Computation and Language · Cross-listed from cs.LG
Balanced Aggregation: Understanding and Fixing Aggregation Bias in GRPO
Zhiyuan Zeng, Jiameng Huang, Zhangyue Yin et al. (+8 more)
Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving
reasoning and code generation in large
language models, and GRPO-style training is widely adopted for its simplicity and effectiveness. However, an important design choice remains underexplored: how
token-level policy gradient terms are aggregated within each sampled group. Standard GRPO uses sequence aggregation, while recent work has advocated
token aggregation as a better alternative. We show that these two rules induce different optimization biases:
token aggregation introduces sign-length coupling, while sequence aggregation implicitly downweights longer responses through sequence-level equal weighting. To address this tension, we propose Balanced Aggregation (BA), a simple drop-in replacement that computes
token-level means separately within the positive and negative subsets and then combines them with sequence-count-based weights. Experiments with Qwen2.5-Math-7B and Qwen3-1.7B on DAPO-17k and Polaris, evaluated on six
reasoning and coding
benchmarks, show that BA consistently improves training stability and final performance over standard
token and sequence aggregation. Our analysis further shows that the relative effectiveness of
token and sequence aggregation is largely governed by response-length variation and the positive-negative length gap, highlighting aggregation as a critical design dimension in GRPO-style RLVR.
Read the full abstract ▾
All authors Zhiyuan Zeng, Jiameng Huang, Zhangyue Yin, Jiashuo Liu, Ziniu Li, Bingrui Li, Yuhao Wu, Yining Zheng, Ge Zhang, Wenhao Huang, Xipeng Qiu
Computation and Language · Cross-listed from cs.LG
RetentiveKV: State-Space Memory for Uncertainty-Aware Multimodal KV Cache Eviction
Sihao Liu, YuFan Xiong, Zhonghua Jiang et al. (+2 more)
Multimodal Large
Language Models face severe challenges in computational efficiency and
memory consumption due to the substantial expansion of the visual KV cache when processing long visual contexts. Existing KV cache compression methods typically rely on the "persistence of importance" hypothesis to prune
tokens. However, this approach proves fragile in
multimodal settings due to two key issues: 1) Visual
tokens display "deferred importance," initially exhibiting low salience but becoming pivotal during later decoding, which can lead to premature eviction. 2) Discrete
pruning disrupts the inherent spatial continuity of visual cues. To address these challenges, we propose RetentiveKV, an
entropy-driven KV cache optimization method that reformulates KV eviction from "discrete context truncation" to "continuous
memory evolution" based on State Space Models. Our method leverages information
entropy to quantify the information potential of low-
attention tokens and integrates
tokens scheduled for eviction into a continuous state space through
entropy-guided state transitions, enabling their dynamic reactivation when semantic relevance arises during subsequent decoding. Extensive experiments on
multimodal benchmarks demonstrate that RetentiveKV achieves 5.0 times KV cache compression and 1.5 times decoding acceleration.
Read the full abstract ▾
All authors Sihao Liu, YuFan Xiong, Zhonghua Jiang, Zhaode Wang, chengfei lv Shengyu Zhang
Computation and Language · Cross-listed from eess.AS
In-Sync: Adaptation of Speech Aware Large Language Models for ASR with Word Level Timestamp Predictions
Xulin Fan, Vishal Sunder, Samuel Thomas et al. (+3 more)
Recent advances in
speech-aware
language models have coupled strong acoustic encoders with large
language models, enabling systems that move beyond transcription to produce richer outputs. Among these, word-level timestamp prediction is critical for applications such as captioning, media search, and
multimodal synchronization, yet it is often handled by external
alignment tools. In this work, we extend an existing
speech-aware
language model to predict timestamps directly alongside transcripts. We introduce a set of novel lightweight training strategies that improve
alignment robustness while preserving recognition quality. Experiments across multiple
datasets show that these strategies not only enhance timestamp accuracy, but also yield gains in overall ASR performance. Together, they demonstrate an efficient and unified approach to
speech recognition with precise timestamp prediction.
