Peer Revenue Publish or profit

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Thursday 7 May

Computation and Language

Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries

Kun Zeng, Yu Huo, Siyu Zhang et al. (+6 more)

Skill-augmented agents increasingly rely on large reusable skill libraries, but retrieving relevant skills is not the same as presenting usable context. Existing methods typically return atomic skills or dependency-aware bundles whose internal roles remain implicit, leaving the agent to infer the execution entry point, support skills, visible requirements, and failure-avoidance guidance. We introduce Group of Skills (GoSkills), an inference-time group-structured retrieval method that changes the agent-facing retrieval object from a flat skill list to a compact, role-labeled execution context. GoSkills builds anchor-centered skill groups from a typed skill graph, expands support groups through a group graph, bottlenecks the selected group plan into a bounded set of atomic skill payloads, and renders a fixed execution contract with Start, Support, Check, and Avoid fields, without changing the downstream agent, skill payloads, or execution environment. Experiments on SkillsBench and ALFWorld show that GoSkills preserves visible-requirement coverage under a small skill budget, improves over flat skill-access baselines, and often improves reward and agent-only runtime relative to structural retrieval references.
Read the full abstract

All authors Kun Zeng, Yu Huo, Siyu Zhang, Zi Ye, Yuecheng Zhuo, Haoyue Liu, Yuquan Lu, Junhao Wen, Xiaoying Tang

Computation and Language · Cross-listed from cs.HC

From Surface Learning to Deep Understanding: A Grounded AI Tutoring System for Moodle

Anna Ostrowska, Michał Kukla, Gabriela Majstrak et al. (+4 more)

This demo paper describes the development of the AI Teaching & Learning Assistant, a modular Moodle plugin that leverages Retrieval-Augmented Generation (RAG) to deliver high-quality, hallucination-free education. The system employs a dual-centric design, providing students with interactive, Socratic-based tutoring and educators with a "human-in-the-loop" workspace for supervised content generation. By grounding Large Language Model (LLM) responses in teacher-provided materials, the assistant addresses the risks of misinformation while encouraging deep conceptual mastery. Evaluation via the Ragas (LLM-as-a-Judge) framework and a preliminary user study confirms its effectiveness, achieving faithfulness scores up to 0.97 and a 4.00/5.00 recommendation rate.
Read the full abstract

All authors Anna Ostrowska, Michał Kukla, Gabriela Majstrak, Jan Opala, Sebastian Pergała, Jan Skwarek, Anna Wróblewska

Computation and Language

MultiSoc-4D: A Benchmark for Diagnosing Instruction-Induced Label Collapse in Closed-Set LLM Annotation of Bengali Social Media

Souvik Pramanik, S.M. Riaz Rahman Antu, Shak Mohammad Abyad et al. (+2 more)

Annotation automation via Large Language Models (LLMs) is the core approach for scaling NLP datasets; however, LLM behavior with respect to closed-set instructions in low-resource languages has not been well studied. We present MultiSoc-4D, a Bengali social media dataset benchmark, which contains 58K+ social media comments from six sources annotated along four dimensions: category, sentiment, hate speech, and sarcasm. By employing a structured pipeline where ChatGPT, Gemini, Claude, and Grok individually annotate separate partitions, while sharing a common validation set of 20%, we diagnose LLM behavior systematically. We discover a prevalent phenomenon called "instruction-induced label collapse", wherein LLMs show a systematic preference towards fallback labels (Other, Neutral, No), leading to high agreement rates but under-detection of minority categories. For example, we find that LLMs failed to detect 79% and 75% of instances with hateful and sarcastic content compared to a human-calibrated reference. Furthermore, we prove that it represents a "label agreement illusion", statistically validated via almost null Fleiss' Kappa ($\kappa \approx -0.001$) on sarcasm detection. Across 40+ LLMs, we benchmark this annotation bias propagation within the training pipeline, regardless of architectural differences. We release MultiSoc-4D as a diagnostic benchmark for annotation biases in Bengali NLP.
Read the full abstract

All authors Souvik Pramanik, S.M. Riaz Rahman Antu, Shak Mohammad Abyad, Md. Ibrahim Khalil, Md. Shahriar Hussain

Computation and Language

Can LLMs Take Retrieved Information with a Grain of Salt?

