Computation and Language
compar:IA: The French Government's LLM arena to collect French-language human prompts and preference data
Lucie Termignon, Simonas Zilinskas, Hadrien Pélissier et al. (+3 more)
Large
Language Models (LLMs) often show reduced performance, cultural
alignment, and safety robustness in non-English
languages, partly because English dominates both pre-training data and human preference
alignment datasets. Training methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) require human preference data, which remains scarce and largely non-public for many
languages beyond English. To address this gap, we introduce compar:IA, an open-source digital public service developed inside the French government and designed to collect large-scale human preference data from a predominantly French-speaking general audience. The platform uses a blind pairwise comparison interface to capture unconstrained, real-world
prompts and user judgments across a diverse set of
language models, while maintaining low participation friction and
privacy-preserving automated filtering. As of 2026-02-07, compar:IA has collected over 600,000 free-form
prompts and 250,000 preference votes, with approximately 89% of the data in French. We release three complementary
datasets – conversations, votes, and reactions – under open licenses, and present initial analyses, including a French-
language model leaderboard and user interaction patterns. Beyond the French context, compar:IA is evolving toward an international digital public good, offering reusable infrastructure for
multilingual model training, evaluation, and the study of human-AI interaction.
Read the full abstract ▾
All authors Lucie Termignon, Simonas Zilinskas, Hadrien Pélissier, Aurélien Barrot, Nicolas Chesnais, Elie Gavoty
Computation and Language
Not All Layers Need Tuning: Selective Layer Restoration Recovers Diversity
Bowen Zhang, Meiyi Wang and Harold Soh
Post-training improves instruction-following and helpfulness of large
language models (LLMs) but often reduces generation diversity, which leads to repetitive outputs in open-ended settings, a phenomenon known as mode collapse. Motivated by evidence that LLM layers play distinct functional roles, we hypothesize that mode collapse can be localized to specific layers and that restoring a carefully chosen range of layers to their pre-trained weights can recover diversity while maintaining high output quality. To validate this hypothesis and decide which layers to restore, we design a proxy task – Constrained Random Character(CRC) – with an explicit validity set and a natural diversity objective. Results on CRC reveal a clear diversity-validity trade-off across restoration ranges and identify configurations that increase diversity with minimal quality loss. Based on these findings, we propose Selective Layer Restoration (SLR), a training-free method that restores selected layers in a post-trained model to their pre-trained weights, yielding a hybrid model with the same architecture and parameter count, incurring no additional inference cost. Across three different tasks (creative writing, open-ended question answering, and multi-step
reasoning) and three different model families (Llama, Qwen, and Gemma), we find SLR can consistently and substantially improve output diversity while maintaining high output quality.
Read the full abstract ▾
Computation and Language
Beyond Static Alignment: Hierarchical Policy Control for LLM Safety via Risk-Aware Chain-of-Thought
Jianfeng Si, Lin Sun, Weihong Lin et al. (+1 more)
Large
Language Models (LLMs) face a fundamental safety-helpfulness trade-off due to static, one-size-fits-all safety policies that lack runtime controllabilityxf, making it difficult to tailor responses to diverse application needs. %As a result, models may over-refuse benign requests or under-constrain harmful ones. We present PACT (
Prompt-configured Action via Chain-of-Thought), a framework for dynamic safety control through explicit, risk-aware
reasoning. PACT operates under a hierarchical policy architecture: a non-overridable global safety policy establishes immutable boundaries for critical risks (e.g., child safety, violent extremism), while user-defined policies can introduce domain-specific (non-global) risk categories and specify label-to-action behaviors to improve utility in real-world deployment settings. The framework decomposes safety decisions into structured Classify$\rightarrow$Act paths that route queries to the appropriate action (comply, guide, or reject) and render the decision-making process transparent. Extensive experiments demonstrate that PACT achieves near state-of-the-art safety performance under global policy evaluation while attaining the best controllability under user-specific policy evaluation, effectively mitigating the safety-helpfulness trade-off. We will release the PACT model suite, training data, and evaluation protocols to facilitate reproducible research in controllable safety
alignment.
Read the full abstract ▾
All authors Jianfeng Si, Lin Sun, Weihong Lin, Xiangzheng Zhang
Computation and Language
Reading Between the Waves: Robust Topic Segmentation Using Inter-Sentence Audio Features
Steffen Freisinger, Philipp Seeberger, Tobias Bocklet et al. (+1 more)
Spoken content, such as online
videos and podcasts, often spans multiple topics, which makes automatic topic
segmentation essential for user
navigation and downstream applications. However, current methods do not fully leverage acoustic features, leaving room for improvement. We propose a multi-modal approach that fine-tunes both a text encoder and a Siamese audio encoder, capturing acoustic cues around sentence boundaries. Experiments on a large-scale
dataset of YouTube
videos show substantial gains over text-only and multi-modal baselines. Our model also proves more resilient to ASR noise and outperforms a larger text-only baseline on three additional
datasets in Portuguese, German, and English, underscoring the value of learned acoustic features for robust topic
segmentation.