Read the full abstract ▾
All authors Xulin Fan, Vishal Sunder, Samuel Thomas, Mark Hasegawa-Johnson, Brian Kingsbury, George Saon
Computation and Language · Cross-listed from cs.CV
Medical thinking with multiple images
Zonghai Yao, Benlu Wang, Yifan Zhang et al. (+8 more)
Large
language models perform well on many medical QA
benchmarks, but real clinical
reasoning often requires integrating evidence across multiple images rather than interpreting a single view. We introduce MedThinkVQA, an expert-annotated
benchmark for thinking with multiple images, where models must interpret each image, combine cross-view evidence, and answer diagnostic questions with intermediate supervision and step-level evaluation. The
dataset contains 8,067 cases, including 720 test cases, with an average of 6.62 images per case, substantially denser than prior work, whose expert-level
benchmarks use at most 1.43 images per case. On the test set, the best closed-source models, Claude-4.6-Opus, Gemini-3-Pro, and GPT-5.2-xhigh, reach only 57.2%, 55.3%, and 54.9% accuracy, while GPT-5-mini and GPT-5-nano reach 39.7% and 30.8%. Strong open-source models lag behind, led by Qwen3.5-397B-A17B at 52.2% and Qwen3.5-27B at 50.6%. Further analysis identifies grounded multi-image
reasoning as the main bottleneck: models often fail to extract, align, and compose evidence across views before higher-level inference can help. Providing expert single-image cues and cross-image summaries improves performance, whereas replacing them with self-generated intermediates reduces accuracy. Step-level analysis shows that over 70% of errors arise from image reading and cross-view integration. Scaling results further show that additional inference-time computation helps only when visual grounding is already reliable; when early evidence extraction is weak, longer
reasoning yields limited or unstable gains and can amplify misread cues. These results suggest that the key challenge is not
reasoning length alone, but reliable mechanisms for grounding, aligning, and composing distributed evidence across real-world
multimodal clinical inputs.
Read the full abstract ▾
All authors Zonghai Yao, Benlu Wang, Yifan Zhang, Junda Wang, Iris Xia, Zhipeng Tang, Shuo Han, Feiyun Ouyang, Zhichao Yang, Arman Cohan, Hong Yu
Computation and Language · Cross-listed from cs.SE
Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
Jiacheng Liu, Xiaohan Zhao, Xinyi Shang et al. (+1 more)
Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its architecture by analyzing the publicly available source code and comparing it with two independent open-source AI
agent systems, OpenClaw and Hermes
Agent, that answer many of similar or even the same design questions. Our analysis identifies five human values, philosophies, and needs that motivate the architecture: human decision authority, safety, security, and
privacy, reliable execution, capability amplification, and contextual adaptability. We then trace them through thirteen design principles to implementation choices. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. Most of the code, however, lives in the systems around this loop: a permission system with seven modes and an ML-based classifier, a five-layer compaction pipeline for context management, four extensibility mechanisms (MCP, plugins, skills, and hooks), a subagent delegation and orchestration mechanism, and append-oriented session storage. Comparisons with OpenClaw and Hermes
Agent show that the same design questions produce different answers across three deployment contexts. Claude Code emphasizes per-action safety, OpenClaw emphasizes perimeter-level access, and Hermes renders per-action approvals across many surfaces. At the runtime layer, Claude Code uses a single CLI loop, OpenClaw embeds the runtime within a gateway control plane, and Hermes uses one process whose role is set by its entry point. At the context and extension layer, Claude Code extends the context window, OpenClaw registers gateway-wide capabilities, and Hermes provides pluggable
memory and model backends. We finally identify six open design directions for future
agent systems, grounded in recent empirical, architectural, and policy literature.
Read the full abstract ▾
All authors Jiacheng Liu, Xiaohan Zhao, Xinyi Shang, Zhiqiang Shen
Computation and Language
Peer-Predictive Self-Training for Language Model Reasoning
Shi Feng, Hanlin Zhang, Fan Nie et al. (+2 more)
Mechanisms for continued self-improvement of
language models without external supervision remain an open challenge. We propose Peer-Predictive Self-Training (PST), a label-free fine-tuning framework in which multiple
language models improve collaboratively by using a cross-model aggregate response as an internal training signal. Given a
prompt, models generate responses sequentially; the final aggregated answer, which is often more reliable than individual responses in practice, serves as an internal reference for learning. We measure how informative each intermediate response is about the aggregate using pointwise mutual information (PMI), and use this signal to scale self-training updates: responses already aligned with the aggregate receive smaller updates, while less informative or misaligned responses receive larger ones. On mathematical
reasoning benchmarks, including SimulEq, MATH-500-Numeric, and MultiArith, PST improves exact-match accuracy by 2.2–4.3 percentage points across Gemma-2-2B, LLaMA-3.2-1B, and Qwen2.5-1.5B, and reduces the average generator–verifier gap (GV-Gap) by 26–40%, while requiring no external supervision, no teacher–student hierarchy, and only cross-model interactions. These results suggest that peer-predictive feedback from cross-model generations can provide an effective mechanism for self-supervised
language-model improvement.