Behzad Shayegh, Mohamed Osama Ahmed, Fred Tung et al. (+1 more)

Large language models have demonstrated impressive retrieval-augmented capabilities. However, a crucial area remains underexplored: their ability to appropriately adapt responses to the certainty of the retrieved information. It is a limitation with real consequences in high-stakes domains like medicine and finance. We evaluate eight LLMs on their context-certainty obedience, measuring how well they adjust responses to match expressed context certainty. Our analysis reveals systematic limitations: LLMs struggle to recall prior knowledge after observing an uncertain context, misinterpret expressed certainties, and overtrust complex contexts. To address these, we propose an interaction strategy combining prior reminders, certainty recalibration, and context simplification. This approach reduces obedience errors by 25% on average, without modifying model weights, demonstrating the efficacy of interaction design in enhancing LLM reliability. Our contributions include a principled evaluation metric, empirical insights into LLMs' uncertainty handling, and a portable strategy to improve context-certainty obedience across diverse LLMs.
Read the full abstract

All authors Behzad Shayegh, Mohamed Osama Ahmed, Fred Tung, Leo Feng

Computation and Language · Cross-listed from cs.DC

Pacing Branch Parallelism in LLM Serving

Swapnil Gandhi, Siva Kumar Sastry Hari, William J. Dally et al. (+1 more)

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the branch externality and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce PACE, a per-step admission controller that treats extra branches as opportunistic width, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, PACE improves goodput by $1.77\times$ over serving without branch parallelism and by $1.48\times$ over eager admission while maintaining 99% SLO attainment. Comparable gains ($1.72$–$2.15\times$) hold on five additional dense and mixture-of-experts models with up to 671B parameters. Code is available at https://github.com/Swapnil-Gandhi/PACE
Read the full abstract

All authors Swapnil Gandhi, Siva Kumar Sastry Hari, William J. Dally, Christos Kozyrakis

Computation and Language

MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text

Chenjun Li, Cheng Wan, Haomiao Chen et al. (+1 more)

Large language models are widely used in everyday writing, making reliable AI-generated text detection crucial for academic integrity, content moderation, and provenance tracking. Yet high AUROC on clean, in-distribution benchmarks is not sufficient. Practical detectors must resist adversarial rewrites, generalize to unseen generators and writing domains, and maintain low false-positive rates (FPR). A pooled AI-versus-human objective does not explicitly require the model to distinguish among generator families, so it may fail to learn the generator-specific structure needed for generalization and attribution. We introduce MELD (Multi-Task Equilibrated Learning Detector), which instead trains on class-balanced, family-specific AI-versus-human tasks while sharing a common representation of human writing. MELD produces detection, generator-family, rewrite-task, and token-level predictions in a single forward pass. Format normalization further makes its scores invariant to the modeled reformatting operations. MELD ranks first among open-source submissions in the public RAID leaderboard and matches or exceeds supervised baselines on five of six held-out evaluation pools. To evaluate transfer to unseen generators, we introduce MELD-eval, a held-out test pool built from four frontier chat models. Without further fine-tuning, MELD achieves 99.7% TPR at 1% FPR on MELD-eval and 98% TPR at the same FPR on a held-out generator whose family is absent from training. Finally, in a case study of 5.9 million scientific texts from 2016–2026, MELD's prediction scores remain stable through 2022 and increase from 2023 onward, coinciding with the widespread adoption of LLM-based writing tools. The model, MELD-eval pool, source code, and live demo are available.
Read the full abstract

All authors Chenjun Li, Cheng Wan, Haomiao Chen, Johannes C. Paetzold

Computation and Language

Reflections and New Directions for Human-Centered Large Language Models

Caleb Ziems, Dora Zhao, Rose E. Wang et al. (+55 more)

Large Language Models (LLMs) are increasingly shaping the private and professional lives of users, with numerous applications in business, education, finance, healthcare, law, and science. With this rise in global influence comes greater urgency to build, evaluate, and deploy these systems in a manner that prioritizes not only technical capabilities but also human priorities. This work presents a framework for developing Human-Centered Large Language Models (HCLLMs), which integrates perspectives from Natural Language Processing (NLP), Human-Computer Interaction (HCI), and responsible AI. Considering the ethics, economics, and technical objectives of language modeling, we argue that model developers need to address human concerns, preferences, values, and goals, not only during a cursory post-training stage, but rather with rigor and care at every stage of the pipeline. This paper offers human-centered insights and recommendations for developers at each stage, from system design to data sourcing, model training, evaluation, and responsible deployment. Then we conclude with a case study, applying these insights to understand the future of work with HCLLMs.
Read the full abstract

