often requires manual trial-and-error. While activation steering (AS) promises a cheap, fast, and controllable alternative to the two existing post-training methods, current AS techniques require hand-crafted prompt pairs or labor-intensive feature annotation, making them more inconvenient than the plug-and-play methods such as Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT). We introduce Painless Activation Steering (PAS), a family of fully automated methods that make AS readily usable with any given labeled dataset, with no need for prompt construction, feature labeling, or human intervention. We evaluate PAS on three open-weight models (Llama3.1-8B-Instruct, DeepSeek-R1-Distill-8B, and Nous-Hermes-2) and 18 tasks; we find that PAS reliably improves performance for behavior tasks, but not for intelligence-oriented tasks. The introspective variant (iPAS) delivers the strongest causal steering effects (10.1% on Bias, 5.2% on Morality, and 34.8% on Alignment). We also show PAS delivers additional gains on top of In-Context Learning (ICL) and SFT. PAS constructs a fast, lightweight activation vector that can be cheaply trained, easily stored, and activated at will. Our results provide a characterization of where AS helps, where it fails, and how to deploy it as a practical, automated LM post-training option.
Authors: Jiakang Wang, Runze Liu, Qingpeng Cai, Lei Lin, Wenping Hu, Xiu Li, Fuzheng Zhang, Guorui Zhou, Kun Gai, Ling Pan
Abstract: Reinforcement learning (RL) has shown great promise in large language models (LLMs) post-training, which typically rely on token-level clipping to maintain stability during optimization. Despite the empirical success of GRPO-style methods, we identify a fundamental and previously overlooked challenge in this popular Outcome-Supervised RL (OSRL) paradigm. We reveal that in OSRL, where advantages are shared across tokens within a response, importance sampling (IS) ratios deviate from their traditional purpose of distribution correction as in classic RL, which become token-level weights that allocate the shared advantage signal across tokens. We show that this hidden role shift induces a critical mismatch for positive-advantage tokens, leading to unbalanced token weighting between positive and negative tokens. Specifically, it suppresses the update of underrepresented tokens that are lagging behind, while over-amplifying already high-probability tokens. This mismatch results in rich-get-richer dynamics that over-reinforce confident tokens, weaken catch-up learning that drive entropy collapse, excessive repetition, and premature convergence. To address this, we propose Asymmetric Importance Sampling Policy Optimization (ASPO), a simple yet effective strategy that reverses the ratio-induced weighting of positive-advantage tokens, while stabilizing extreme updates and maintaining gradient flow. This mismatch correction aligns their update direction with the learning dynamics of negative ones. Comprehensive experiments across math reasoning and coding benchmarks demonstrate that ASPO significantly mitigates entropy collapse, improves training stability, and enhances performance over strong GRPO-based baselines. Our analysis provides new insights into the role of token-level weighting in OSRL and highlights the critical importance of correcting ratio-induced weighting in LLM RL.
Authors: Yutao Wu, Xiao Liu, Yinghui Li, Yifeng Gao, Yifan Ding, Jiale Ding, Xiang Zheng, Xingjun Ma
Abstract: Knowledge poisoning poses a critical threat to Retrieval-Augmented Generation (RAG) systems by injecting adversarial content into knowledge bases, tricking Large Language Models (LLMs) into producing attacker-controlled outputs grounded in manipulated context. Prior work highlights LLMs' susceptibility to misleading or malicious retrieved content. However, real-world fact-checking scenarios are more challenging, as credible evidence typically dominates the retrieval pool. To investigate this problem, we extend knowledge poisoning to the fact-checking setting, where retrieved context includes authentic supporting or refuting evidence. We propose \textbf{ADMIT} (\textbf{AD}versarial \textbf{M}ulti-\textbf{I}njection \textbf{T}echnique), a few-shot, semantically aligned poisoning attack that flips fact-checking decisions and induces deceptive justifications, all without access to the target LLMs, retrievers, or token-level control. Extensive experiments show that ADMIT transfers effectively across 4 retrievers, 11 LLMs, and 4 cross-domain benchmarks, achieving an average attack success rate (ASR) of 86\% at an extremely low poisoning rate of $0.93 \times 10^{-6}$, and remaining robust even in the presence of strong counter-evidence. Compared with prior state-of-the-art attacks, ADMIT improves ASR by 11.2\% across all settings, exposing significant vulnerabilities in real-world RAG-based fact-checking systems.
Authors: Seungbeen Lee, Jinhong Jeong, Donghyun Kim, Yejin Son, Youngjae Yu
Abstract: Our ability to interpret others' mental states through nonverbal cues (NVCs) is fundamental to our survival and social cohesion. While existing Theory of Mind (ToM) benchmarks have primarily focused on false-belief tasks and reasoning with asymmetric information, they overlook other mental states beyond belief and the rich tapestry of human nonverbal communication. We present Motion2Mind, a framework for evaluating the ToM capabilities of machines in interpreting NVCs. Leveraging an expert-curated body-language reference as a proxy knowledge base, we build Motion2Mind, a carefully curated video dataset with fine-grained nonverbal cue annotations paired with manually verified psychological interpretations. It encompasses 222 types of nonverbal cues and 397 mind states. Our evaluation reveals that current AI systems struggle significantly with NVC interpretation, exhibiting not only a substantial performance gap in Detection, as well as patterns of over-interpretation in Explanation compared to human annotators.
Authors: Rulin Shao, Akari Asai, Shannon Zejiang Shen, Hamish Ivison, Varsha Kishore, Jingming Zhuo, Xinran Zhao, Molly Park, Samuel G. Finlayson, David Sontag, Tyler Murray, Sewon Min, Pradeep Dasigi, Luca Soldaini, Faeze Brahman, Wen-tau Yih, Tongshuang Wu, Luke Zettlemoyer, Yoon Kim, Hannaneh Hajishirzi, Pang Wei Koh
Abstract: Deep research agents perform multi-step research to produce long-form, well-attributed answers. However, most open deep research agents are trained on easily verifiable short-form QA tasks via reinforcement learning with verifiable rewards, which does not extend to realistic long-form tasks. We address this with Reinforcement Learning with Evolving Rubrics (RLER), where rubrics are constructed and maintained to co-evolve with the policy model during training. This allows the rubrics to incorporate newly explored information from search and contrasting model responses, enabling better fact checking and more discriminative on-policy feedback. Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first fully open model that is directly trained for open-ended, long-form deep research. Across four long-form deep research benchmarks in science, healthcare, and general domains, DR Tulu substantially outperforms existing open deep research agents (by 15.6% over Tongyi DR on average) and matches or exceeds proprietary deep research agents (by 0.7% over OpenAI DR on average), while being significantly smaller and cheaper per query (1000x cheaper than OpenAI DR per query).
Authors: Jiacheng Guo, Suozhi Huang, Zixin Yao, Yifan Zhang, Yifu Lu, Jiashuo Liu, Zihao Li, Nicholas Deng, Qixin Xiao, Jia Tian, Kanghong Zhan, Tianyi Li, Xiaochen Liu, Jason Ge, Chaoyang He, Kaixuan Huang, Lin Yang, Wenhao Huang, Mengdi Wang
Abstract: This paper introduces CryptoBench, the first expert-curated, dynamic benchmark designed to rigorously evaluate the real-world capabilities of Large Language Model (LLM) agents in the uniquely demanding and fast-paced cryptocurrency domain. Unlike general-purpose agent benchmarks for search and prediction, professional crypto analysis presents specific challenges: \emph{extreme time-sensitivity}, \emph{a highly adversarial information environment}, and the critical need to synthesize data from \emph{diverse, specialized sources}, such as on-chain intelligence platforms and real-time Decentralized Finance (DeFi) dashboards. CryptoBench thus serves as a much more challenging and valuable scenario for LLM agent assessment. To address these challenges, we constructed a live, dynamic benchmark featuring 50 questions per month, expertly designed by crypto-native professionals to mirror actual analyst workflows. These tasks are rigorously categorized within a four-quadrant system: Simple Retrieval, Complex Retrieval, Simple Prediction, and Complex Prediction. This granular categorization enables a precise assessment of an LLM agent's foundational data-gathering capabilities alongside its advanced analytical and forecasting skills. Our evaluation of ten LLMs, both directly and within an agentic framework, reveals a performance hierarchy and uncovers a failure mode. We observe a \textit{retrieval-prediction imbalance}, where many leading models, despite being proficient at data retrieval, demonstrate a pronounced weakness in tasks requiring predictive analysis. This highlights a problematic tendency for agents to appear factually grounded while lacking the deeper analytical capabilities to synthesize information.