Read the full abstract ▾
All authors Shi Feng, Hanlin Zhang, Fan Nie, Sham Kakade, Yiling Chen
Computation and Language
AgentSPEX: An Agent SPecification and EXecution Language
Pengcheng Wang, Jerry Huang, Jiarui Yao et al. (+7 more)
Language-model
agent systems commonly rely on reactive prompting, in which a single instruction guides the model through an open-ended sequence of
reasoning and tool-use steps, leaving control flow and intermediate state implicit and making
agent behavior potentially difficult to control. Orchestration frameworks such as LangGraph, DSPy, and CrewAI impose greater structure through explicit workflow definitions, but tightly couple workflow logic with Python, making
agents difficult to maintain and modify. In this paper, we introduce AgentSPEX, an
Agent SPecification and EXecution
Language for specifying LLM-
agent workflows with explicit control flow and modular structure, along with a customizable
agent harness. AgentSPEX supports typed steps, branching and loops, parallel execution, reusable submodules, and explicit state management, and these workflows execute within an
agent harness that provides tool access, a sandboxed virtual environment, and support for checkpointing, verification, and logging. Furthermore, we provide a visual editor with synchronized graph and workflow views for authoring and inspection. We include ready-to-use
agents for deep research and scientific research, and we evaluate AgentSPEX on 7
benchmarks. Finally, we show through a user study that AgentSPEX provides a more interpretable and accessible workflow-authoring paradigm than a popular existing
agent framework.
Read the full abstract ▾
All authors Pengcheng Wang, Jerry Huang, Jiarui Yao, Rui Pan, Peizhi Niu, Yaowenqi Liu, Ruida Wang, Renhao Lu, Yuwei Guo, Tong Zhang
Computation and Language · Cross-listed from cs.AI
WebXSkill: Skill Learning for Autonomous Web Agents
Zhaoyang Wang, Qianhui Wu, Xuchao Zhang et al. (+12 more)
Autonomous web
agents powered by large
language models (LLMs) remain brittle on long-horizon browser workflows. A key bottleneck is a grounding gap in existing skill formulations: textual workflow skills provide natural
language guidance but cannot be directly executed, while code-based skills execute without giving the
agent step-level guidance for adaptation or recovery. We introduce WebXSkill, a framework that bridges this gap with executable skills, each pairing a parameterized action program with step-level natural-
language guidance. WebXSkill operates in three stages: skill extraction mines reusable action subsequences from readily available synthetic
agent trajectories and abstracts them into parameterized skills, skill organization indexes them into a URL-based graph for context-aware
retrieval, and skill deployment exposes two complementary modes, grounded mode for fully automated execution and guided mode where skills serve as step-by-step instructions the
agent follows with its native planning. WebXSkill demonstrates consistent improvements on WebArena, WebVoyager, and Online-Mind2Web. We further find that better skill deployment mode depends on a model's plan and execution capability. The code is available at https://github.com/aiming-lab/WebXSkill.
Read the full abstract ▾
All authors Zhaoyang Wang, Qianhui Wu, Xuchao Zhang, Chaoyun Zhang, Wenlin Yao, Fazle Elahi Faisal, Baolin Peng, Si Qin, Suman Nath, Qingwei Lin, Chetan Bansal, Dongmei Zhang, Saravan Rajmohan, Jianfeng Gao, Huaxiu Yao
Computation and Language
Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus
John E. Ortega, Rodolfo Zevallos and Fabricio Carraro
We present a unified pipeline for synthesizing high-quality Quechua and Spanish
speech for the Peruvian Constitution using three state-of-the-art text-to-
speech (TTS) architectures: XTTS v2, F5-TTS, and DiFlow-TTS. Our models are trained on independent Spanish and Quechua
speech datasets with heterogeneous sizes and recording conditions, and leverage bilingual and
multilingual TTS capabilities to improve synthesis quality in both
languages. By exploiting cross-lingual transfer, our framework mitigates data scarcity in Quechua while preserving naturalness in Spanish. We release trained checkpoints, inference code, and synthesized audio for each constitutional article, providing a reusable resource for
speech technologies in indigenous and
multilingual contexts. This work contributes to the development of inclusive TTS systems for political and legal content in low-resource settings.
Read the full abstract ▾
Computation and Language
English is Not All You Need: Systematically Exploring the Role of Multilinguality in LLM Post-Training
Mehak Dhaliwal, Shashwat Chaurasia, Yao Qin et al. (+2 more)
Despite the widespread
multilingual deployment of large
language models, post-training pipelines remain predominantly English-centric, contributing to performance disparities across
languages. We present a systematic, controlled study of the interplay between training
language coverage, model scale, and task domain, based on 220 supervised fine-tuning runs on parallel translated
multilingual data mixtures spanning mathematical
reasoning and API calling tasks, with models up to 8B parameters. We find that English-only post-training is typically suboptimal: incorporating even a single non-English
language improves both English performance and cross-lingual generalization. Increasing
language diversity during post-training generally yields further gains, particularly for low-resource
languages, while performance on high-resource
languages tends to plateau rather than degrade. Moreover, greater
language diversity enables strong zero-shot transfer to unseen
languages, reducing the need for direct inclusion, though gains remain limited for typologically distant, low-resource
languages.