Read the full abstract ▾
All authors Steffen Freisinger, Philipp Seeberger, Tobias Bocklet, Korbinian Riedhammer
Computation and Language
FairJudge: An Adaptive, Debiased, and Consistent LLM-as-a-Judge
Bo Yang, Lanfei Feng, Yunkui Chen et al. (+3 more)
Existing LLM-as-a-Judge systems suffer from three fundamental limitations: limited adaptivity to task- and domain-specific evaluation criteria, systematic biases driven by non-semantic cues such as position, length, format, and model provenance, and evaluation inconsistency that leads to contradictory judgments across different evaluation modes (e.g., pointwise versus pairwise). To address these issues, we propose FairJudge, an adaptive, debiased, and consistent LLM-as-a-Judge. Unlike prior approaches that treat the judge as a static evaluator, FairJudge models judging behavior itself as a learnable and regularized policy. From a data-centric perspective, we construct a high-information-density judging
dataset that explicitly injects supervision signals aligned with evaluation behavior. Building on this
dataset, we adopt a curriculum-style SFT-DPO-GRPO training paradigm that progressively aligns rubric adherence, bias mitigation, and cross-mode consistency, while avoiding catastrophic forgetting. Experimental results on multiple internal and public
benchmarks show that FairJudge consistently improves agreement and F1, reduces non-semantic biases, and outperforms substantially larger instruction-tuned LLMs. All resources will be publicly released after acceptance to facilitate future research.
Read the full abstract ▾
All authors Bo Yang, Lanfei Feng, Yunkui Chen, Yu Zhang, Xiao Xu, Shijian Li
Computation and Language
Do Prompts Guarantee Safety? Mitigating Toxicity from LLM Generations through Subspace Intervention
Himanshu Singh, Ziwei Xu, A. V. Subramanyam et al. (+1 more)
Large
Language Models (LLMs) are powerful text generators, yet they can produce toxic or harmful content even when given seemingly harmless
prompts. This presents a serious safety challenge and can cause real-world harm. Toxicity is often subtle and context-dependent, making it difficult to detect at the
token level or through coarse sentence-level signals. Moreover, efforts to mitigate toxicity often face a trade-off between safety and the coherence, or fluency of the generated text. In this work, we present a targeted subspace intervention strategy for identifying and suppressing hidden toxic patterns from underlying model representations, while preserving overall ability to generate safe fluent content. On the RealToxicityPrompts, our method achieves strong mitigation performance compared to existing baselines, with minimal impact on inference complexity. Across multiple LLMs, our approach reduces toxicity of state-of-the-art detoxification systems by 8-20%, while maintaining comparable fluency. Through extensive quantitative and qualitative analyses, we show that our approach achieves effective toxicity reduction without impairing generative performance, consistently outperforming existing baselines.
Read the full abstract ▾
All authors Himanshu Singh, Ziwei Xu, A. V. Subramanyam, Mohan Kankanhalli
Computation and Language
Echoes as Anchors: Probabilistic Costs and Attention Refocusing in LLM Reasoning
Zhuoyuan Hao, Zhuo Li, Wu Li et al. (+3 more)
Recent work improves
reasoning in large
reasoning models (LRMs) by test-time scaling through self-consistency and parallel thinking, adding generic "thinking
tokens", or presenting the question twice, as in re-reading (RE2). However, these approaches either inject task-agnostic
tokens or mandate heuristics that do not explain, and often ignore, the spontaneous repetition that many LRMs exhibit at the head of their long chain-of-thought (CoT). In this work, we investigate the model's tendency to restate the question, which we dub Echo of
Prompt (EOP), as a mechanism that shapes computation from the start and can be viewed as
prompt repetition inside the
reasoning trace rather than in the input. We formalize its probabilistic cost by casting echo removal as rejection sampling, and define the Echo Likelihood Gap as a computable proxy linking early repetition to likelihood gains and downstream accuracy. We find that (1) in a causal intervention, reinserting an echo into failed GSM8K traces raises exact match by 7.9-10.4 points for two
reasoning models but not for a non-
reasoning base model; and (2) likelihood analyses controlling for length and suffix indicate that EOP's benefits go beyond verbosity, while layer-wise
attention studies show that EOP increases
attention from answer
tokens to the answer prefix in middle layers, consistent with an
attention refocusing mechanism. Building on these findings, we develop Echo-Distilled SFT (ED-SFT) to instill echoing before
reasoning via supervised fine-tuning, and training-free Echoic Prompting (EP) to ground the model on the
prompt midway through the trace. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and observe gains over baselines, including the insertion of thinking
tokens, in most settings. Our code is publicly released at https://github.com/hhh2210/echoes-as-anchors.