Authors: Jianxiang Zang, Yongda Wei, Ruxue Bai, Shiyu Jiang, Nijia Mo, Binhong Li, Qiang Sun, Hui Liu
Abstract: Reliable reward models (RMs) are critical for ensuring the safe alignment of large language models (LLMs). However, current RM evaluation methods focus solely on preference perception accuracies in given specific scenarios, obscuring the critical vulnerabilities of RMs in real-world scenarios. We identify the true challenge lies in assessing a novel dimension: Suitability, defined as conditional reliability under specific real-world perturbations. To this end, we introduce Reward Auditor, a hypothesis-testing framework specifically designed for RM suitability inference. Rather than answering "How accurate is the RM's preference perception for given samples?", it employs scientific auditing to answer: "Can we infer RMs exhibit systematic vulnerabilities in specific real-world scenarios?". Under real-world perturbed scenarios, Reward Auditor quantifies statistical significance and effect size by auditing distribution degradation of RM preference perception confidence. This enables inference of both the certainty and severity of RM vulnerabilities across diverse real-world scenarios. This lays a solid foundation for building next-generation LLM alignment systems that are verifiably safe, more robust, and trustworthy.
Authors: Guoshenghui Zhao, Huawei Lin, Weijie Zhao
Abstract: Removing specific data influence from large language models (LLMs) remains challenging, as retraining is costly and existing approximate unlearning methods are often unstable. The challenge is exacerbated when the forget set is small or imbalanced. We introduce RapidUn, an influence-driven and parameter-efficient unlearning framework. It first estimates per-sample influence through a fast estimation module, then maps these scores into adaptive update weights that guide selective parameter updates -- forgetting harmful behavior while retaining general knowledge. On Mistral-7B and Llama-3-8B across Dolly-15k and Alpaca-57k, RapidUn achieves up to 100 times higher efficiency than full retraining and consistently outperforms Fisher, GA, and LoReUn on both in-distribution and out-of-distribution forgetting. These results establish influence-guided parameter reweighting as a scalable and interpretable paradigm for LLM unlearning.
Authors: Hengxing Cai, Yijie Rao, Ligang Huang, Zanyang Zhong, Jinhan Dong, Jingjun Tan, Changhao Nai, Jue Hou, Wenhao Lu, Renxin Zhong
Abstract: Existing UAV vision-and-language navigation (VLN) benchmarks rarely provide realistic aerial scenes, natural process-level instructions, and sufficient scale simultaneously, making it difficult to systematically train and evaluate UAV VLN agents under realistic settings. To address this, we propose \textbf{AirNav}, a large-scale benchmark built on real urban aerial data, comprising 137K navigation samples with natural and diverse instructions generated via a human--LLM collaborative pipeline with 10 user personas. We conduct a systematic evaluation of representative approaches on AirNav, ranging from traditional models to multimodal large language models (MLLMs), under unified metrics with open-source implementations. We further propose \textbf{AirVLN-R1}, trained via supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), achieving state-of-the-art performance with a 51.82\% success rate on the test-unseen split. Real-world experiments on a physical UAV platform provide preliminary evidence of sim-to-real transferability, and our dataset and code are publicly available.
Authors: Manar Ali, Judith Sieker, Sina Zarrie{\ss}, Hendrik Buschmeier
Abstract: In human conversation, both interlocutors play an active role in maintaining mutual understanding. When listeners are uncertain about what speakers mean, for example, they can request clarification. It is an open question for language models whether they can assume a similar listener role, recognizing and expressing their own uncertainty through clarification. We argue that reference games are a suitable testbed to approach this question as they are controlled, self-contained, and make clarification needs explicit and measurable. To test this, we evaluate three vision-language models comparing a baseline reference resolution task to an experiment where the models are instructed to request clarification when uncertain. The results suggest that even in such simple tasks, models often struggle to recognize internal uncertainty and translate it into adequate clarification behavior. This demonstrates the value of reference games as testbeds for interaction qualities of (vision and) language models.
Authors: Boxiang Zhao, Qince Li, Zhonghao Wang, Zelin Cao, Yi Wang, Peng Cheng, Bo Lin
Abstract: As Large Language Models (LLMs) evolve into the core of Web-based autonomous agents and complex Web Information Systems, their ability to faithfully translate natural language into rigorous structured formats has become paramount, as this capability is critical for Web API invocation and data exchange. However, evaluating this structural fidelity in Web-native payloads remains a challenge: traditional text metrics fail to capture topological consistency in semi-structured Web data, while manual evaluation is prohibitively costly. To address this, we propose Structure-BiEval, a novel self-supervised framework for quantitative, annotation-free assessment tailored for Web data engineering. By leveraging deterministic Intermediate Representations, our framework effectively decouples structure from content, utilizing Content Semantic Accuracy and Normalized Tree Edit Distance as precise metrics. We empirically benchmark 15 state-of-the-art LLMs across dual Web structural topologies, namely Hierarchical Data (Web backend payloads) and Tabular Data (Web frontend presentation). The results reveal substantial variability in structural performance, including cases where mid-sized models unexpectedly outperform larger counterparts in Web data formatting. Furthermore, our findings show that deep recursive nesting poses a consistent challenge for Web agents across varying parameter scales.
Authors: Shengrui Li, Fei Zhao, Kaiyan Zhao, Jieying Ye, Haifeng Liu, Fangcheng Shi, Zheyong Xie, Yao Hu, Shaosheng Cao
Abstract: Determining an effective data mixture is a key factor in Large Language Model (LLM) pre-training, where models must balance general competence with proficiency on hard tasks such as math and code. However, identifying an optimal mixture remains an open challenge, as existing approaches either rely on unreliable tiny-scale proxy experiments or require prohibitively expensive large-scale exploration. To address this, we propose Decouple Searching from Training Mix (DeMix), a novel framework that leverages model merging to predict optimal data ratios. Instead of training proxy models for every sampled mixture, DeMix trains component models on candidate datasets at scale and derives data mixture proxies via weighted model merging. This paradigm decouples search from training costs, enabling evaluation of unlimited sampled mixtures without extra training burden and thus facilitating better mixture discovery through more search trials. Extensive experiments demonstrate that DeMix breaks the trade-off between sufficiency, accuracy and efficiency, obtaining the optimal mixture with higher benchmark performance at lower search cost. Additionally, we release the DeMix Corpora, a comprehensive 22T-token dataset comprising high-quality pre-training data with validated mixtures to facilitate open research. Our code and DeMix Corpora is available at https://github.com/Lucius-lsr/DeMix.
Authors: Changyue Wang, Weihang Su, Qingyao Ai, Yiqun Liu
Abstract: Scaling training data and model parameters has long driven progress in large language models (LLMs), but this paradigm is increasingly constrained by the scarcity of high-quality data and diminishing returns from rising computational costs. As a result, recent work is increasing the focus on continual learning from real-world deployment, where user interaction logs provide a rich source of authentic human feedback and procedural knowledge. However, learning from user logs is challenging due to their unstructured and noisy nature. Vanilla LLM systems often struggle to distinguish useful feedback signals from noisy user behavior, and the disparity between user log collection and model optimization (e.g., the off-policy optimization problem) further strengthens the problem. To this end, we propose UNO (User log-driveN Optimization), a unified framework for improving LLM systems (LLMsys) with user logs. UNO first distills logs into semi-structured rules and preference pairs, then employs query-and-feedback-driven clustering to manage data heterogeneity, and finally quantifies the cognitive gap between the model's prior knowledge and the log data. This assessment guides the LLMsys to adaptively filter out noisy feedback and construct different modules for primary and reflective experiences extracted from user logs, thereby improving future responses. Extensive experiments show that UNO achieves state-of-the-art effectiveness and efficiency, significantly outperforming Retrieval Augmented Generation (RAG) and memory-based baselines. We have open-sourced our code at https://github.com/bebr2/UNO .