Read the full abstract ▾
All authors Mehak Dhaliwal, Shashwat Chaurasia, Yao Qin, Dezhi Hong, Thomas Butler
Computation and Language
L2D-Clinical: Learning to Defer for Adaptive Model Selection in Clinical Text Classification
Rishik Kondadadi and John E. Ortega
Clinical text classification requires choosing between specialized fine-tuned models (BERT variants) and general-purpose large
language models (LLMs), yet neither dominates across all instances. We introduce Learning to Defer for clinical text (L2D-Clinical), a framework that learns when a BERT classifier should defer to an LLM based on uncertainty signals and text characteristics. Unlike prior L2D work that defers to human experts assumed universally superior, our approach enables adaptive deferral-improving accuracy when the LLM complements BERT. We evaluate on two English clinical tasks: (1) ADE detection (ADE Corpus V2), where BioBERT (F1=0.911) outperforms the LLM (F1=0.765), and (2) treatment outcome classification (MIMIC-IV with multi-LLM consensus ground truth), where GPT-5-nano (F1=0.967) outperforms ClinicalBERT (F1=0.887). On ADE, L2D-Clinical achieves F1=0.928 (+1.7 points over BERT) by selectively deferring 7% of instances where the LLM's high recall compensates for BERT's misses. On MIMIC, L2D-Clinical achieves F1=0.980 (+9.3 points over BERT) by deferring only 16.8% of cases to the LLM. The key insight is that L2D-Clinical learns to selectively leverage LLM strengths while minimizing API costs.
Read the full abstract ▾
Computation and Language
Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size
Dikshant Kukreja, Kshitij Sah, Gautam Gupta et al. (+5 more)
Larger
language models become simultaneously better and worse at handling contextual information – better at ignoring false claims, worse at ignoring irrelevant
tokens. We formalize this apparent paradox through the first scaling laws for contextual entrainment, the tendency of models to favor
tokens that appeared in context regardless of relevance. Analyzing the Cerebras-GPT (111M-13B) and Pythia (410M-12B) model families, we find entrainment follows predictable power-law scaling, but with opposite trends depending on context type: semantic contexts show decreasing entrainment with scale, while non-semantic contexts show increasing entrainment. Concretely, the largest models are four times more resistant to counterfactual misinformation than the smallest, yet simultaneously twice as prone to copying arbitrary
tokens. These diverging trends, which replicate across model families, suggest that semantic filtering and mechanical copying are functionally distinct behaviors that scale in opposition – scaling alone does not resolve context sensitivity, it reshapes it.
Read the full abstract ▾
All authors Dikshant Kukreja, Kshitij Sah, Gautam Gupta, Avinash Anand, Rajiv Ratn Shah, Zhengkui Wang, Aik Beng Ng, Erik Cambria
Computation and Language · Cross-listed from cs.CV
Bahey Tharwat, Giorgos Kordopatis-Zilos, Pavel Suma et al. (+2 more)
Multimodal Large
Language Models (MLLMs) have demonstrated strong cross-modal
reasoning capabilities, yet their potential for
vision-only tasks remains underexplored. We investigate MLLMs as training-free similarity
estimators for instance-level image-to-image
retrieval. Our approach
prompts the model with paired images and converts next-
token probabilities into similarity scores, enabling zero-shot re-ranking within large-scale
retrieval pipelines. This design avoids specialized architectures and fine-tuning, leveraging the rich visual discrimination learned during
multimodal pre-training. We address scalability by combining MLLMs with
memory-efficient indexing and top-$k$ candidate re-ranking. Experiments across diverse
benchmarks show that MLLMs outperform task-specific re-rankers outside their native domains and exhibit superior robustness to clutter, occlusion, and small objects. Despite strong results, we identify failure modes under severe appearance changes, highlighting opportunities for future research. Our findings position MLLMs as a promising alternative for open-world large-scale image
retrieval.
Read the full abstract ▾
All authors Bahey Tharwat, Giorgos Kordopatis-Zilos, Pavel Suma, Ian Reid, Giorgos Tolias
Computation and Language
Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs
Vishal Pramanik, Maisha Maliha, Nathaniel D. Bastian et al. (+1 more)
Attribution methods seek to explain
language model predictions by quantifying the contribution of input
tokens to generated outputs. However, most existing techniques are designed for encoder-based architectures and rely on linear approximations that fail to capture the causal and semantic complexities of autoregressive generation in decoder-only models. To address these limitations, we propose Hessian-Enhanced
Token Attribution (HETA), a novel attribution framework tailored for decoder-only
language models. HETA combines three complementary components: a semantic transition vector that captures
token-to-
token influence across layers, Hessian-based sensitivity scores that model second-order effects, and KL divergence to measure information loss when
tokens are masked. This unified design produces context-aware, causally faithful, and semantically grounded attributions. Additionally, we introduce a curated
benchmark dataset for systematically evaluating attribution quality in generative settings. Empirical evaluations across multiple models and
datasets demonstrate that HETA consistently outperforms existing methods in attribution faithfulness and
alignment with human annotations, establishing a new standard for interpretability in autoregressive
language models.