Read the full abstract ▾
All authors Zhuoyuan Hao, Zhuo Li, Wu Li, Fangming Liu, Min Zhang, Jing Li
Computation and Language · Cross-listed from cs.HC
Personality as Relational Infrastructure: User Perceptions of Personality-Trait-Infused LLM Messaging
Dominik P. Hofer, David Haag, Rania Islambouli et al. (+1 more)
Digital behaviour change systems increasingly rely on repeated, system-initiated messages to support users in everyday contexts. LLMs enable these messages to be personalised consistently across interactions, yet it remains unclear whether such personalisation improves individual messages or instead shapes users' perceptions through patterns of exposure. We explore this question in the context of LLM-generated JITAIs, which are short, context-aware messages delivered at moments deemed appropriate to support behaviour change, using physical activity as an application domain. In a controlled retrospective study, 90 participants evaluated messages generated using four LLM strategies: baseline prompting, few-shot prompting, fine-tuned models, and
retrieval augmented generation, each implemented with and without Big Five Personality Traits to produce personality-aligned communication across multiple scenarios. Using ordinal multilevel models with within-between decomposition, we distinguish trial-level effects, whether personality information improves evaluations of individual messages, from person-level exposure effects, whether participants receiving higher proportions of personality-informed messages exhibit systematically different overall perceptions. Results showed no trial-level associations, but participants who received higher proportions of BFPT-informed messages rated the messages as more personalised, appropriate, and reported less negative affect. We use Communication Accommodation Theory for post-hoc analysis. These results suggest that personality-based personalisation in behaviour change systems may operate primarily through aggregate exposure rather than per-message optimisation, with implications for how adaptive systems are designed and evaluated in sustained human-AI interaction. In-situ longitudinal studies are needed to validate these findings in real-world contexts.
Read the full abstract ▾
All authors Dominik P. Hofer, David Haag, Rania Islambouli, Jan D. Smeddinck
Computation and Language
Inference-Time Rethinking with Latent Thought Vectors for Math Reasoning
Deqian Kong, Minglu Zhao, Aoyang Qin et al. (+10 more)
Standard chain-of-thought
reasoning generates a solution in a single forward pass, committing irrevocably to each
token and lacking a mechanism to recover from early errors. We introduce Inference-Time Rethinking, a generative framework that enables iterative self-correction by decoupling declarative latent thought vectors from procedural generation. We factorize
reasoning into a continuous latent thought vector (what to reason about) and a decoder that verbalizes the trace conditioned on this vector (how to reason). Beyond serving as a declarative buffer, latent thought vectors compress the
reasoning structure into a continuous representation that abstracts away surface-level
token variability, making gradient-based optimization over
reasoning strategies well-posed. Our prior model maps unstructured noise to a learned
manifold of valid
reasoning patterns, and at test time we employ a Gibbs-style procedure that alternates between generating a candidate trace and optimizing the latent vector to better explain that trace, effectively navigating the latent
manifold to refine the
reasoning strategy. Training a 0.2B-parameter model from scratch on GSM8K, our method with 30 rethinking iterations surpasses baselines with 10 to 15 times more parameters, including a 3B counterpart. This result demonstrates that effective mathematical
reasoning can emerge from sophisticated inference-time computation rather than solely from massive parameter counts.
Read the full abstract ▾
All authors Deqian Kong, Minglu Zhao, Aoyang Qin, Bo Pang, Chenxin Tao, David Hartmann, Edouardo Honig, Dehong Xu, Amit Kumar, Matt Sarte, Chuan Li, Jianwen Xie, Ying Nian Wu
Computation and Language
Baichuan-M3: Modeling Clinical Inquiry for Reliable Medical Decision-Making
Baichuan-M3 Team: Chengfeng Dou, Fan Yang, Fei Li et al. (+14 more)
We introduce Baichuan-M3, a medical-enhanced large
language model engineered to shift the paradigm from passive question-answering to active, clinical-grade decision support. Addressing the limitations of existing systems in open-ended consultations, Baichuan-M3 utilizes a specialized training pipeline to model the systematic workflow of a physician. Key capabilities include: (i) proactive information acquisition to resolve ambiguity; (ii) long-horizon
reasoning that unifies scattered evidence into coherent diagnoses; and (iii) adaptive
hallucination suppression to ensure factual reliability. Empirical evaluations demonstrate that Baichuan-M3 achieves state-of-the-art results on HealthBench, the newly introduced HealthBench-Hallu and ScanBench, significantly outperforming GPT-5.2 in clinical inquiry, advisory and safety. The models are publicly available at https://huggingface.co/collections/baichuan-inc/baichuan-m3.
Read the full abstract ▾
All authors Baichuan-M3 Team: Chengfeng Dou, Fan Yang, Fei Li, Jiyuan Jia, Qiang Ju, Shuai Wang, Tianpeng Li, Xiangrong Zeng, Yijie Zhou, Hongda Zhang, Jinyang Tai, Linzhuang Sun, Peidong Guo, Yichuan Mo, Xiaochuan Wang, Hengfu Cui, Zhishou Zhang
Computation and Language · Cross-listed from cs.CV
SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs
Niccolo Avogaro, Nayanika Debnath, Li Mi et al. (+6 more)
Despite recent successes, test-time scaling – i.e., dynamically expanding the
token budget during inference as needed – remains brittle for
vision-
language models (VLMs). Unstructured visual
reasoning chains entangle perception and
reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers.
Reasoning also requires expensive reinforcement learning with hand-crafted rewards. Here, we introduce SPARC (Separating Perception And
Reasoning Circuits), a modular framework that explicitly decouples visual perception from
reasoning. Inspired by sequential sensory-to-cognitive processing in the
brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its
reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric compute allocation (e.g., prioritizing perceptual processing under distribution shift), and supports selective optimization (e.g., improving the perceptual stage alone when it is the bottleneck for end-to-end performance). It also accommodates compressed contexts by running global search at lower image resolutions and allocating high-resolution processing only to selected regions, thereby reducing visual
token count and compute. SPARC outperforms monolithic baselines and strong visual-grounding approaches across challenging visual
reasoning tasks, such as improving Qwen3VL 4B on the $V^*$ VQA
benchmark by 6.7 points and surpassing "thinking with images" by 4.6 points in an OOD setting with a $200\times$ lower
token budget.