All authors Caleb Ziems, Dora Zhao, Rose E. Wang, Matthew Jörke, Ahmad Rushdi, Advit Deepak, Sunny Yu, Anshika Agarwal, Harshvardhan Agarwal, Gabriela Aranguiz-Dias, Aditri Bhagirath, Justine Breuch, Huanxing Chen, Ruishi Chen, Sarah Chen, Haocheng Fan, William Fang, Cat Gonzales Fergesen, Daniel Frees, Tian Gao, Ziqing Huang, Vishal Jain, Yucheng Jiang, Kirill Kalinin, Su Doga Karaca, Arpandeep Khatua, Teland La, Isabelle Levent, Miranda Li, Xinling Li, Yongce Li, Angela Liu, Minsik Oh, Nathan J. Paek, Anthony Qin, Emily Redmond, Michael J. Ryan, Aadesh Salecha, Xiaoxian Shen, Pranava Singhal, Shashanka Subrahmanya, Mei Tan, Irawadee Thawornbut, Michelle Vinocour, Xiaoyue Wang, Zheng Wang, Henry Jin Weng, Pawan Wirawarn, Shirley Wu, Sophie Wu, Yichen Xie, Patrick Ye, Sean Zhang, Yutong Zhang, Cathy Zhou, Yiling Zhao, James Landay, Diyi Yang

Computation and Language

MIST: Multimodal Interactive Speech-based Tool-calling Conversational Assistants for Smart Homes

Maximillian Chen, Xuanming Zhang, Michael Peng et al. (+3 more)

The rise of Internet of Things (IoT) devices in the physical world necessitates voice-based interfaces capable of handling complex user experiences. While modern Large Language Models (LLMs) already demonstrate strong tool-usage capabilities, modeling real-world IoT devices presents a difficult, understudied challenge which combines modeling spatiotemporal constraints with speech inputs, dynamic state tracking, and mixed-initiative interaction patterns. We introduce MIST (the Multimodal Interactive Speech-based Tool-calling Dataset), a synthetic multi-turn, voice-driven code generation task that operates over IoT devices. We find that there is a significant gap between open- and closed-weight multimodal LLMs on MIST, and that even frontier closed-weight LLMs have substantial headroom. We release MIST and an extensible data generation framework to build related datasets in order to facilitate research on mixed-initiative voice assistants which reason about physical world constraints.
Read the full abstract

All authors Maximillian Chen, Xuanming Zhang, Michael Peng, Zhou Yu, Alexandros Papangelis, Yohan Jo

Computation and Language

TajPersLexon: A Tajik-Persian Lexical Resource and Hybrid Model for Cross-Script Low-Resource NLP

Mullosharaf K. Arabov

This work introduces TajPersLexon, a curated Tajik–Persian parallel lexical resource of 40,112 word and short-phrase pairs for cross-script lexical retrieval, transliteration, and alignment in low-resource settings. We conduct a comprehensive CPU-only benchmark comparing three methodological families: (i) a lightweight hybrid pipeline, (ii) neural sequence-to-sequence models, and (iii) retrieval methods. Our evaluation establishes that the task is essentially solvable, with neural and retrieval baselines achieving 98-99% top-1 accuracy. Crucially, we demonstrate that while large multilingual sentence transformers fail on this exact lexical matching, our interpretable hybrid model offers a favorable accuracy-efficiency trade-off for practical applications, achieving 96.4% accuracy in an OCR post-correction task. All experiments use fixed random seeds for full reproducibility. The dataset, code, and models will be publicly released.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Benchmarked Yet Not Measured – Generative AI Should be Evaluated Against Real-World Utility

Ishani Mondal and Shweta Bhardwaj

Generative AI systems achieve impressive performance on standard benchmarks yet fail to deliver real-world utility, a disconnect we identify across 28 deployment cases spanning education, healthcare, software engineering, and law. We argue that this benchmark utility gap arises from three recurring failures in evaluation practice: proxy displacement, temporal collapse, and distributional concealment. Motivated by these observations, we argue that generative AI evaluation requires a paradigm shift from static benchmark-centered transparency toward stakeholder, goal, and context-conditioned utility transparency grounded in human outcome trajectories. Existing evaluations primarily characterize properties of model outputs, while deployment success depends on whether interaction with AI improves stakeholders' ability to achieve their goals over time. The missing construct is therefore utility: the change in a stakeholder's capability induced through sustained interaction with an AI system within a deployment context. To operationalize this perspective, we propose SCU-GenEval, a four-stage evaluation framework consisting of stakeholder-goal mapping, construct-indicator specification, mechanism modeling, and longitudinal utility measurement. To make these stages practically deployable, we introduce three supporting instruments: structured deployment protocols, context-conditioned user simulators, and persona- and goal-conditioned proxy metrics. We conclude with domain-specific calls to action, arguing that progress in generative AI must be evaluated through measurable improvements in human outcomes rather than benchmark performance alone.
Read the full abstract