Read the full abstract ▾
All authors Vishal Pramanik, Maisha Maliha, Nathaniel D. Bastian, Sumit Kumar Jha
Computation and Language
Evaluating the Evaluator: Problems with SemEval-2020 Task 1 for Lexical Semantic Change Detection
Bach Phan-Tat, Kris Heylen, Dirk Geeraerts et al. (+2 more)
This discussion paper re-examines SemEval-2020 Task 1, the most influential shared
benchmark for lexical semantic change detection, through a three-part evaluative framework: operationalisation, data quality, and
benchmark design. First, at the level of operationalisation, we argue that the
benchmark models semantic change mainly as gain, loss, or redistribution of discrete senses. While practical for annotation and evaluation, this framing is too narrow to capture gradual, constructional, collocational, and discourse-level change. Also, the gold labels are outcomes of annotation decisions, clustering procedures, and threshold settings, which could potentially limit the validity of the task. Second, at the level of data quality, we show that the
benchmark is affected by substantial corpus and preprocessing problems, including OCR noise, malformed characters, truncated sentences, inconsistent lemmatisation, POS-tagging errors, and missed targets. These issues can distort model behaviour, complicate linguistic analysis, and reduce reproducibility. Third, at the level of bench-mark design, we argue the small curated target sets and limited
language coverage reduce realism and increase statistical uncertainty. Taken together, these limitations suggest that the
benchmark should be treated as a useful but partial test bed rather than a definitive measure of progress. We therefore call for future
datasets and shared tasks to adopt broader theories of semantic change, document pre-processing transparently, expand cross-linguistic coverage, and use more realistic evaluation settings. Such steps are necessary for more valid, interpretable, and generalisable progress in lexical semantic change detection
Read the full abstract ▾
All authors Bach Phan-Tat, Kris Heylen, Dirk Geeraerts, Stefano De Pascale, Dirk Speelmana
Computation and Language
InfiniteScienceGym: An Unbounded, Procedurally-Generated Benchmark for Scientific Analysis
Oliver Bentham and Vivek Srikumar
Large
language models are emerging as scientific assistants, but evaluating their ability to reason from empirical data remains challenging.
Benchmarks derived from published studies and human annotations inherit publication bias, known-knowledge bias, label noise, and substantial storage requirements. We present InfiniteScienceGym, a procedurally generated
benchmark of scientific repositories paired with a verifiable question-answering task. From a seed, the simulator deterministically generates a self-contained repository with realistic directory structure, files, and tabular data, and a privileged QA generator produces both answerable and unanswerable questions with exact ground truth. This makes it possible to evaluate evidence-grounded
reasoning, abstention, and tool-mediated analysis in a controlled setting without distributing a large static corpus. InfiniteScienceGym complements real scientific
benchmarks by targeting blind spots and failure modes that are hard to evaluate using published
datasets alone. Evaluating both proprietary and open-weight models, we find that none achieve more than 50% accuracy overall, that recognizing unanswerable questions remains a major weakness, and that stronger models tend to use tools more effectively rather than simply consuming more
tokens.
Read the full abstract ▾
Computation and Language
Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization
Shiping Gao, Hongzhan Chen, Xiaojun Quan et al. (+2 more)
Process reward models (PRMs) provide fine-grained supervision for
reasoning, but reliable PRMs often require step annotations or heavy verification pipelines, making them costly to scale and refresh during online RL. Implicit PRMs reduce this cost by training log-likelihood-ratio rewards from trajectory-level outcome labels. However, the log-ratio is constrained only as a sequence-level aggregate during training, while inference decomposes it into
token- or step-level scores for partial prefixes. This train-inference mismatch leaves local credits weakly identified, so distribution-wide scoring can amplify misleading advantages. We propose Implicit Prefix-Value Reward Model (IPVRM), which directly learns the probability of eventual correctness for each prefix from outcome labels. Step signals are then obtained as temporal-difference (TD) differences between consecutive prefix values, aligning the training target with inference-time use. IPVRM markedly improves step-verification F1 on ProcessBench. To exploit these prefix values during policy optimization, we further introduce Distribution-Level RL (DistRL), which applies TD advantages to both sampled
tokens and high-probability candidate
tokens, providing dense counterfactual updates without additional rollouts. Experiments show that DistRL brings limited gains with unreliable implicit rewards, but consistently improves downstream
reasoning when paired with IPVRM. The implementation of our method is available at https://github.com/gaoshiping/IPVRM .