Read the full abstract ▾
All authors Niccolo Avogaro, Nayanika Debnath, Li Mi, Thomas Frick, Junling Wang, Zexue He, Hang Hua, Konrad Schindler, Mattia Rigotti
Computation and Language · Cross-listed from cs.CR
"Do Not Mention This to the User": Detecting and Understanding Malicious Agent Skills in the Wild
Yi Liu, Zhihao Chen, Yanjun Zhang et al. (+4 more)
LLM-based coding
agents increasingly rely on third-party extensions called skills, which bundle natural
language instructions and helper scripts that execute with full user privileges. Community registries have emerged to distribute these skills, but the security implications remain unstudied due to the absence of labeled threat data. This paper presents a systematic security analysis of 98,380 skills collected from two major registries. Through a combination of static pattern matching and dynamic behavioral verification, we identify 157 skills exhibiting confirmed malicious behavior, encompassing 632 distinct vulnerabilities across 13 attack techniques. Our analysis reveals that these threats are deliberate rather than accidental: each malicious skill contains an average of 4.03 vulnerabilities spanning multiple attack phases. We identify two dominant attack strategies with statistically significant negative correlation – credential theft via remote code execution, and
agent manipulation through
adversarial instructions embedded in documentation. Over half of all confirmed cases originate from a single threat actor employing templated brand impersonation at scale. We further observe that attack sophistication correlates with concealment investment, with advanced skills universally employing undocumented capabilities while also exploiting platform-native trust mechanisms. Following responsible disclosure, registry maintainers removed all 157 (100%) of the reported skills. Our
dataset and detection pipeline are publicly available to facilitate future research on securing LLM
agent ecosystems.
Read the full abstract ▾
All authors Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning, Leo Yu Zhang
Computation and Language
MTQE.en-he: Machine Translation Quality Estimation for English-Hebrew
Andy Rosenbaum, Assaf Siani and Ilan Kernerman
We release MTQE.en-he: to our knowledge, the first publicly available English-Hebrew
benchmark for Machine Translation Quality Estimation. MTQE.en-he contains 959 English segments from WMT24++, each paired with a machine translation into Hebrew, and Direct Assessment scores of the translation quality annotated by three human experts. We
benchmark ChatGPT prompting, TransQuest, and CometKiwi and show that ensembling the three models outperforms the best single model (CometKiwi) by 6.4 percentage points Pearson and 5.6 percentage points Spearman. Fine-tuning experiments with TransQuest and CometKiwi reveal that full-model updates are sensitive to overfitting and distribution collapse, yet parameter-efficient methods (LoRA, BitFit, and FTHead, i.e., fine-tuning only the classification head) train stably and yield improvements of 2-3 percentage points. MTQE.en-he and our experimental results enable future research on this under-resourced
language pair.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research
Yishan Li, Wentong Chen, Yukun Yan et al. (+12 more)
Generating deep research reports requires large-scale information acquisition and the synthesis of insight-driven analysis, posing a significant challenge for current
language models. Most existing approaches follow a plan-then-write paradigm, whose performance heavily depends on the quality of the initial outline. However, constructing a comprehensive outline itself demands strong
reasoning ability, causing current deep research systems to rely almost exclusively on closed-source or online large models. This reliance raises practical barriers to deployment and introduces safety and
privacy concerns for user-authored data. In this work, we present AgentCPM-Report, a lightweight yet high-performing local solution composed of a framework that mirrors the human writing process and an 8B-parameter deep research
agent. Our framework uses a Writing As
Reasoning Policy (WARP), which enables models to dynamically revise outlines during report generation. Under this policy, the
agent alternates between Evidence-Based Drafting and
Reasoning-Driven Deepening, jointly supporting information acquisition, knowledge refinement, and iterative outline
evolution. To effectively equip small models with this capability, we introduce a Multi-Stage Agentic Training strategy, consisting of cold-start, atomic skill RL, and holistic pipeline RL. Experiments on DeepResearch Bench, DeepConsult, and DeepResearch Gym demonstrate that AgentCPM-Report outperforms leading closed-source systems, with substantial gains in Insight.
Read the full abstract ▾
All authors Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun
Computation and Language · Cross-listed from cs.AI
LogicSkills: A Structured Benchmark for Formal Reasoning in Large Language Models
Brian Rabern, Philipp Mondorf and Barbara Plank
Large
language models perform well on many logical
reasoning benchmarks, but it remains unclear which core logical skills they truly master. To address this, we introduce LogicSkills, a
benchmark that isolates three fundamental logical skills: (i) $\textit{formal symbolization}\unicode{x2014}{}$translating premises into first-order logic; (ii) $\textit{countermodel construction}\unicode{x2014}$showing that an argument is logically invalid by constructing a finite countermodel; and (iii) $\textit{validity assessment}\unicode{x2014}$determining whether a conclusion follows from a set of premises. Items are drawn from the two-variable fragment of first-order logic without identity and are presented in both English and a Carrollian nonce-word
language. All instances are solver-verified with Z3 for correctness and non-triviality. Across conventional instruction-tuned LLMs, performance is high on $\textit{validity assessment}$ but substantially lower on $\textit{formal symbolization}$ and $\textit{countermodel construction}$, highlighting that high task-level accuracy can mask weaknesses in core logical skills. In contrast, recent
reasoning-tuned models perform strongly across all three tasks, suggesting a more systematic logical skill profile.