Computation and Language

IntentGrasp: A Comprehensive Benchmark for Intent Understanding

Yuwei Yin, Chuyuan Li and Giuseppe Carenini

Accurately understanding the intent behind speech, conversation, and writing is crucial to the development of helpful Large Language Model (LLM) assistants. This paper introduces IntentGrasp, a comprehensive benchmark for evaluating the intent understanding capability of LLMs. Derived from 49 high-quality, open-licensed corpora spanning 12 diverse domains, IntentGrasp is constructed through source datasets curation, intent label contextualization, and task format unification. IntentGrasp contains a large-scale training set of 262,759 instances and two evaluation sets: an All Set of 12,909 test cases and a more balanced and challenging Gem Set of 470 cases. Extensive evaluations on 20 LLMs across 7 families (including frontier models such as GPT-5.4, Gemini-3.1-Pro, and Claude-Opus-4.7) demonstrate unsatisfactory performance, with scores below 60% on All Set and below 25% on Gem set. Notably, 17 out of 20 tested models perform worse than a random-guess baseline (15.2%) on Gem Set, while the estimated human performance is ~81.1%, showing substantial room for improvement. To enhance such ability, this paper proposes Intentional Fine-Tuning (IFT), which fine-tunes the models on the training set in IntentGrasp, yielding significant gains of 30+ F1 points on All Set and 20+ points on Gem Set. Tellingly, the leave-one-domain-out (Lodo) experiments further demonstrate the strong cross-domain generalizability of IFT, verifying that it is a promising approach to substantially enhancing the intent understanding of LLMs. Overall, by benchmarking and boosting intent understanding ability, this study sheds light on a promising path towards more intentional, capable, and safe AI assistants for human benefits and social good.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

ProtSent: Protein Sentence Transformers

Dan Ofer, Oriel Perets, Michal Linial et al. (+1 more)

Protein language models produce rich per-residue representations, but the sequence embeddings obtained by mean-pooling them are never trained to place related proteins near one another, so the geometry that retrieval, clustering and nearest-neighbor transfer depend on is never optimized. We present ProtSent, a contrastive post-training framework that supervises a protein language model with three types of biological relation at once, Pfam family co-membership, AlphaFold DB structural clusters and STRING interactions, and with deep mutational scanning fitness ordering as an optional fourth. It needs no structure at training or inference, trains on corpora filtered at 40% sequence identity against the benchmark test sets they most resemble, and is applied to ESM-2 (35M and 150M parameters) and ESM-C (300M). Post-training makes structural family membership recoverable from the frozen embedding without a trained readout: on SCOPe-40, the adjusted Rand index of the 35M model rises from 0.054 (near chance) to 0.507, and family retrieval gains +0.185, +0.190 and +0.424 Recall@1 over the three frozen backbones, with paired bootstrap intervals excluding zero. On the CATH midnight-zone benchmark, homologous-superfamily accuracy rises from 40.7% to 56.7% at 35M and from 43.3% to 62.7% at 150M, more than the 12 points a CATH-supervised model gains over its own base. Against a maximally sensitive HMMER search, alignment keeps the best single hit at both ESM-2 scales, but the embedding leads at ranking depth (MAP 0.705 against 0.607 at 150M) and returns a ranking for every query, including those where alignment finds none. Under a trained linear probe, however, the same embeddings fall below the untuned backbone, as do structure-distilled embeddings against theirs, so the gain in neighborhood structure comes at a cost in linear-probe accuracy. Code and models: https://github.com/oriel9p/ProtSent
Read the full abstract

All authors Dan Ofer, Oriel Perets, Michal Linial, Nadav Rappoport

Computation and Language

VITA-QinYu: Expressive Spoken Language Model for Role-Playing and Singing

Jiacheng Xu, Heting Gao, Liufei Xie et al. (+8 more)