Authors: Szilvia Ujv\'ary, Louis B\'ethune, Pierre Ablin, Jo\~ao Monteiro, Marco Cuturi, Michael Kirchhof
Abstract: Language models have consistently grown to compress more world knowledge into their parameters, but the knowledge that can be pretrained into them is upper-bounded by their parameter size. Especially the capacity of Small Language Models (SLMs) is limited, leading to factually incorrect generations. This problem is often mitigated by giving the SLM access to an outside source: the ability to query a larger model, documents, or a database. Under this setting, we study the fundamental question of \emph{which tokens an SLM can and should learn} during pretraining, versus \emph{which ones it should delegate} via a \texttt{
Authors: Tunyu Zhang, Xinxi Zhang, Ligong Han, Haizhou Shi, Xiaoxiao He, Zhuowei Li, Hao Wang, Kai Xu, Akash Srivastava, Chengzhi Mao, Hao Wang, Vladimir Pavlovic, Dimitris N. Metaxas
Abstract: Diffusion large language models (DLLMs) have emerged as powerful generative models with the promise of fast text generation through parallel decoding. However, realizing this potential in practice remains challenging: reducing the number of decoding steps, typically causes a substantial degradation in output quality due to token factorization error. To alleviate this, we propose a self-distillation framework that trains a few-step student to match the generative trajectory of a full-step teacher. We theoretically and empirically show that trajectory-level supervision mitigates this factorization error, thereby enabling effective few-step decoding. We further incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that encourages mode-seeking toward the teacher's modes, yielding stronger performance on challenging reasoning tasks. Across reasoning and code-generation benchmarks, our method substantially narrows the gap between few-step and full-step decoding. The source code is available at https://github.com/Tyrion58/T3D.
Authors: Jonathan Steinberg, Oren Gal
Abstract: Vision-language models (VLMs) can read text from images, but where does this optical character recognition (OCR) information enter the language processing stream? We investigate the OCR routing mechanism across three architecture families (Qwen3-VL, Phi-4, InternVL3.5) using causal interventions. By computing activation differences between original images and text-inpainted versions, we identify architecture-specific OCR bottlenecks whose dominant location depends on the vision-language integration strategy: DeepStack models (Qwen) show peak sensitivity at mid-depth (about 50%) for scene text, while single-stage projection models (Phi-4, InternVL) peak at early layers (6-25%), though the exact layer of maximum effect varies across datasets. The OCR signal is remarkably low-dimensional: PC1 captures up to 72.9% of variance. Crucially, principal component analysis (PCA) directions learned on one dataset transfer to others, demonstrating shared text-processing pathways. Surprisingly, in models with modular OCR circuits (notably Qwen3-VL-4B), OCR removal can improve counting performance (up to +6.9 percentage points), suggesting OCR interferes with other visual processing in sufficiently modular architectures.
Authors: Zeyi Lu, Angelica Henestrosa, Pavel Chizhov, Ivan P. Yamshchikov
Abstract: Large Language Models (LLMs) often exhibit highly agreeable and reinforcing conversational styles, also known as AI-sycophancy. Although this pattern arises from training objectives that reward user satisfaction over accuracy, it may become problematic when interacting with user prompts that reflect negative social tendencies. Such responses risk amplifying harmful behavior rather than mitigating it. In this study, we examine how LLMs respond to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, and Psychopathy) using a curated dataset. Our analysis reveals differences across models, whereby all models predominantly exhibit corrective behavior, while showing reinforcing output in certain cases. Model behavior also depends on the severity level and differs in the sentiment of the response. Our findings raise implications for designing safer conversational systems that can detect and respond appropriately when users escalate from benign to harmful requests.
Authors: Priyaranjan Pattnayak, Sanchari Chowdhuri
Abstract: As large language models (LLMs) are deployed in multilingual settings, their safety behavior in culturally diverse, low-resource languages remains poorly understood. We present the first systematic evaluation of LLM safety across 12 Indic languages, spoken by over 1.2 billion people but underrepresented in LLM training data. Using a dataset of 6,000 culturally grounded prompts spanning caste, religion, gender, health, and politics, we assess 10 leading LLMs on translated variants of the prompt. Our analysis reveals significant safety drift: cross-language agreement is just 12.8\%, and \texttt{SAFE} rate variance exceeds 17\% across languages. Some models over-refuse benign prompts in low-resource scripts, overflag politically sensitive topics, while others fail to flag unsafe generations. We quantify these failures using prompt-level entropy, category bias scores, and multilingual consistency indices. Our findings highlight critical safety generalization gaps in multilingual LLMs and show that safety alignment does not transfer evenly across languages. We release \textsc{IndicSafe}, the first benchmark to enable culturally informed safety evaluation for Indic deployments, and advocate for language-aware alignment strategies grounded in regional harms.
Authors: Suyuan Liu, Molly Babel
Abstract: We introduce the Mandarin-English Language Interview (MELI) Corpus, an open-source resource of 29.8 hours of speech from 51 Mandarin-English bilingual speakers. MELI combines matched sessions in Mandarin and English with two speaking styles: read sentences and spontaneous interviews about language varieties, standardness, and learning experiences. Audio was recorded at 44.1 kHz (16-bit, stereo). Interviews were fully transcribed, force-aligned at word and phone levels, and anonymized. Descriptively, the Mandarin component totals ~14.7 hours (mean duration 17.3 minutes) and the English component ~15.1 hours (mean duration 17.8 minutes). We report token/type statistics for each language and document code-switching patterns (frequent in Mandarin sessions; more limited in English sessions). The corpus design supports within-/cross-speaker, within/cross-language acoustic comparison and links acoustics to speakers' stated language attitudes, enabling both quantitative and qualitative analyses. The MELI Corpus will be released with transcriptions, alignments, metadata, scans of labelled maps and documentation under a CC BY-NC 4.0 license.
Authors: Fan Zhang (Steve), Mingzi Song (Steve), Rania Elbadry (Steve), Yankai Chen (Steve), Shaobo Wang (Steve), Yixi Zhou (Steve), Xunwen Zheng (Steve), Yueru He (Steve), Yuyang Dai (Steve), Georgi Georgiev (Steve), Ayesha Gull (Steve), Muhammad Usman Safder (Steve), Fan Wu (Steve), Liyuan Meng (Steve), Fengxian Ji (Steve), Junning Zhao (Steve), Xueqing Peng (Steve), Jimin Huang (Steve), Yu Chen (Steve), Xue (Steve), Liu, Preslav Nakov, Zhuohan Xie
Abstract: Financial reporting systems increasingly leverage Large Language Models (LLMs) to extract and summarize corporate disclosures. However, most existing approaches assume a single-market setting and overlook structural differences across jurisdictions. Variations in accounting taxonomies, tagging infrastructures (e.g., XBRL vs.\ PDF), and aggregation conventions introduce substantial challenges for semantic alignment and reliable verification. Here, we aim to bridge this gap. We present FinReporting, an agentic workflow for localized cross-jurisdiction financial reporting. The system constructs a unified canonical ontology spanning the income statement, balance sheet, and cash flow statement, and decomposes reporting into auditable stages, including filing acquisition, extraction, canonical mapping, and anomaly logging. Rather than treating LLMs as free-form generators, FinReporting employs them as constrained verifiers operating under explicit decision rules with evidence grounding. Evaluated on annual filings from the USA, Japan, and China, FinReporting improves consistency and reliability under heterogeneous reporting regimes. We further release an interactive demo that enables cross-market inspection and supports structured export of localized financial statements. Our demo is available at url{https://huggingface.co/spaces/BoomQ/FinReporting-Demo. A video describing our system is available at https://www.youtube.com/watch?v=f65jdEL31Kk.
URLs: https://huggingface.co/spaces/BoomQ/FinReporting-Demo., https://www.youtube.com/watch?v=f65jdEL31Kk.