Read the full abstract ▾
Computation and Language
Completing Missing Annotation: Multi-Agent Debate for Accurate and Scalable Relevant Assessment for IR Benchmarks
Minjeong Ban, Jeonghwan Choi, Hyangsuk Min et al. (+4 more)
Information
retrieval (IR) evaluation remains challenging due to incomplete IR
benchmark datasets that contain unlabeled relevant chunks. While LLMs and LLM-human hybrid strategies reduce costly human effort, they remain prone to LLM overconfidence and ineffective AI-to-human escalation. To address this, we propose DREAM, a multi-round debate-based relevance assessment framework with LLM
agents, built on opposing initial stances and iterative reciprocal critique. Through our agreement-based debate, it yields more accurate labeling for certain cases and more reliable AI-to-human escalation for uncertain ones, achieving 95.2% labeling accuracy with only 3.5% human involvement. Using DREAM, we build BRIDGE, a refined
benchmark that mitigates evaluation bias and enables fairer retriever comparison by uncovering 29,824 missing relevant chunks. We then re-
benchmark IR systems and extend evaluation to RAG, showing that unaddressed holes not only distort retriever rankings but also drive
retrieval-generation misalignment. The relevance assessment framework is available at https: //github.com/DISL-Lab/DREAM-ICLR-26; and the BRIDGE
dataset is available at https://github.com/DISL-Lab/BRIDGE-
Benchmark.
Read the full abstract ▾
All authors Minjeong Ban, Jeonghwan Choi, Hyangsuk Min, Nicole Hee-Yeon Kim, Minseok Kim, Jae-Gil Lee, Hwanjun Song
Computation and Language · Cross-listed from cs.HC
Designing Computational Tools for Exploring Causal Relationships in Qualitative Data
Han Meng, Qiuyuan Lyu, Peinuan Qin et al. (+4 more)
Exploring causal relationships for qualitative data analysis in HCI and social science research enables the understanding of user needs and theory building. However, current computational tools primarily characterize and categorize qualitative data; the few systems that analyze causal relationships either inadequately consider context, lack credibility, or produce overly complex outputs. We first conducted a formative study with 15 participants interested in using computational tools for exploring causal relationships in qualitative data to understand their needs and derive design guidelines. Based on these findings, we designed and implemented QualCausal, a system that extracts and illustrates causal relationships through interactive causal network construction and multi-view visualization. A feedback study (n = 15) revealed that participants valued our system for reducing the analytical burden and providing cognitive scaffolding, yet navigated how such systems fit within their established research paradigms, practices, and habits. We discuss broader implications for designing computational tools that support qualitative data analysis.
Read the full abstract ▾
All authors Han Meng, Qiuyuan Lyu, Peinuan Qin, Yitian Yang, Renwen Zhang, Wen-Chieh Lin, Yi-Chieh Lee
Computation and Language
Revisiting the Shape Convention of Transformer Language Models
Feng-Ting Liao, Guan-Ting Yi, Tzu-Quan Lin et al. (+2 more)
The architectural shape of dense
Transformers has remained remarkably stable: narrow-wide-narrow feed-forward networks (FFNs) consume most non-
embedding parameters. Motivated by theoretical and empirical evidences that residual wide-narrow-wide (hourglass) MLPs remain expressive despite bottlenecks, we revisit whether this architectural convention is necessary for dense
language models. We study Hourglass
Transformers, which replace the conventional FFN with residual stacks of hourglass sub-MLPs and use hourglass
attention to decouple residual-stream width from
attention width. This exposes a practical depth-width trade-off: compressing the FFN intermediate dimension allows wider hidden states and fewer layers at matched parameter budgets. Across model scales from 113M to 8B parameters, Hourglass
Transformers achieve
language-modeling and downstream performance comparable to conventional
Transformers, while improving training compute efficiency by $8.7\%$ at matched average downstream accuracy across the 906M, 3B, and 8B scales. After long-context extension, the 8B Hourglass model also outperforms its matched conventional baseline across 4k-64k context lengths. At 64k context, the reduced
attention layer count lowers both computation and KV-cache requirements, yielding up to $1.93\times$ faster
token decoding and $50\%$ lower KV-cache
memory at the 1B scale. These results identify hourglass structures as a practical architecture-efficiency alternative for compute- and latency-conscious
Transformer design.
Read the full abstract ▾
All authors Feng-Ting Liao, Guan-Ting Yi, Tzu-Quan Lin, Meng-Hsi Chen, Da-shan Shiu
Computation and Language
Improve Large Language Model Systems with User Logs
Changyue Wang, Weihang Su, Qingyao Ai et al. (+4 more)
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 .