Human speech conveys expressiveness beyond linguistic content, including personality, mood, or performance elements, such as a comforting tone or humming a song, which we formalize as role-playing and singing. We present VITA-QinYu, the first expressive end-to-end (E2E) spoken language model (SLM) that goes beyond natural conversation to support both role-playing and singing generation. VITA-QinYu adopts a hybrid speech-text paradigm that extends interleaved text-audio modeling with multi-codebook audio tokens, a design enabling richer paralinguistic representation while preserving a clear separation between modalities to avoid interference. We further develop a comprehensive data generation pipeline to synthesize a total of 15.8K hours of natural conversation, role-playing, and singing data for training. VITA-QinYu demonstrates superior expressiveness, outperforming peer SLMs by 7 percentage points on objective role-playing benchmarks, and surpassing peer models by 0.13 points on a 5-point MOS scale for singing. Simultaneously, it achieves state-of-the-art conversational accuracy and fluency, exceeding prior SLMs by 1.38 and 4.98 percentage points on the C3 and URO benchmarks, respectively. We open-source our code and models and provide an easy-to-use demo with full-stack support for streaming and full-duplex interaction.
Read the full abstract

All authors Jiacheng Xu, Heting Gao, Liufei Xie, Zhenchuan Yang, Lijiang Li, Yiting Chen, Bin Zhang, Meng Chen, Chaoyu Fu, Weifeng Zhao, Wenjiang Zhou

Computation and Language · Cross-listed from cs.CR

When Routine Chats Turn Toxic: Unintended Long-Term State Poisoning in Personalized Agents

Xiaoyu Xu, Minxin Du, Qipeng Xie et al. (+3 more)

Personalized LLM agents maintain persistent cross-session state to support long-horizon collaboration. Yet, this persistence introduces a subtle but critical security vulnerability: routine user-agent interactions can gradually reshape an agent's long-term state, inadvertently weakening future confirmation boundaries, expanding tool-use defaults, and escalating autonomous behavior over time. We formalize this risk as unintended long-term state poisoning. To systematically study it, we introduce the Unintended Long-Term State Poisoning Bench (ULSPB), a bilingual benchmark comprising $350$ settings spanning five assistance categories, seven interaction patterns, 24-turn routine interactions, and matched single-injection counterparts. Furthermore, we define the Harm Score (HS), a state-centric metric that quantifies authorization drift, tool-use escalation, and unchecked autonomy. Experiments on OpenClaw with four backbone LLMs demonstrate that, while single-injection is generally effective, routine conversations alone can substantially poison long-term state, primarily corrupting memory-centric artifacts. Evaluations seeded with real-world user interactions confirm that this risk is not a mere artifact of synthetic prompts. To mitigate this threat, we propose StateGuard, a lightweight, post-execution defense that audits state diffs at the writeback boundary and selectively rolls back dangerous edits. Across all evaluated models, StateGuard reduces HS to near zero and lowers false-negative rates, with acceptable high false-positive rates under a safety-first writeback defense and minimal overhead.
Read the full abstract

All authors Xiaoyu Xu, Minxin Du, Qipeng Xie, Haobin Ke, Qingqing Ye, Haibo Hu

Computation and Language · Cross-listed from cs.AI

When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment

Long Zhang, Wei-neng Chen, Feng-feng Wei et al. (+1 more)

Language models often generate reasoning before giving a final answer, but the visible answer does not reveal when the model's answer preference became stable. We study this question through a narrow computable object: finite-answer preference stabilization. For a model state and specified answer verbalizers, we project the model's own continuation probabilities onto a finite answer set; in binary tasks this yields an exact log-odds code, $\delta(\xi)=S_\theta(\mathrm{yes}\mid\xi)-S_\theta(\mathrm{no}\mid\xi)$. This target defines parser-based answer onset, retrospective stabilization time, and lead without relying on greedy rollouts or learned probes. In controlled delayed-verdict tasks with Qwen3-4B-Instruct, the contextual finite-answer projection stabilizes before the answer is parseable, with 17–31 token mean lead in the main templates and positive, shorter lead in a parser-clean replication. The signal tracks the model's eventual output rather than truth, is linearly recoverable from compact hidden summaries, is partly separable from cursor progress, and transfers as shared information without a single invariant coordinate. Diagnostics separate the measurement from online stopping, verbalizer-free belief, and causal answer control; exact steering shows local sensitivity of $\delta$ but not reliable generation control.
Read the full abstract

All authors Long Zhang, Wei-neng Chen, Feng-feng Wei, Zi-bo Qin

Computation and Language · Cross-listed from cs.AI

From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

Jinghao Luo, Yuchen Tian, Chuxue Cao et al. (+6 more)