Authors: Alireza Dadgarnia, Soroush Tabesh, Mahdi Nikdan, Michael Helcig, Eldar Kurtic, Maximilian Kleinegger, Dan Alistarh
Abstract: Quantization has become a standard tool for efficient LLM deployment, especially for local inference, where models are now routinely served at 2-3 bits per parameter. The state of the art is currently split into simple scalar quantization techniques, such as GPTQ or AWQ, which are widely deployed but plateau in accuracy at 3-4 bits per parameter (bpp), and "second-generation" vector- or trellis-quantized methods, such as QTIP, GPTVQ and AQLM, which push the accuracy frontier but are notoriously hard to implement and to scale. In this paper, we ask whether this gap is fundamental, or whether a carefully optimized $\textit{scalar}$ quantizer can recover most of it. We answer in the affirmative, by introducing GSQ (Gumbel-Softmax Quantization), a post-training scalar quantization method which jointly learns the per-coordinate grid assignments and the per-group scales using a Gumbel-Softmax relaxation of the discrete grid. GSQ matches the cardinality of the relaxation to the small number of levels available in the target bit-width regime (e.g., 3-8 levels for ternary and 3 bpp, respectively), making optimization tractable. Practically, on the standard Llama-3.1-8B/70B-Instruct models, GSQ closes most of the gap between scalar quantization and the QTIP frontier at 2 and 3 bits, while using a symmetric scalar grid with group-wise quantization, and thus remains compatible with existing scalar inference kernels. We further show that the same discrete-assignment optimization can be applied to practical GGUF K-Quant checkpoints: starting from publicly released GGUF models, GSQ improves accuracy while projecting the result back into the same deployment format. Finally, GSQ scales to trillion-scale Mixture-of-Experts models such as Kimi-K2.5, where vector-quantized methods are difficult to apply. The source code is publicly available at https://github.com/IST-DASLab/GSQ.
Authors: Jincheng Ren, Siwei Wu, Yizhi Li, Kang Zhu, Shu Xu, Boyu Feng, Ruibin Yuan, Wei Zhang, Riza Batista-Navarro, Jian Yang, Chenghua Lin
Abstract: As terminal agents scale to long-horizon, multi-turn workflows, a key bottleneck is not merely limited context length, but the accumulation of noisy terminal observations in the interaction history. Retaining raw observations preserves useful environment feedback, but also leads to context saturation and high token cost; conversely, naive compression may discard task-critical signals needed for subsequent actions. Because terminal environments are highly heterogeneous across repositories, commands, and execution states, heuristic-based or fixed-prompt compression methods are difficult to generalize. We propose TACO, a plug-and-play, training-free, self-evolving Terminal Agent Compression framework for existing terminal agents. TACO automatically discovers, refines, and reuses structured compression rules from interaction trajectories, enabling workflow-adaptive filtering of low-value terminal outputs while preserving task-relevant observations. Experiments on TerminalBench (TB 1.0 and TB 2.0) and four additional terminal-related benchmarks, including SWE-Bench Lite, CompileBench, DevEval, and CRUST-Bench, show that TACO consistently improves task performance and token efficiency across agent scaffolds and backbone models. On TerminalBench, TACO yields 1%-4% accuracy gains across strong agentic models and improves accuracy by around 2%-3% under the same token budget. On additional terminal-related benchmarks, it reduces total token consumption while maintaining or improving task success rates. These results suggest that self-evolving, workflow-adaptive observation compression is an effective path toward more reliable and efficient long-horizon terminal agents. The code is publicly available at https://github.com/multimodal-art-projection/TACO.
Authors: Mihailo \v{S}kori\'c, Cosimo Palma
Abstract: This paper presents a pipeline designed to transform raw Wikimedia dumps into quality textual corpora for seven South Slavic languages. The work is divided into two major phases. The first involves extracting and cleaning text from raw dumps of Wikipedia, Wikisource, Wikibooks, Wikinews, and Wikiquote. This step requires careful handling of raw wiki markup to isolate, first of all, textual articles, and then usable natural language text within them. The second phase addresses the challenge of questionable or low-quality articles, which are often generated from databases or structured knowledge bases. These articles are characterised by repetitive patterns, generic phrasing, and minimal to no original content. To mitigate their impact, a n-gram-based filtering strategy was employed to detect high levels of textual redundancy between articles and then remove such articles from the corpora entirely. The resulting datasets aim to provide linguistically rich texts suitable for training language models or conducting comparative research across South Slavic languages. By combining systematic extraction with quality control, this work contributes to the creation of reliable, high-information corpora that reflect the authentic cultural contexts of languages. While focused on the South Slavic case in the paper, the approach is mostly language-agnostic and can be generalised to other languages.
Authors: Guoshenghui Zhao, Tan Yu, Weijie Zhao
Abstract: Diffusion large language models generate text through multi-step denoising, where hallucination signals may emerge throughout the trajectory rather than only in the final output. Existing detectors mainly rely on output uncertainty or coarse trace statistics, which often fail to capture the richer hidden dynamics of D-LLMs. We propose HIVE, a hidden-evidence verification framework that extracts compressed hidden evidence from denoising trajectories, selects informative step-layer evidence, and conditions a verifier language model on the selected evidence through prefix embeddings. HIVE produces both a continuous hallucination score from verifier decision logits and structured verification outputs, including hallucination types, evidence pairs, and short rationales. Across two D-LLMs and three QA benchmarks, HIVE consistently outperforms eight strong baselines and achieves up to 0.9236 AUROC and 0.9537 AUPRC. Ablation studies further confirm the importance of hidden-evidence conditioning, learned evidence selection, two-stream evidence representation, and step-layer embeddings. These results suggest that selected hidden evidence from denoising trajectories provides a stronger and more usable hallucination signal than output-only uncertainty or coarse trace statistics.
Authors: Jasper Dekoninck, Nikola Jovanovi\'c, Tim Gehrunger, K\'ari R\"ognvaldsson, Ivo Petrov, Chenhao Sun, Martin Vechev
Abstract: Large language models (LLMs) are becoming increasingly capable mathematical collaborators, but static benchmarks are no longer sufficient for evaluating progress: they are often narrow in scope, quickly saturated, and rarely updated. This makes it hard to compare models reliably and track progress over time. Instead, we need evaluation platforms: continuously maintained systems that run, aggregate, and analyze evaluations across many benchmarks to give a comprehensive picture of model performance within a broad domain. In this work, we build on the original MathArena benchmark by substantially broadening its scope from final-answer olympiad problems to a continuously maintained evaluation platform for mathematical reasoning with LLMs. MathArena now covers a much wider range of tasks, including proof-based competitions, research-level arXiv problems, and formal proof generation in Lean. Additionally, we maintain a clear evaluation protocol for all models and regularly design new benchmarks as model capabilities improve to ensure that MathArena remains challenging. Notably, the strongest model, GPT-5.5, now reaches 98% on the 2026 USA Math Olympiad and 74% on research-level questions, showing that frontier models can now comfortably solve extremely challenging mathematical problems. This highlights the importance of continuously maintained evaluation platforms like MathArena to track the rapid progress of LLMs in mathematical reasoning.
Authors: Junnan Liu, Linhao Luo, Thuy-Trang Vu, Gholamreza Haffari
Abstract: Recent advances in large language models (LLMs) have demonstrated remarkable reasoning capabilities, largely stimulated by Reinforcement Learning with Verifiable Rewards (RLVR). However, existing RL algorithms face a fundamental limitation: their exploration remains largely constrained by the inherent capability boundary of the policy model. Although recent methods introduce external expert demonstrations to extend this boundary, they typically rely on complete trajectory-level guidance, which is sample-inefficient, information-sparse, and may confine exploration to a static guidance space. Inspired by the potential of multi-agent systems, we propose $\textbf{AIPO}$, an enhanced reinforcement learning framework that improves LLM reasoning through active multi-agent interaction during exploration. Specifically, AIPO enables the policy model to proactively consult three functional collaborative agents, $\textit{Verify Agent}$, $\textit{Knowledge Agent}$, and $\textit{Reasoning Agent}$, when encountering reasoning bottlenecks, thereby receiving fine-grained and targeted guidance to actively expand its capability boundary during training. We further introduce a tailored importance sampling coefficient together with a clipping strategy to mitigate the off-policy bias and gradient vanishing issues that arise when learning from agent-provided feedback. After training, the policy model performs reasoning independently without relying on collaborative agents. Extensive experiments on diverse reasoning benchmarks, including AIME, MATH500, GPQA-Diamond, and LiveCodeBench, show that AIPO consistently improves reasoning performance, generalizes robustly across different policy models and RLVR algorithms, and effectively expands the reasoning capability boundary of the policy model.