Read the full abstract ▾
All authors Changyue Wang, Weihang Su, Qingyao Ai, Xingzhao Yue, Rui Zhang, Xiaojia Chang, Yiqun Liu
Computation and Language
Daisuke Oba, Hiroki Furuta and Naoaki Okazaki
Masked
diffusion language models generate text through iterative masked-
token filling, but terminal-only rewards on final completions provide coarse credit assignment for the intermediate filling decisions that shape the generation process. We propose
Diffusion-State Policy Optimization (DiSPO), a plug-in credit-assignment layer that directly optimizes intermediate filling decisions. At selected intermediate masked states, DiSPO branches by resampling the currently masked positions from rollout-cached logits, scores the resulting completions, and updates only the newly filled
tokens, requiring no additional multi-step
diffusion rollouts or optimizer steps. We formalize a fixed-state objective for branched completions and derive a policy-gradient
estimator that reuses the same rollouts as terminal-feedback policy optimization. Experiments on LLaDA-8B-Instruct show that DiSPO consistently improves terminal-feedback baselines, including diffu-GRPO and SPG, on math and planning
benchmarks under matched rollout compute and optimizer steps, supporting its use as a general plug-in for masked
diffusion policy optimization. Our project page is available at https://daioba.github.io/dispo .
Read the full abstract ▾
Computation and Language
RelayGen: Intra-Generation Model Switching for Efficient Reasoning
Jiwon Song, Yoongon Kim and Jae-Joon Kim
Large
reasoning models (LRMs) achieve strong performance on complex
reasoning tasks by generating long, multi-step
reasoning trajectories, but inference-time scaling incurs substantial deployment cost. A key challenge is that generation difficulty varies within a single output, whereas existing efficiency-oriented approaches either ignore this intra-generation variation or rely on supervised
token-level routing with high system complexity. We present RelayGen, a training-free, segment-level runtime model switching framework that exploits difficulty variation in long-form
reasoning. Through offline analysis of generation uncertainty using
token probability margins, we show that coarse-grained segment-level control is sufficient to capture difficulty transitions within a
reasoning trajectory. RelayGen identifies model-specific switch cues that signal transitions to lower-difficulty segments and dynamically delegates their continuation to a smaller model, while preserving high-difficulty
reasoning on the large model. Across multiple
reasoning benchmarks, RelayGen substantially reduces inference latency while preserving most of the accuracy of large models. When combined with speculative decoding, RelayGen achieves up to 2.2$\times$ end-to-end speedup with less than 2% accuracy degradation, without requiring additional training or learned routing components.
Read the full abstract ▾
Computation and Language
An evidence-guided reinforcement learning method to improve psychiatric reasoning in small language models
Xinxin Lin, Guangxin Dai, Yi Zhong et al. (+25 more)
Privacy and computational constraints limit the use of large
language models in psychiatry, while adapting small
language models (SLMs) often requires substantial data and expert annotation. We developed ClinMPO, an evidence-guided reinforcement-learning framework guided by the psychiatrist-defined Clinical Psychiatry Thinking Strategy (CPTS). ClinMPO uses ClinRM, a reward model trained on 18,569 question–answer pairs from 4,474 psychiatry articles. We evaluated four Qwen3 sizes on 1,737 model-screened questions. ClinMPO outperformed Base, supervised fine-tuning and standard group relative policy optimization across scales. From responses by 300 senior pre-licensure medical students, we established the human baseline, a medical-student reference. The 4B model approached this baseline, whereas the 8B model surpassed it and ranked first among 31 models and post-training variants. ClinMPO improved performance across two complementary schemes covering ICD-11 diagnostic categories and psychiatric practice competencies. Blinded assessment by three clinicians showed improved rationale quality across CPTS criteria. These findings highlight how existing clinical evidence and specialist knowledge can be incorporated into the development of medical AI systems through evidence-guided learning.
Read the full abstract ▾
All authors Xinxin Lin, Guangxin Dai, Yi Zhong, Xiang Li, Xue Xiao, Jian Liu, Yixin Zhang, Lingming Hu, Zhengdong Wu, Yongbo Zheng, Runchuan Zhu, Ming Zhao, Huizi Yu, Yi Zhang, Fangting Lu, Shuo Wu, Jun Zhao, Ping Yin, Joey W. Y. Chan, Ngan Yin Chan, Yumei Wang, Lejin Yang, Yanqiu Xing, Sijing Chen, Yun Kwok Wing, Lin Lu, Xin Ma, Lizhou Fan
Computation and Language
CORE: Comprehensive Ontological Relation Evaluation for Large Language Models
Satyam Dwivedi, Sanjukta Ghosh, Shivam Dwivedi et al. (+7 more)
Large
Language Models (LLMs) perform well on many
reasoning benchmarks, yet existing evaluations rarely assess their ability to distinguish between meaningful semantic relations and genuine unrelatedness. We introduce CORE (Comprehensive Ontological Relation Evaluation), a
dataset of 225K multiple-choice questions spanning 74 disciplines, together with a general-domain open-source
benchmark of 203 rigorously validated questions (Cohen's Kappa = 1.0) covering 24 semantic relation types with equal representation of unrelated pairs. A human baseline from 1,000+ participants achieves 92.6% accuracy (95.1% on unrelated pairs). In contrast, 29 state-of-the-art LLMs achieve 48.25-70.9% overall accuracy, with near-ceiling performance on related pairs (86.5-100%) but severe degradation on unrelated pairs (0-41.35%), despite assigning similar confidence (92-94%). Expected Calibration Error increases 2-4x on unrelated pairs, and a mean semantic collapse rate of 37.6% indicates systematic generation of spurious relations. On the CORE 225K MCQs
dataset, accuracy further drops to approximately 2%, highlighting substantial challenges in domain-specific semantic
reasoning. We identify unrelatedness
reasoning as a critical, under-evaluated frontier for LLM evaluation and safety.