Large Language Model (LLM)-based agents have fundamentally reshaped artificial intelligence by integrating external tools and planning capabilities. While memory mechanisms have emerged as the architectural cornerstone of these systems, current research remains fragmented, oscillating between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis and a coherent evolutionary perspective. To bridge this gap, this survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing the development process into three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). We first formally define these three stages before analyzing the three core drivers of this evolution: the necessity for long-range consistency, the challenges in dynamic environments, and the ultimate goal of continual learning. Furthermore, we specifically explore two transformative mechanisms in the frontier Experience stage: proactive exploration and cross-trajectory abstraction. By synthesizing these disparate views, this work offers robust design principles and a clear roadmap for the development of next-generation LLM agents.
Read the full abstract

All authors Jinghao Luo, Yuchen Tian, Chuxue Cao, Ziyang Luo, Hongzhan Lin, Kaixin Li, Chuyi Kong, Ruichao Yang, Jing Ma

Computation and Language

EMO: Pretraining Mixture of Experts for Emergent Modularity

Ryan Wang, Akshita Bhagia and Sewon Min

Large language models are typically deployed as monolithic systems, requiring the full model even when applications need only a narrow subset of capabilities, e.g., code, math, or domain-specific knowledge. Mixture-of-Experts (MoEs) seemingly offer a potential alternative by activating only a subset of experts per input, but in practice, restricting inference to a subset of experts for a given domain leads to severe performance degradation. This limits their practicality in memory-constrained settings, especially as models grow larger and sparser. We introduce EMO, an MoE designed for modularity-the independent use and composition of expert subsets-without requiring human-defined priors. Our key idea is to encourage tokens from similar domains to rely on similar experts. Since tokens within a document often share a domain, EMO restricts them to select experts from a shared pool, while allowing different documents to use different pools. This simple constraint enables coherent expert groupings to emerge during pretraining using document boundaries alone. We pretrain a 1B-active, 14B-total EMO on 1T tokens. As a full model, it matches standard MoE performance. Crucially, it enables selective expert use: retaining only 25% (12.5%) of experts incurs just a 1% (3%) absolute drop, whereas standard MoEs break under the same setting. We further find that expert subsets in EMO specialize at semantic levels (e.g., domains such as math or code), in contrast to the low-level syntactic specialization observed in standard MoEs. Altogether, our results demonstrate a path toward modular, memory-efficient deployment of large, sparse models and open new opportunities for composable architectures.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Verifier-Backed Hard Problem Generation for Mathematical Reasoning

Yuhang Lai, Jiazhan Feng, Yee Whye Teh et al. (+1 more)

Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.
Read the full abstract

All authors Yuhang Lai, Jiazhan Feng, Yee Whye Teh, Ning Miao

Computation and Language · Cross-listed from cs.LG

When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels

Sushant Gautam, Finn Schwall, Annika Willoch Olstad et al. (+6 more)

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($\eta^2 \approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.
Read the full abstract

All authors Sushant Gautam, Finn Schwall, Annika Willoch Olstad, Fernando Vallecillos Ruiz, Birk Torpmann-Hagen, Sunniva Maria Stordal Bjørklund, Leon Moonen, Klas Pettersen, Michael A. Riegler

Computation and Language

Improving Reasoning Ability via Asynchronous On-Policy Self-Distillation under Positive Rollouts

Mingwei Xu and Hao Fang

Distillation and reinforcement learning through verifiable rewards (RLVR) have achieved progress in enhancing the reasoning ability of large language models (LLMs). However, we note that negative rollouts may admit no gradation of failure severity, and the combinatorial vastness makes penalizing a few sampled negatives unlikely to cover a meaningful reward signal under sparse binary rewards. In this work, we propose Positive-Only Policy Optimization (POPO), an on-policy self-distillation integrated RLVR framework in which learning occurs exclusively on online positive rollouts. Specifically, POPO utilizes bounded importance sampling over the positive rollout set. Thus, no disjoint negative rollouts are used for gradient guidance during post-training. We show that implicit negative gradients can emerge naturally through reinforcing the positive probability via rollout redistribution. Next, POPO stabilizes the policy optimization through self-distillation. First, it applies a Siamese policy network with a momentum-based adaptation law for asynchronous policy evolution. Second, we replace the KL-divergence with a bounded similarity penalty term in the Siamese representation space. We conduct extensive experiments using publicly available, well-established text-LLM models across all-level mathematical benchmarks (MATH-500, AMC23, AIME 2024/2025, and Olympiad). Our experiment demonstrates that POPO achieves superior performance compared to GRPO. Notably, we show that POPO can achieve 36.67% in AIME 2025 with Qwen-Math-7B, outperforming GRPO 30.00%. Our ablation and sweep studies further illustrate the necessity and robustness.
Read the full abstract