Read the full abstract ▾
All authors Shiping Gao, Hongzhan Chen, Xiaojun Quan, Qifan Wang, Lifu Huang
Computation and Language · Cross-listed from cs.CV
SceneCritic: A Symbolic Evaluator for 3D Indoor Scene Synthesis
Kathakoli Sengupta, Kai Ao and Paola Cascante-Bonilla
Large
Language Models (LLMs) and
Vision-
Language Models (VLMs) increasingly generate indoor scenes through intermediate structures such as layouts and scene graphs, yet evaluation still relies on LLM or VLM judges that score rendered views, making judgments sensitive to viewpoint,
prompt phrasing, and
hallucination. When the evaluator is unstable, it becomes difficult to determine whether a model has produced a spatially plausible scene or whether the output score reflects the choice of viewpoint, rendering, or
prompt. We introduce SceneCritic, a symbolic evaluator for floor-plan-level layouts. SceneCritic's constraints are grounded in SceneOnto, a structured spatial ontology we construct by aggregating indoor scene priors from 3D-FRONT, ScanNet, and Visual Genome. SceneOnto traverses this ontology to jointly verify semantic, orientation, and
geometric coherence across object relationships, providing object-level and relationship-level assessments that identify specific violations and successful placements. Furthermore, we pair SceneCritic with an iterative refinement test bed that probes how models build and revise spatial structure under different critic modalities: a rule-based critic using collision constraints as feedback, an LLM critic operating on the layout as text, and a VLM critic operating on rendered observations. Through extensive experiments, we show that (a) SceneCritic aligns substantially better with human judgments than VLM-based evaluators, (b) text-only LLMs can outperform VLMs on semantic layout quality, and (c) image-based VLM refinement is the most effective critic modality for semantic and orientation correction.
Read the full abstract ▾
Computation and Language
Toward Autonomous Long-Horizon Engineering for ML Research
Guoxin Chen, Jie Chen, Lei Chen et al. (+7 more)
Agentic systems increasingly automate pieces of AI research. Yet turning underspecified research objectives into runnable, experimentally validated ML systems remains a central bottleneck. We study this operational setting as long-horizon ML research engineering: converting a research specification into a runnable ML system through repeated implementation, experimentation, and refinement. The central challenge is to sustain cumulative project progress across heterogeneous stages under delayed, confounded feedback. We introduce AiScientist, a multi-
agent system built around thin control over thick state: a lightweight hierarchical research team coordinates through a File-as-Bus workspace that preserves decision-relevant artifacts across roles and invocations. On PaperBench, AiScientist improves over the strongest matched baselines by 9.92 and 11.15 points with Gemini-3-Flash and GLM-5, respectively. On MLE-Bench Lite, it reaches 81.82 Any Medal% under both backbones, improving over the strongest matched baselines by 4.55 and 16.67 points, and exceeding a Codex/GPT-5.5 xhigh frontier harness reference by 13.64 Any Medal points. Ablations and process analyses show that durable project state is central to later-round refinement: removing File-as-Bus lowers PaperBench score by 6.41 points and MLE-Bench Lite Any Medal% by 31.82 points. These results suggest that long-horizon AI research is not only a problem of stronger local
reasoning, but a systems problem of maintaining cumulative, inspectable project progress.
Read the full abstract ▾
All authors Guoxin Chen, Jie Chen, Lei Chen, Jiale Zhao, Fanzhe Meng, Wayne Xin Zhao, Ruihua Song, Cheng Chen, Ji-Rong Wen, Kai Jia
Computation and Language · Cross-listed from cs.LG
Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Yaxuan Li, Yuxin Zuo, Bingxiang He et al. (+8 more)
On-policy
distillation (OPD) has become a core technique in the post-training of large
language models, yet its training dynamics remain poorly understood. This paper provides a systematic investigation of OPD dynamics and mechanisms. We first identify that two conditions govern whether OPD succeeds or fails: (i) the student and teacher should share compatible thinking patterns; and (ii) even with consistent thinking patterns and higher scores, the teacher must offer genuinely new capabilities beyond what the student has seen during training. We validate these findings through weak-to-strong reverse
distillation, showing that same-family 1.5B and 7B teachers are distributionally indistinguishable from the student's perspective. Probing into the
token-level mechanism, we show that successful OPD is characterized by progressive
alignment on high-probability
tokens at student-visited states, a small shared
token set that concentrates most of the probability mass (97%-99%). We further propose two practical strategies to recover failing OPD: off-policy cold start and teacher-aligned
prompt selection. Finally, we show that OPD's apparent free lunch of dense
token-level reward comes at a cost, raising the question of whether OPD can scale to long-horizon
distillation.