Read the full abstract ▾
All authors Satyam Dwivedi, Sanjukta Ghosh, Shivam Dwivedi, Nishi Kumari, Anil Thakur, Anurag Purushottam, Deepak Alok, Praveen Gatla, Manjuprasad B, Bipasha Patgiri
Computation and Language
TrailBlazer: History-Guided Reinforcement Learning for Black-Box LLM Jailbreaking
Sung-Hoon Yoon, Ruizhi Qian, Minda Zhao et al. (+2 more)
Large
Language Models (LLMs) have become integral to many domains, making their safety a critical priority. Prior jailbreaking research has explored diverse approaches, including
prompt optimization, automated red teaming, obfuscation, and reinforcement learning (RL) based methods. However, most existing techniques fail to effectively leverage vulnerabilities revealed in earlier interaction turns, resulting in inefficient and unstable attacks. Since jailbreaking involves sequential interactions in which each response influences future actions, reinforcement learning provides a natural framework for this problem. Motivated by this, we propose a history-aware RL-based jailbreak framework that analyzes and reweights vulnerability signals from prior steps to guide future decisions. We show that incorporating historical information alone improves jailbreak success rates. Building on this insight, we introduce an
attention-based reweighting mechanism that highlights critical vulnerabilities within the interaction history, enabling more efficient exploration with fewer queries. Extensive experiments on AdvBench and HarmBench demonstrate that our method achieves state-of-the-art jailbreak performance while significantly improving query efficiency. These results underscore the importance of historical vulnerability signals in reinforcement learning-driven jailbreak strategies and offer a principled pathway for advancing
adversarial research on LLM safeguards.
Read the full abstract ▾
All authors Sung-Hoon Yoon, Ruizhi Qian, Minda Zhao, Weiyue Li, Mengyu Wang
Computation and Language
Investigating the structure of emotions by analyzing similarity and association of emotion words
Fumitaka Iwaki and Tatsuji Takahashi
In the field of natural
language processing, some studies have attempted sentiment analysis on text by handling
emotions as explanatory or response variables. One of the most popular
emotion models used in this context is the wheel of
emotion proposed by Plutchik. This model schematizes human
emotions in a circular structure, and represents them in two or three dimensions. However, the validity of Plutchik's wheel of
emotion has not been sufficiently examined. This study investigated the validity of the wheel by creating and analyzing a semantic networks of
emotion words. Through our experiments, we collected data of similarity and association of ordered pairs of
emotion words, and constructed networks using these data. We then analyzed the structure of the networks through community detection, and compared it with that of the wheel of
emotion. The results showed that each network's structure was, for the most part, similar to that of the wheel of
emotion, but locally different.
Read the full abstract ▾
Computation and Language
On the Wings of Imagination: Conflicting Script-based Multi-role Framework for Humor Caption Generation
Wenbo Shang, Yuxi Sun, Jing Ma et al. (+1 more)
Humor is a commonly used and intricate human
language in daily life. Humor generation, especially in multi-modal scenarios, is a challenging task for large
language models (LLMs), which is typically as funny caption generation for images, requiring visual understanding, humor
reasoning, creative imagination, and so on. Existing LLM-based approaches rely on
reasoning chains or self-improvement, which suffer from limited creativity and interpretability. To address these bottlenecks, we develop a novel LLM-based humor generation mechanism based on a fundamental humor theory, GTVH. To produce funny and script-opposite captions, we introduce a humor-theory-driven multi-role LLM collaboration framework augmented with humor
retrieval (HOMER). The framework consists of three LLM-based roles: (1) conflicting-script extractor that grounds humor in key script oppositions, forming the basis of caption generation; (2)
retrieval-augmented hierarchical imaginator that identifies key humor targets and expands the creative space of them through diverse associations structured as imagination trees; and (3) caption generator that produces funny and diverse captions conditioned on the obtained knowledge. Extensive experiments on two New Yorker Cartoon benchmarking
datasets show that HOMER outperforms state-of-the-art baselines and powerful LLM
reasoning strategies on multi-modal humor captioning.
Read the full abstract ▾
All authors Wenbo Shang, Yuxi Sun, Jing Ma, Xin Huang
Computation and Language
Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding
Daisuke Oba, Danushka Bollegala, Masahiro Kaneko et al. (+1 more)
Masked
Diffusion Language Models generate sequences via iterative sampling that progressively unmasks
tokens. However, they still recompute the
attention and feed-forward blocks for every
token position at every step – even when many unmasked
tokens are essentially fixed, resulting in substantial waste in compute. We propose SureLock: when the posterior at an unmasked position has stabilized across steps (our sure condition), we lock that position – thereafter skipping its query projection and feed-forward sublayers – while caching its
attention keys and values so other positions can continue to attend to it. This reduces the dominant per-iteration computational cost from $O(N^2d)$ to $O(MNd)$ where $N$ is the sequence length, $M$ is the number of unlocked
token positions, and $d$ is the model dimension. In practice, $M$ decreases as the iteration progresses, yielding substantial savings. On LLaDA-8B, SureLock reduces algorithmic FLOPs by 30–50% relative to the same sampler without locking, while maintaining comparable generation quality. We also provide a theoretical analysis to justify the design rationale of SureLock: monitoring only the local KL at the lock step suffices to bound the deviation in final
token probabilities. Our project page is available at https://daioba.github.io/surelock .