Computation and Language

StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction

Xiangyuan Xue, Yifan Zhou, Zidong Wang et al. (+5 more)

Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.
Read the full abstract

All authors Xiangyuan Xue, Yifan Zhou, Zidong Wang, Shengji Tang, Philip Torr, Wanli Ouyang, Lei Bai, Zhenfei Yin

Computation and Language · Cross-listed from cs.LG

Recursive Agent Optimization

Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang et al. (+2 more)

We introduce Recursive Agent Optimization (RAO), a reinforcement learning approach for training recursive agents: agents that can spawn and delegate sub-tasks to new instantiations of themselves recursively. Recursive agents implement an inference-time scaling algorithm that naturally allows agents to scale to longer contexts and generalize to more difficult problems via divide-and-conquer. RAO provides a method to train models to best take advantage of such recursive inference, teaching agents when and how to delegate and communicate. We find that recursive agents trained in this way enjoy better training efficiency, can scale to tasks that go beyond the model's context window, generalize to tasks much harder than the ones the agent was trained on, and can enjoy reduced wall-clock time compared to single-agent systems.
Read the full abstract

All authors Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, Graham Neubig

Computation and Language · Cross-listed from cs.AI

How Deep Can LLMs Learn to Reason? Expressiveness Is Key

Tianle Wang, Zhaoyang Wang, Guangchen Lan et al. (+4 more)

Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. Observed LLM shortcomings in long-horizon reasoning have raised the prospect that they are fundamental to the autoregressive transformer architecture. To address this, we introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying formal system. Our proposed framework supports a wide range of logics: from simple implication-only logic ("if-then") towards more expressive first-order reasoning with conjunction (”and”), disjunction (”or”), negation (”not”), and universal quantification ("for all"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \propto D^{\gamma}$, $R^{2} > 0.99$), and the scaling exponent $\gamma$ increases monotonically with expressiveness, from $1.05$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency. More broadly, our results suggest that LLM shortcomings in long-horizon reasoning are not fundamental to the underlying architecture, and could be addressed by improved training methodology and data.
Read the full abstract

All authors Tianle Wang, Zhaoyang Wang, Guangchen Lan, Xinpeng Wei, Sipeng Zhang, Guanwen Qiu, Abulhair Saparov

Computation and Language

Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents

Hailey Onweller, Elias Lumer, Austin Huber et al. (+3 more)

Large language models (LLMs) power deep research agents that synthesize information from hundreds of web sources into cited reports, yet these citations cannot be reliably verified. Current approaches either trust models to self-cite accurately, risking bias, or employ retrieval-augmented generation (RAG) that does not validate source accessibility, relevance, or factual consistency. We introduce the first source attribution evaluation framework that uses a reproducible AST parser to extract and evaluate inline citations from LLM-generated Markdown reports at scale. Unlike methods that verify claims in isolation, our framework closes the loop by retrieving the actual cited content, enabling human or model evaluators to judge each citation against its source. Citations are evaluated along three dimensions. (1) Link Works verifies URL accessibility, (2) Relevant Content measures topical alignment, and (3) Fact Check validates factual accuracy against source content. We benchmark 14 closed-source and open-source LLMs across three evaluation dimensions using rubric-based LLM-as-a-judge evaluators calibrated through human review. Our results reveal that even the strongest frontier models maintain link validity above 94% and relevance above 80%, yet achieve only 39-77% factual accuracy, while fewer than half of open-source models successfully generate cited reports in a one-shot setting. Ablation studies on research depth show that Fact Check accuracy drops by approximately 42% on average across two frontier models as tool calls scale from 2 to 150, demonstrating that more retrieval does not produce more accurate citations. These findings reveal a critical disconnect between surface-level citation quality and factual reliability, and our framework provides the evaluation infrastructure to assess the disconnect.
Read the full abstract

All authors Hailey Onweller, Elias Lumer, Austin Huber, Pia Ramchandani, Vamse Kumar Subbiah, Corey Feld

Computation and Language

Parser agreement and disagreement in L2 Korean UD: Implications for human-in-the-loop annotation

Hakyung Sung and Gyu-Ho Shin

We propose a simplified human-in-the-loop workflow for second language (L2) Korean morphosyntactic annotation by leveraging agreement between two domain-adapted parsers. We first evaluate whether parser agreement can serve as a proxy for annotation correctness by comparing it with independent human judgments. The results show strong correspondence between parser and human judgments, supporting the feasibility of semi-automatic L2-Korean UD annotation. Further analysis demonstrates that parser disagreements cluster in linguistically predictable domains such as grammatical-relation distinctions and clause-boundary ambiguity. While many disagreement cases are tractable for iterative model refinement, others reflect deeper representational challenges inherent in parsing and tagging L2-Korean corpora.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