Read the full abstract ▾
All authors Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, Ning Ding
Computation and Language
One Token Away from Collapse: The Fragility of Instruction-Tuned Helpfulness
Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu et al. (+1 more)
Instruction-tuned large
language models produce helpful, structured responses, but how robust is this helpfulness under trivial constraints? We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses, losing 14–48% of comprehensiveness across seven models spanning five families (7B–70B, open- and closed-weight). A blinded human evaluation with 10 STEM-trained evaluators confirms genuine content loss, with information criteria degrading $1.5$–$2.3\times$ more than surface criteria, a finding corroborated by over 4,100 automated pairwise comparisons (77–100% baseline preference) across three LLM judges from two model families. Diagnostic analysis identifies this as a planning failure: two-pass generation recovers 59–96% of response length, and linear probes on
prompt representations predict response length with $R^2 = 0.51$–$0.94$ before generation begins. The same probes yield negative $R^2$ on base models, confirming that instruction tuning introduces the representational structure underlying the collapse. Base models show no systematic degradation under identical constraints, demonstrating that instruction tuning couples task competence to narrow surface-form templates. The effect extends to realistic deployment constraints (preamble suppression, corporate tone guidelines, legal compliance hedging, accessibility requirements) causing comparable degradation ($-$22% to $-$34%), with suppressing the conversational opener alone (“Certainly!”) causing 40% collapse on our most fragile model despite restricting only the opening
tokens. We further show that standard independent LLM-as-judge evaluation detects only a 3.5% quality drop where pairwise evaluation reveals 23%, exposing a methodological blind spot in current evaluation practice.
Read the full abstract ▾
All authors Erfan Baghaei Potraghloo, Seyedarmin Azizi, Souvik Kundu, Massoud Pedram
Computation and Language
PolicyLLM: Towards Excellent Comprehension of Public Policy for Large Language Models
Han Bao, Penghao Zhang, Yue Huang et al. (+9 more)
Large
Language Models (LLMs) are increasingly integrated into real-world decision-making, including in the domain of public policy. Yet, their ability to comprehend and reason about policy-related content remains underexplored. To fill this gap, we present \textbf{PolicyBench}, the first large-scale cross-system
benchmark (US-China) evaluating policy comprehension, comprising 21K cases across a broad spectrum of policy areas, capturing the diversity and complexity of real-world governance. Following Bloom's taxonomy, the
benchmark assesses three core capabilities: (1) Memorization: factual recall of policy knowledge, (2) Understanding: conceptual and contextual
reasoning, and (3) Application: problem-solving in real-life policy scenarios. Building on this
benchmark, we further propose \textbf{PolicyMoE}, a domain-specialized Mixture-of-Experts (MoE) model with expert modules aligned to each cognitive level. The proposed models demonstrate stronger performance on application-oriented policy tasks than on memorization or conceptual understanding, and yields the highest accuracy on structured
reasoning tasks. Our results reveal key limitations of current LLMs in policy understanding and suggest paths toward more reliable, policy-focused models.
Read the full abstract ▾
All authors Han Bao, Penghao Zhang, Yue Huang, Zhengqing Yuan, Yanchi Ru, Rui Su, Yujun Zhou, Xiangqi Wang, Kehan Guo, Nitesh V Chawla, Yanfang Ye, Xiangliang Zhang
Computation and Language
Accelerating Speculative Decoding with Block Diffusion Draft Trees
Liran Ringel and Yaniv Romano
Speculative decoding accelerates autoregressive
language models by using a lightweight drafter to propose multiple future
tokens, which the target model then verifies in parallel. DFlash shows that a block
diffusion drafter can generate an entire draft block in a single forward pass and achieve state-of-the-art speculative decoding performance, outperforming strong autoregressive drafters such as EAGLE-3. Vanilla DFlash, however, still verifies only a single drafted trajectory per round, potentially limiting its acceptance length. We introduce DDTree (
Diffusion Draft Tree), a method that constructs a draft tree directly from the per-position distributions of a block
diffusion drafter. Under a fixed node budget, DDTree uses a simple best-first heap
algorithm to select the continuations that are most likely to match the target model according to a surrogate defined by the draft model's output. The resulting tree is verified efficiently in a single target model forward pass using an ancestor-only
attention mask. Because DDTree builds on DFlash, a leading draft model for speculative decoding, these gains place DDTree among the leading approaches to speculative decoding.