Authors: Yuzhuang Xu, Xu Han, Yuxuan Li, Pengzhan Li, Wanxiang Che
Abstract: Large language models (LLMs) achieve strong performance but incur high deployment costs, motivating extremely low-bit but lossy quantization. Existing quantization algorithms mainly focus on improving the numerical accuracy of forward computation to eliminate performance degradation. In this paper, we show that extremely quantized LLMs suffer from systematic smoothness degradation beyond numerical precision loss. Through a smoothness proxy, we observe that such degradation becomes increasingly severe as the quantization bit-width decreases. Furthermore, based on sequence neighborhood modeling, we find that quantized models exhibit a rapid reduction of effective token candidates within the prediction neighborhood, which directly leads to a sparser decoding tree and degraded generation quality. To validate it, we introduce a simple smoothness-preserving principle in both post-training quantization and quantization-aware training, and demonstrate that preserving smoothness brings additional gains beyond numerical accuracy. The core goal of this paper is to highlight smoothness preservation as an important design consideration for future extreme quantization methods. Code is available at https://github.com/xuyuzhuang11/FINE.
Authors: Xinyu Zhang, Zhicheng Dou, Deyang Li, Jianjun Tao, Shuo Cheng, Ruifeng Shi, Fangchao Liu, Enrui Hu, Yangkai Ding, Hongbo Wang, Qi Ye, Xuefeng Jin, Zhangchun Zhao
Abstract: As artificial intelligence engineering paradigms shift from single-agent Prompt and Context Engineering toward multi-agent \textbf{Coordination Engineering}, the ability to codify and systematically improve how multiple agents collaborate has emerged as a critical bottleneck. While single-agent skills can now be distributed as portable assets, multi-agent coordination protocols remain locked within framework-internal code or static configurations, preventing them from being shared across systems or autonomously improved over time. We propose \textbf{Swarm Skills}, a portable specification that extends the Anthropic Skills standard with multi-agent semantics. Swarm Skills turns multi-agent workflows into first-class, distributable assets that consist of roles, workflows, execution bounds, and a built-in semantic structure for self-evolution. To operationalize the specification's evolving nature, we present a companion self-evolution algorithm that automatically distills successful execution trajectories into new Swarm Skills and continuously patches existing ones based on multi-dimensional scoring (Effectiveness, Utilization, and Freshness), eliminating the need for human-in-the-loop oversight during the refinement process. Through an architectural compatibility analysis and a comprehensive qualitative case study using the open-source JiuwenSwarm reference implementation, we demonstrate how Swarm Skills achieves zero-adapter cross-agent portability via progressive disclosure, enabling agent teams to self-evolve their coordination strategies without framework lock-in.
Authors: Reza Khanmohammadi, Erfan Miahi, Simerjot Kaur, Charese H. Smiley, Ivan Brugere, Kundan Thind, Mohammad M. Ghassemi
Abstract: Large vision-language models suffer from visual ungroundedness: they can produce a fluent, confident, and even correct response driven entirely by language priors, with the image contributing nothing to the prediction. Existing confidence estimation methods cannot detect this, as they observe model behavior under normal inference with no mechanism to determine whether a prediction was shaped by the image or by text alone. We introduce BICR (Blind-Image Contrastive Ranking), a model-agnostic confidence estimation framework that makes this contrast explicit during training by extracting hidden states from a frozen LVLM twice: once with the real image-question pair, and once with the image blacked out while the question is held fixed. A lightweight probe is trained on the real-image hidden state and regularized by a ranking loss that penalizes higher confidence on the blacked-out view, teaching it to treat visual grounding as a signal of reliability at zero additional inference cost. Evaluated across five modern LVLMs and seven baselines on a benchmark covering visual question answering, object hallucination detection, medical imaging, and financial document understanding, BICR achieves the best cross-LVLM average on both calibration and discrimination simultaneously, with statistically significant discrimination gains robust to cluster-aware analysis at 4-18x fewer parameters than the strongest probing baseline.
Authors: Xubo Lin, Zezhi Deng, Shihao Wang, Grace Hui Yang, Yang Deng
Abstract: Most existing dialogue systems are user-driven, primarily designed to fulfill user requests. However, in many critical real-world scenarios, a conversational agent must proactively extract information to achieve its own objectives rather than merely respond. To address this gap, we introduce Inquisitive Conversational Agents (ICAs) and develop an ICA specifically tailored to U.S. Supreme Court oral arguments. We propose a Dual Hierarchical Reinforcement Learning framework featuring two cooperating RL agents, each with its own policy, to coordinate strategic dialogue management and fine-grained utterance generation. By learning when and how to ask probing questions, the agent emulates judicial questioning patterns and systematically uncovers crucial information to fulfill its legal objectives. Evaluations on a U.S. Supreme Court dataset show that our method outperforms various baselines across multiple metrics. It represents an important first step toward broader high-stakes, domain-specific applications.
Authors: Mokshit Surana, Archit Rathod, Akshaj Satishkumar
Abstract: Large Language Models (LLMs) trained on web-scale corpora inherently absorb toxic patterns from their training data. This leads to toxic degeneration where even innocuous prompts can trigger harmful outputs. This phenomenon poses significant risks for real-world deployments. Thus, necessitating effective mitigation strategies that should maintain model utility while ensuring safety. In this comprehensive replication study, we evaluate the efficacy of DExperts (Decoding-time Experts), which is an inference-time mitigation technique that steers generation without requiring model retraining. We structured our research into three systematic phases: (1) establishing baseline toxicity measurements using RealToxicityPrompts on standard GPT-2 models; then (2) implementing and evaluating DExperts to mitigate explicit toxicity; and finally (3) stress-testing the method against implicit hate speech using the adversarial ToxiGen dataset. Our empirical results confirm that while DExperts achieves near-perfect safety rates (100%) on explicit toxicity benchmarks, it exhibits brittleness against adversarial, implicit hate speech, with safety rates dropping to 98.5%. Furthermore, we quantify a critical trade-off. The method introduces a 10x latency penalty (from 0.2s to 2.0s per generation), posing challenges for real-time deployment scenarios. This study contributes to the growing body of work on AI safety by highlighting the robustness gap between explicit and implicit toxicity mitigation. We emphasize the need for more sophisticated approaches that generalize across diverse hate-speech patterns without incurring prohibitive computational costs.
Authors: Sinclair Schneider, Florian Steuber, Gabi Dreo Rodosek
Abstract: We present a new computational framework for detecting and structuring manipulative political narratives. A task that became more important due to the shift of political discussions to social media. One of the primary challenges thereby is differentiating between manipulative political narratives and legitimate critiques. Some posts may also reframe actual events within a manipulative context. To achieve good clustering results, we filter manipulative posts beforehand using a detailed few-shot prompt that combines documented campaign narratives with legitimate criticisms to differentiate them. This prompt enables a reasoning model to assign labels, retaining only manipulative narrative posts for further processing. The remaining posts are subsequently embedded and dimensionality-reduced using UMAP, before HDBSCAN is applied to uncover narrative groups. A key advantage of this unsupervised approach is its independence from a predefined list of target categories, enabling it to uncover new narrative clusters. Finally, a reasoning model is employed to uncover the narrative behind each cluster. This approach, applied to over 1.2 million social media posts, effectively identified 41 distinct manipulative narrative clusters by integrating prompt-based filtering with unsupervised clustering.