Zhexuan Wang, Xuebo Liu, Li Wang et al. (+4 more)

Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at https://github.com/wangzx1219/MASPO.
Read the full abstract

All authors Zhexuan Wang, Xuebo Liu, Li Wang, Zifei Shan, Yutong Wang, Zhenxi Song, Min Zhang

Computation and Language

Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance

Jan Fillies, Ronald E. Robertson and Jeffrey Hancock

As large language models (LLMs) increasingly mediate both content generation and moderation, linguistic evasion strategies known as Algospeak have intensified the coevolution between evaders and detectors. This research formalizes the underlying dynamics grounded in a joint action model: when Algospeak increases, detectability and understandability decrease. Further, the concept of Majority Understandable Modulation (MUM) is introduced and defined as the modulation level at which additional evasive alteration increases detector evasion but loses comprehension for the majority of recipients. To empirically probe this trade-off, we introduce a reproducible framework that can be used to create meaning-preserving, Algospeak-style variants, based on an existing taxonomy and with tunable modulation levels. Using COVID-19 disinformation as a first proof-by-example setting, we construct a reference dataset of 700 modulated items, drawn from twenty base sentences across five modulation levels and seven strategies. We then run two linked evaluations with seven different language models: one testing for interpretation through meaning recovery and one for disinformation detection through classification. Curve fitting over modulation levels yields an estimate of the Majority Understandable Modulation threshold and enables sensitivity analyses across strategies and models, see Figure 1. Results reveal the characteristic relationships between understandability and modulation. This study lays the groundwork for understanding the dynamics behind Algospeak and provides the framework, dataset, and experimental setups described.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

When and Why SignSGD Outperforms SGD: A Theoretical Study Based on $\ell_1$-norm Lower Bounds

Hongyi Tao, Dingzhi Yu and Lijun Zhang

Sign-based optimization algorithms, such as SignSGD and Muon, have garnered significant attention for their remarkable performance in training large foundation models. Despite this empirical success, we still lack a theoretical understanding of when and why these sign-based methods outperform vanilla SGD. The core obstacle is that under standard smoothness and finite variance conditions, SGD is known to be minimax optimal for finding stationary points measured by $\ell_2$-norms, thereby fundamentally precluding any complexity gains for sign-based methods in standard settings. To overcome this barrier, we analyze sign-based optimizers leveraging $\ell_1$-norm stationarity, $\ell_\infty$-smoothness, and a separable noise model, which can better capture the coordinate-wise nature of signed updates. Under this distinct problem geometry, we derive matched upper and lower bounds for SignSGD and explicitly characterize the problem class in which SignSGD provably dominates SGD. Specifically, we compare the upper bound of SignSGD with the lower bound of SGD, illustrating that SignSGD effectively reduces the complexity by a factor of $d$ under sparse noise, where $d$ is the problem dimension. Furthermore, we elevate this framework to the matrix domain, providing an equivalent optimal lower bound for the Muon optimizer, proving that extending the sign operator to matrices preserves this optimal scaling with dimensionality. Finally, we bridge our theoretical bounds to practice, demonstrating that the theoretical superiority of SignSGD accurately predicts its faster convergence during the pretraining of a 124M parameter GPT-2 model.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

SkillOS: Learning Skill Curation for Self-Evolving Agents

Siru Ouyang, Jun Yan, Yanfei Chen et al. (+13 more)

LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.
Read the full abstract

All authors Siru Ouyang, Jun Yan, Yanfei Chen, Rujun Han, Zifeng Wang, Bhavana Dalvi Mishra, Rui Meng, Chun-Liang Li, Yizhu Jiao, Kaiwen Zha, Maohao Shen, Vishy Tirumalashetty, George Lee, Jiawei Han, Tomas Pfister, Chen-Yu Lee

Computation and Language

UniSD: Towards a Unified Self-Distillation Framework for Large Language Models

Yiqiao Jin, Yiyang Wang, Lucheng Fu et al. (+8 more)

Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.
Read the full abstract

All authors Yiqiao Jin, Yiyang Wang, Lucheng Fu, Yijia Xiao, Yinyi Luo, Haoxin Liu, B. Aditya Prakash, Chao Zhang, Josiah Hester, Jindong Wang, Srijan Kumar