Read the full abstract ▾
All authors Daisuke Oba, Danushka Bollegala, Masahiro Kaneko, Naoaki Okazaki
Computation and Language
FMBench: Adaptive Large Language Model Output Formatting
Yaoting Wang, Yun Zhou and Henghui Ding
Producing outputs that satisfy both semantic intent and format constraints is essential for deploying large
language models in user-facing and system-integrated workflows. In this work, we focus on Markdown formatting, which is ubiquitous in assistants, documentation, and tool-augmented pipelines but still prone to subtle, hard-to-detect errors (e.g., broken lists, malformed tables, inconsistent headings, and invalid code blocks) that can significantly degrade downstream usability. We present FMBench, a
benchmark for adaptive Markdown output formatting that evaluates models under a wide range of instruction-following scenarios with diverse structural requirements. FMBench emphasizes real-world formatting behaviors such as multi-level organization, mixed content (natural
language interleaved with lists/tables/code), and strict adherence to user-specified layout constraints. To improve Markdown compliance without relying on hard decoding constraints, we propose a lightweight
alignment pipeline that combines supervised fine-tuning (SFT) with reinforcement learning fine-tuning. Starting from a base model, we first perform SFT on instruction-response pairs, and then optimize a composite objective that balances semantic fidelity with structural correctness. Experiments on two model families (OpenPangu and Qwen) show that SFT consistently improves semantic
alignment, while reinforcement learning provides additional gains in robustness to challenging Markdown instructions when initialized from a strong SFT policy. Our results also reveal an inherent trade-off between semantic and structural objectives, highlighting the importance of carefully designed rewards for reliable formatted generation. Code is available at: https://github.com/FudanCVL/FMBench.
Read the full abstract ▾
Computation and Language
ReBeCA: Causality-Inspired Behavior Analysis for Understanding the Self-Reflection of Language Models
Tianqiang Yan, Lizhen Qu, Sihan Shang et al. (+6 more)
Self-reflection is a basic self-improvement skill of
language agents, yet it does not reliably improve their outputs, where the underlying behavioral mechanism remains opaque. We introduce **ReBeCA** (self-**Re**flection **Be**havior explained through **C**ausality-inspired **A**nalysis), a framework for identifying and validating outcome-relevant reflective behaviors from recorded trajectories. ReBeCA uses **reflective codes**, an inductive coding schema, to represent each trajectory as a sequence of reflective behaviors. It learns dependencies among these behaviors and the outcome with causal discovery, selects the behaviors that best predict the outcome of held-out trajectories, and tests them for consistency across data subsets and by
prompt-based steering. In a controlled case study of four Qwen3 models under fixed-round Self-Refine on mathematical
reasoning and translation, ReBeCA identifies compact, round-specific behavioral sets with superior held-out predictive fit and uses
prompt-based steering on unseen problems to reveal how individual and combined behavioral directives affect performance (Holm-adjusted $p = .020$ on Qwen3-4B). ReBeCA offers a reusable framework that turns
agent trajectories into interpretable, verifiable behavioral hypotheses without requiring model internals or preselecting the behaviors to test.
Read the full abstract ▾
All authors Tianqiang Yan, Lizhen Qu, Sihan Shang, Yuheng Li, Song Qiu, Hao Peng, Wenjian Luo, Jue Xie, Yuan Gao
Computation and Language
Cost-Aware Model Selection for Text Classification: Multi-Objective Trade-offs Between Fine-Tuned Encoders and LLM Prompting in Production
Alberto Andres Valdes Gonzalez
Large
language models (LLMs) such as GPT-4o and Claude Sonnet 4.5 have demonstrated strong capabilities in open-ended
reasoning and generative
language tasks, leading to their widespread adoption across a broad range of NLP applications. However, for structured text classification problems with fixed label spaces, model selection is often driven by predictive performance alone, overlooking operational constraints encountered in production systems. In this work, we present a systematic comparison of two contrasting paradigms for text classification: zero- and few-shot
prompt-based large
language models, and fully fine-tuned encoder-only architectures. We evaluate these approaches across four canonical
benchmarks (IMDB, SST-2, AG News, and DBPedia), measuring predictive quality (macro F1), inference latency, and monetary cost. We frame model evaluation as a multi-objective decision problem and analyze trade-offs using Pareto frontier projections and a parameterized utility function reflecting different deployment regimes. Our results show that fine-tuned encoder-based models from the BERT family achieve competitive, and often superior, classification performance while operating at one to two orders of magnitude lower cost and latency compared to zero- and few-shot LLM prompting. Overall, our findings suggest that indiscriminate use of large
language models for standard text classification workloads can lead to suboptimal system-level outcomes. Instead, fine-tuned encoders emerge as robust and efficient components for structured NLP pipelines, while LLMs are better positioned as complementary elements within hybrid architectures. We release all code,
datasets, and evaluation protocols to support reproducibility and cost-aware NLP system design.
Read the full abstract ▾