Authors: Xiang Shen, Yuhang Zhou, Yifan Wu, Zhuokai Zhao, Siyu Lin, Lei Huang, Qianqian Zhong, Lizhu Zhang, Benyu Zhang, Xiangjun Fan, Hong Yan
Abstract: Memory-augmented LLM agents have advanced personalized recommendation, yet existing approaches universally adopt flat memory representations that conflate ephemeral signals with stable preferences, and none provides a complete lifecycle governing how memory should evolve. We propose MARS (Memory-Augmented Agentic Recommender System), a framework that treats recommendation as a partially observable problem and maintains a structured belief state that progressively abstracts noisy behavioral observations into a compact estimate of user preferences. MARS organizes this belief state into three tiers: event memory buffers raw signals, preference memory maintains fine-grained mutable chunks with explicit strength and evidence tracking, and profile memory distills all preferences into a coherent natural language narrative. A complete lifecycle of six operations -- extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis -- is adaptively scheduled by an LLM-based planner rather than fixed-interval heuristics. Experiments on four InstructRec benchmark domains show that MARS achieves state-of-the-art performance with average improvements of 26.4% in HR@1 and 10.3% in NDCG@10 over the strongest baselines with further gains from agentic scheduling in evolving settings.
Authors: Yihang Chen, Pin Qian, Su Wang, Sipeng Zhang, Huan Xu, Shuhuai Lin, Xinpeng Wei
Abstract: The Context-Compliance Regime in Retrieval-Augmented Generation (RAG) occurs when retrieved context dominates the final answer even when it conflicts with the model's parametric knowledge. Accuracy alone does not reveal how retrieved context causally shapes answers under such conflict. We introduce Context-Driven Decomposition (CDD), a belief-decomposition probe that operates at inference time and serves as an intervention mechanism for controlled retrieval conflict. Across Epi-Scale stress tests, TruthfulQA misconception injection, and cross-model reruns, CDD exposes three patterns. P1: context compliance is measurable in an upper-bound adversarial setting, where Standard RAG reaches 15.0% accuracy on TruthfulQA misconception injection (N=500). P2: adversarial accuracy gains transfer across model families -- CDD improves accuracy on Gemini-2.5-Flash and on Claude Haiku/Sonnet/Opus -- but rationale-answer causal coupling does not transfer. CDD reaches 64.1% mistake-injection causal sensitivity on Gemini-2.5-Flash, while sensitivities for all three Claude variants fall in the [-3%, +7%] range, suggesting that the Claude-side accuracy gains operate through a mechanism distinct from the explicit conflict-resolution trace. P3: explicit conflict decomposition improves robustness under temporal drift and noisy distractors, with CDD reaching 71.3% on temporal shifts and 69.9% on distractor evidence on the full Epi-Scale adversarial benchmark. These three patterns identify context-compliance as a structural axis along which standard RAG can be probed and intervened on, distinct from retrieval-quality or single-method robustness questions, and motivate releasing Epi-Scale for systematic study across model families and retrieval pipelines.
Authors: ZiYi Dong, Yuliang Huang, Weijian Deng, Xiangyang Ji, Liang Lin, Pengxu Wei
Abstract: This work reformulates language generation as a stochastic optimal control problem, providing a unified theoretical perspective to analyze autoregressive and diffusion models and explain their limitations (Efficiency-Fidelity Paradox, Irreversibility Error Propagation, Optimization Tractability and Fidelity) in terms of combination of trajectory singularity, adjoint state vanishing, and gradient absence. To address these issues, we approximate the solution to the Hamilton-Jacobi-Bellman (HJB) equation, yielding an optimal policy that acts as a closed-loop controller. To bypass the intractability of directly solving the HJB PDE, we employ Flow Matching as the optimal trajectory solver within the rectified latent control space. This allows our Manta-LM with Global Integral Operator to approximate the global vector field, effectively realizing a model that simultaneously achieves high-fidelity text generation and efficient, low-cost parallel sampling. Empirically, our method achieves strong performance on language modeling and conditional generation tasks, while exhibiting improved stability, efficiency, and controllability.
Authors: Jie Jiang, Xing Sun, Ruotian Chen, Jianan Su, Kaixin Shen
Abstract: Speculative decoding accelerates LLM inference by having a lightweight draft model propose speculative windows of candidate tokens for parallel verification by a larger target model. In practice, speculative efficiency is often bottlenecked by hard-to-draft positions, where an early mismatch truncates the accepted prefix and invalidates the rest of the speculative window. Most learning-based drafters are still optimized with token-level supervised objectives, even though speculative utility is inherently window-level and prefix-sensitive. We propose PPOW (Performance-Driven Policy Optimization with Adaptive Windowing), a reinforcement learning framework that shifts drafter optimization from token-level imitation to window-level optimization. PPOW combines a Cost-Aware Speedup Reward, a Distribution-Based Proximity Reward, and Adaptive Divergence-Aware Windowing, which prioritizes informative windows with high confidence-weighted draft-target divergence. PPOW achieves average acceptance lengths of 6.29-6.52 and speedups of 3.39-4.36$\times$ across multiple model families and benchmarks under a unified decoding protocol. These results show that performance-driven window-level optimization is a practical approach to improving speculative decoding efficiency.
Authors: Alex L. Zhang, Thomas L. Griffiths, Karthik R. Narasimhan, Ofir Press
Abstract: Vision-language models (VLMs) have achieved strong results on coding and math benchmarks that are challenging for humans, yet their ability to perform tasks that come naturally to humans--such as perception, spatial navigation, and memory management--remains understudied. Real video games are crafted to be intuitive for humans to learn and master by leveraging innate inductive biases, making them an ideal testbed for evaluating such capabilities in VLMs. To this end, we introduce VideoGameBench, a benchmark consisting of 10 popular video games from the 1990s that VLMs directly interact with in real-time. VideoGameBench challenges models to complete entire games with access to only raw visual inputs and a high-level description of objectives and controls, a significant departure from existing setups that rely on game-specific scaffolding and auxiliary information. We keep three of the games secret to encourage solutions that generalize to unseen environments. Our experiments show that frontier vision-language models struggle to progress beyond the beginning of each game. We find inference latency to be a major limitation of frontier models in the real-time setting; therefore, we introduce VideoGameBench Lite, a setting where the game pauses while waiting for the LM's next action. The best performing models, Gemini 2.5 Pro and Claude 3.7 Sonnet, complete only 0.48% of VideoGameBench and 1.6% of VideoGameBench Lite. We hope that the formalization of the human skills mentioned above into this benchmark motivates progress in these research directions.
Authors: Zeyu Huang, Tianhao Cheng, Zihan Qiu, Zili Wang, Yinghui Xu, Edoardo M. Ponti, Ivan Titov
Abstract: Existing LLMs-post-training techniques are broadly categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). Each paradigm presents a distinct trade-off: (1) SFT excels at mimicking demonstration data, but can lead to problematic generalization as a form of behavior cloning. (2) Conversely, RFT can significantly enhance a model's performance but is prone to learning unexpected behaviors, and its performance is sensitive to the initial policy. In this paper, we propose a unified view of these methods and introduce Prefix-RFT, a hybrid approach that synergizes learning from both demonstration and exploration. Using mathematical reasoning problems as a test bed, we empirically demonstrate that Prefix-RFT is simple yet effective. Not only does it surpass the performance of standalone SFT and RFT, but it also outperforms parallel mixed-policy RFT methods. Our analysis highlights the complementary nature of SFT and RFT, validating that Prefix-RFT effectively harmonizes them. Further ablation studies confirm the method's robustness to variations in the quality and quantity of demonstration data.
Authors: Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, Jacob Andreas
Abstract: When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models. Code, models, and further info is available at https://rl-calibration.github.io/.
Authors: Jessica M. Lundin, Usman Nasir Nakakana, Guillaume Chabot-Couture
Abstract: Rigorous evaluation of domain-specific language models requires benchmarks that are comprehensive, contamination-resistant, and maintainable. Static, manually curated datasets do not satisfy these properties. We present a graph-based evaluation harness that transforms structured clinical guidelines into a queryable knowledge graph and dynamically instantiates evaluation queries via graph traversal. The framework provides three guarantees: (1) complete coverage of guideline relationships; (2) surface-form contamination resistance through combinatorial variation; and (3) validity inherited from expert-authored graph structure. Applied to the WHO IMCI guidelines, the harness generates clinically grounded multiple-choice questions spanning symptom recognition, treatment, severity classification, and follow-up care. Evaluation across five language models reveals systematic capability gaps. Models perform well on symptom recognition but show lower accuracy on treatment protocols and clinical management decisions. The framework supports continuous regeneration of evaluation data as guidelines evolve and generalizes to domains with structured decision logic. This provides a scalable foundation for evaluation infrastructure.
Authors: Parth Asawa, Alan Zhu, Abigail O'Neill, Matei Zaharia, Alexandros G. Dimakis, Joseph E. Gonzalez
Abstract: Frontier language models are deployed as black-box services, where model weights cannot be modified and customization is limited to prompting. We introduce Advisor Models, a method to train small open-weight models to generate dynamic, per-instance natural language advice that improves the capabilities of black-box frontier models. Advisor Models improve GPT-5.2's performance on RuleArena (Taxes) by 27.4%, reduce Gemini 3 Pro's steps taken in SWE agent tasks by 24.6%, and outperform static prompt optimizers in personalizing GPT-5 to user preferences (85-100% vs. 40-60%). We also find that advisors are transferable: an advisor trained with a low-cost student model still transfers improvements to a frontier model. Moreover, Advisor Models are robust: we observe no degradation on other benchmarks than the pipeline is trained on. Our method shows how to perform parametric optimization for black-box frontier models in a practical and cost-effective way.
Authors: Tien Dang, The-Hai Nguyen, Dinh Mai Phuong, Nguyen Minh Phuong, Anh Bui, Hoang Thanh-Tung, Le-Minh Nguyen, Naoya Inoue
Abstract: We consider Representation Misdirection (RM), a class of large language model (LLM) unlearning methods that achieve forgetting by redirecting the forget-representations, that is, latent representations of forget-samples, toward a target vector. Despite being important, the roles of the target vector used in RM, however, remain underexplored. Here, we approach and revisit RM through the lens of the Linear Representation Hypothesis. Specifically, if one can identify a one-dimensional representation corresponding to a high-level concept, the Linear Representation Hypothesis enables linear operations on this concept vector within the forget-representation space. Under this view, we hypothesize that, beyond forgetting, machine unlearning via RM elicits controllable emergent side behaviors and stronger side capabilities corresponding to the high-level concept. Our hypothesis is empirically validated across a wide range of tasks, including behavioral control (e.g., controlling unlearned models' truthfulness, sentiment, refusal, and language) and capability enhancement (e.g., improving unlearned models' in-context learning (ICL) capability). Our findings reveal that this phenomenon could be either a hidden risk if misused or a mechanism that can be harnessed for developing unlearned models that require stronger capabilities and controllable behaviors.
Authors: Yixuan Even Xu, John Kirchenbauer, Yash Savani, Asher Trockman, Alexander Robey, Tom Goldstein, Fei Fang, J. Zico Kolter
Abstract: Model distillation enables efficient emulation of frontier large language models (LLMs), creating a need for robust mechanisms to detect when a third-party student model has trained on a teacher model's outputs. However, existing fingerprinting techniques that could be used to detect such distillation rely on heuristic perturbations that impose a steep trade-off between generation quality and fingerprinting strength, often requiring significant degradation of utility to ensure the fingerprint is effectively internalized by the student. We introduce antidistillation fingerprinting (ADFP), a principled approach that aligns the fingerprinting objective with the student's learning dynamics. Building upon the gradient-based framework of antidistillation sampling, ADFP utilizes a proxy model to identify and sample tokens that directly maximize the expected detectability of the fingerprint in the student after fine-tuning, rather than relying on the incidental absorption of the un-targeted biases of a more naive watermark. Experiments on GSM8K, OASST1, and MBPP demonstrate that ADFP achieves a significant Pareto improvement over state-of-the-art baselines, yielding stronger detection confidence with minimal impact on utility across mathematical reasoning, dialogue, and code generation, even when the student model's architecture is unknown.
Authors: Atharva Sehgal, James Hou, Akanksha Sarkar, Ishaan Mantripragada, Swarat Chaudhuri, Jennifer J. Sun, Yisong Yue
Abstract: Large language model (LLM) coding agents increasingly operate at the repository level, motivating benchmarks that evaluate their ability to optimize entire codebases under realistic constraints. Existing code benchmarks largely rely on synthetic tasks, binary correctness signals, or single-objective evaluation, limiting their ability to assess holistic optimization behavior. We introduce FormulaCode, a benchmark for evaluating agentic optimization on large, real-world codebases with fine-grained, multi-objective performance metrics. FormulaCode comprises 957 performance bottlenecks mined from scientific Python repositories on GitHub, each paired with expert-authored patches and, on average, 264.6 community-maintained performance workloads per task, enabling the holistic ability of LLM agents to optimize codebases under realistic correctness and performance constraints. Our evaluations reveal that repository-scale, multi-objective optimization remains a major challenge for frontier LLM agents. Project website at: https://formula-code.github.io
Authors: Pegah Ramezani, Thomas Kinfe, Andreas Maier, Achim Schilling, Patrick Krauss
Abstract: Understanding how the brain processes linguistic constructions is a central challenge in cognitive neuroscience and linguistics. Recent computational studies show that artificial neural language models spontaneously develop differentiated representations of Argument Structure Constructions (ASCs), generating predictions about when and how construction-level information emerges during processing. The present study tests these predictions in human neural activity using electroencephalography (EEG). Ten native English speakers listened to 200 synthetically generated sentences across four construction types (transitive, ditransitive, caused-motion, resultative) while neural responses were recorded. Analyses using time-frequency methods, feature extraction, and machine learning classification revealed construction-specific neural signatures emerging primarily at sentence-final positions, where argument structure becomes fully disambiguated, and most prominently in the alpha band. Pairwise classification showed reliable differentiation, especially between ditransitive and resultative constructions, while other pairs overlapped. Crucially, the temporal emergence and similarity structure of these effects mirror patterns in recurrent and transformer-based language models, where constructional representations arise during integrative processing stages. These findings support the view that linguistic constructions are neurally encoded as distinct form-meaning mappings, in line with Construction Grammar, and suggest convergence between biological and artificial systems on similar representational solutions. More broadly, this convergence is consistent with the idea that learning systems discover stable regions within an underlying representational landscape - recently termed a Platonic representational space - that constrains the emergence of efficient linguistic abstractions.
Authors: Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev, Vyacheslav Zhdanovskiy, Yegor Yershov
Abstract: With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input prompt. We create and release the Text2JSON benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.
Authors: Yiqun Sun, Pengfei Wei, Lawrence B. Hsieh
Abstract: Retrieval-Augmented Generation (RAG) grounds LLM responses in external evidence but treats the model as a passive consumer of search results, with no view of how the corpus is organized or what it has not yet seen. We present Corpus2Skill, which distills a document corpus offline into a hierarchical skill directory and lets an LLM agent navigate it at serve time, drilling from a bird's-eye view through progressively finer summaries down to documents, and backtracking when a branch is unproductive. On an enterprise customer-support benchmark, Corpus2Skill improves both answer quality and grounding over single-shot dense, hybrid, hierarchical-retrieval, and agentic RAG baselines at a moderate cost tradeoff. A ten-subset generalization study further shows that corpus navigation is not a universal replacement for retrieval: it consistently helps on single-domain corpora with a recoverable topical taxonomy, but flat retrieval remains preferable on open-domain factoid pools or homogeneous-tabular corpora that defeat top-level clustering. We characterize this scope distinction and discuss it as a design guideline for knowledge-grounded systems. Code is available at https://github.com/dukesun99/Corpus2Skill.
Authors: Gabriel Garcia
Abstract: Large language models often fail at simple counting tasks, even when items to count are in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert representations to the correct output tokens. Across three model families: Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find evidence for the second explanation. Linear probes recover the correct count from intermediate layers with $R^2>0.99$, showing that the information is present. However, the internal directions that encode counts are nearly orthogonal to digit-token output-head rows ($|\cos| \leq 0.032$). In other words, the model stores the count in a form that the digit logits do not naturally read out. We localize this failure with two interventions. Updating only the digit rows of the output head (36,864 parameters) substantially improves constrained digit prediction (60.7--100.0% on four tasks), but it does not fix unconstrained generation (0%); we do not claim that digit-row repair fixes open-ended text. By contrast, small LoRA on attention Q/V (7.67M parameters) improves upstream routing and achieves 83.1%$\pm$7.2% in true greedy autoregressive generation (deployable fix). Logit-lens at layer 35 (entity counting; correct-digit rank): (i) median over 3 seeds drops from order-$10^4$ to 1; (ii) seed 42 shows $54{,}332 \to 838$ (median top-1 while one seed stays far below). Norm, logit-lens, and cross-task analyses generalize the bottleneck to counting, addition, and list length; nulls on MMLU and GSM8K and limited DROP transfer. These results identify counting failure as a geometric readout bottleneck, not an internal-representation failure: the model knows the count but the output pathway is misaligned with tokens needed to express it.
Authors: Daniel Goldstein, Eugene Cheah
Abstract: We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and linear RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. It can also be implemented in a hybrid solution in tandem with LRNN layers in place of traditional attention, to supplement the LRNN with improved sublinear memory growth context length usage and long context decoding. We release our code at https://github.com/featherless-ai/KVM-paper and trained models at https://huggingface.co/collections/featherless-ai/kvm-paper under the Apache 2.0 license.
URLs: https://github.com/featherless-ai/KVM-paper, https://huggingface.co/collections/featherless-ai/kvm-paper
Authors: Gabriel Garcia
Abstract: Corruption studies, the standard tool for evaluating chain-of-thought (CoT) faithfulness, infer which steps are ``computationally important'' from accuracy loss when steps are corrupted. We show that when benchmark chains end with an explicit terminal answer line, as in GSM8K and MATH, these tests largely measure \emph{answer placement} rather than where intermediate computation is carried out. Using matched GSM8K examples, removing only the final answer statement while preserving all reasoning collapses suffix sensitivity by about $19\times$ for Qwen~2.5-3B ($N{=}300$, $p{=}0.022$). Conflicting-answer prompts, which contain correct reasoning but a wrong explicit final answer, drive accuracy to zero or near-zero at 7B across five open-weight model families; wrong-answer following is strong at 3B--7B and attenuates sharply at larger scales. Replications on MATH, within-stable comparisons at 7B, and suffix-free chains show the same pattern in different guises: corruption sensitivity tracks the location of explicit answer text, not a fixed computational depth in the reasoning. Generation-time probes indicate that final answers are rarely early-determined during generation (${<}5\%$ early commitment), yet consumption-time behavior systematically follows explicit answer text. The confound is therefore largely a readout effect when the chain is consumed. We propose a three-prerequisite protocol (question-only control, format characterization, and an all-position sweep) as a practical minimum for future corruption-based faithfulness studies.
Authors: Seonghoon Yu, Dongjun Nam, Byung-Kwan Lee, Jeany Son
Abstract: Recent think-answer approaches in VLMs, such as Qwen3-VL-Thinking, boost reasoning performance by leveraging intermediate thinking steps before the final answer, but their computational cost becomes substantial, especially for larger VLMs. To distill such capabilities into compact think-answer VLMs, a primary objective is to improve the student's ability to utilize visual evidence throughout its reasoning trace, as long think-answer traces suffer from visual forgetting issues. To this end, we introduce a novel think-answer distillation framework that encourages the student to anchor its thinking on visual information by masking the student's salient reasoning prefixes. To compensate for such masked textual cues, the student is encouraged to rely more on visual evidence as an alternative source of information during distillation. Our masking strategies include: 1) token-wise salient reasoning-prefix masking, which masks high-influence reasoning prefixes selectively for each next-token prediction, and 2) self-paced masking budget scheduling, which gradually increases the masking scale according to distillation difficulty, measured by the discrepancy between teacher--student distributions. In the distillation phase, the student is guided by our salient reasoning-prefix mask, which blocks both future tokens and salient reasoning cues, in place of the standard causal mask used for auto-regressive language modeling. Experimental results show that our approach outperforms recent open-source VLMs, VLM distillation, and self-distillation methods on multimodal reasoning benchmarks, while further analyzes confirm enhanced visual utilization along the student thinking process.
Authors: Anastasiia Sedova, Skyler Seto, Natalie Schluter, Pierre Ablin
Abstract: As language models scale, the amount of data they require grows -- yet many target data sources, such as low-resource languages or specialized domains, are inherently limited in size. A common strategy is to mix this scarce but valuable target data with abundant generic data, which presents a fundamental trade-off: too little target data in the mixture underexposes the model to the target domain, while too much target data repeats the same examples excessively, yielding diminishing returns and eventual overfitting. We study this trade-off across more than 2,000 language-model training runs spanning multiple model and target dataset sizes, as well as several data types, including multilingual, domain-specific, and quality-filtered mixtures. Across all settings, we find that repetition is a central driver of target-domain performance, and that mixture training tolerates much higher repetition than single-source training: scarce target corpora can be reused 15-20 times, with the optimal number of repetitions depending on the target data size, compute budget, and model scale. Next, we introduce a repetition-aware mixture scaling law that accounts for the decreasing value of repeated target tokens and the regularizing role of generic data. Optimizing the scaling law provides a principled way to compute effective mixture configurations, yielding practical mixture recommendations for pretraining under data constraints.
Authors: Jerem\'ias Figueiredo Paschmann, Juan Kaplan, Francisco Nattero, Santiago Barron, Juan Wisznia, Luciano del Corro
Abstract: Pairwise Ranking Prompting (PRP) elicits pairwise preference judgments from an LLM, which are then aggregated into a ranking, usually via classical sorting algorithms. However, judgments are noisy, order-sensitive, and sometimes intransitive, so sorting assumptions do not match the setting. Because sorting aims to recover a full permutation, truncating it to meet a call budget does not produce a dependable top-K. We thus reframe PRP reranking as active learning from noisy pairwise comparisons and show that active rankers are drop-in replacements that improve NDCG@10 per call in the call-constrained regime. Our noise-robust framework also introduces a randomized-direction oracle that uses a single LLM call per pair. This approach converts systematic position bias into zero-mean noise, enabling unbiased aggregate ranking without the cost of bidirectional calls.
Authors: Joy Bose
Abstract: Legal reasoning is not semantic similarity search. A court judgment encodes constrained symbolic reasoning: precedent propagation, procedural state transitions, and statute-bound inference. These are properties that vector-based retrieval-augmented generation (RAG) cannot faithfully represent. Hallucinated precedents, outdated statute citations, and unsupported reasoning chains remain persistent failure modes in LLM-based legal AI, with real consequences for access to justice in high-caseload jurisdictions such as India. This paper presents Falkor-IRAC, a graph-constrained generation framework for Indian legal AI that grounds generation in structured reasoning over an IRAC (Issue, Rule, Analysis, Conclusion) knowledge graph. Judgments from the Supreme Court and High Courts of India are ingested as IRAC node structures enriched with procedural state transitions, precedent relationships, and statutory references, stored in FalkorDB for low-latency agentic traversal. At inference time, LLM-generated answers are accepted only if a valid supporting path can be traced through the graph, a check performed by a falsifiability oracle called the Verifier Agent. The system also detects doctrinal conflicts as a first-class output rather than silently resolving them. Falkor-IRAC is evaluated using graph-native metrics: citation grounding accuracy, path validity rate, hallucinated precedent rate, and conflict detection rate. These metrics are argued to be more appropriate for legal reasoning evaluation than BLEU and ROUGE. On a proof-of-concept corpus of 51 Supreme Court judgments, the Verifier Agent correctly validated citations on completed queries and correctly rejected fabricated citations. Evaluation against vector-only RAG baselines is left for future work. The companion InIRAC dataset, 500+ structured Indian court judgments with IRAC annotations, is released alongside this paper.