Read the full abstract ▾
Computation and Language
GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner et al. (+2 more)
Optical character recognition (OCR) has advanced rapidly with the rise of
vision-
language models, yet evaluation has remained concentrated on a small cluster of high- and mid-resource scripts. We introduce GlotOCR Bench, a comprehensive
benchmark evaluating OCR generalization across 100+ Unicode scripts. Our
benchmark comprises clean and degraded image variants rendered from real
multilingual texts. Images are rendered using fonts from the Google Fonts repository, shaped with HarfBuzz and rasterized with FreeType, supporting both LTR and RTL scripts. Samples of rendered images were manually reviewed to verify correct rendering across all scripts. We evaluate a broad suite of open-weight and proprietary
vision-
language models and find that most perform well on fewer than ten scripts, and even the strongest frontier models fail to generalize beyond thirty scripts. Performance broadly tracks script-level pretraining coverage, suggesting that current OCR systems rely on
language model pretraining as much as on visual recognition. Models confronted with unfamiliar scripts either produce random noise or hallucinate characters from similar scripts they already know. We release the
benchmark and pipeline for reproducibility. Pipeline Code: https://github.com/cisnlp/glotocr-bench,
Benchmark: https://hf.co/
datasets/cis-lmu/glotocr-bench.
Read the full abstract ▾
All authors Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner, François Yvon, Hinrich Schütze
Computation and Language
MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models
Chung-Ming Chien, Manu Orsini, Eugene Kharitonov et al. (+3 more)
Speech-to-
speech language models have recently emerged to enhance the naturalness of conversational AI. In particular, full-duplex models are distinguished by their real-time interactivity, including handling of pauses, interruptions, and backchannels. However, improving their factuality remains an open challenge. While scaling the model size could address this gap, it would make real-time inference prohibitively expensive. In this work, we propose MoshiRAG, a modular approach that combines a compact full-duplex interface with selective
retrieval to access more powerful knowledge sources. Our asynchronous framework enables the model to identify knowledge-demanding queries and ground its responses in external information. By leveraging the natural temporal gap between response onset and the delivery of core information, the
retrieval process can be completed while maintaining a natural conversation flow. With this approach, MoshiRAG achieves factuality comparable to the best publicly released non-duplex
speech language models while preserving the interactivity inherent to full-duplex systems. Moreover, our flexible design supports plug-and-play
retrieval methods without retraining and demonstrates strong performance on out-of-domain mathematical
reasoning tasks.
Read the full abstract ▾
All authors Chung-Ming Chien, Manu Orsini, Eugene Kharitonov, Neil Zeghidour, Karen Livescu, Alexandre Défossez
Computation and Language
MetFuse: Figurative Fusion between Metonymy and Metaphor
Saptarshi Ghosh and Tianyu Jiang
Metonymy and metaphor often co-occur in natural
language, yet computational work has studied them largely in isolation. We introduce a framework that transforms a literal sentence into three figurative variants: metonymic, metaphoric, and hybrid. Using this framework, we construct MetFuse, the first dedicated
dataset of figurative fusion between metonymy and metaphor, containing 1,000 human-verified meaning-aligned quadruplets totaling 4,000 sentences. Extrinsic experiments on eight existing
benchmarks show that augmenting training data with MetFuse consistently improves both metonymy and metaphor classification, with hybrid examples yielding the largest gains on metonymy tasks. Using this
dataset, we also analyze how the presence of one figurative type impacts another. Our findings show that both human annotators and large
language models better identify metonymy in hybrid sentences than in metonymy-only sentences, demonstrating that the presence of a metaphor makes a metonymic noun more explicit. Our
dataset is publicly available at: https://github.com/cincynlp/MetFuse.
Read the full abstract ▾
Computation and Language
Round-Trip Translation Reveals What Frontier Multilingual Benchmarks Miss
Ronald Skorobogat, Ameya Prabhu and Matthias Bethge
Multilingual benchmarks guide the development of frontier models. Yet
multilingual evaluations reported by frontier models are structured similar to popular
reasoning and knowledge
benchmarks, but across many
languages. We show such
benchmarks, and consequently
multilingual evaluations, measure mathematical
reasoning and factual recall, not
multilingual proficiency. For example, thinking variants dramatically outperform instruct variants on these
benchmarks, yet often perform worse on real-world
multilingual tasks, such as LMArena. We propose a simple alternative: evaluate
multilingual capability via round-trip translation. Given text in a source
language, translate it to a target
language and back; semantic gaps between the original and result expose failures in
multilingual generation capabilities. Round-trip translation correlates almost perfectly (h̊o} = 0.94) with user ratings on LMArena with our
benchmark, requires no human reference translations, and does not require a more capable
multilingual judge than tested models. Lastly, we introduce Lost in Translation (LiT), a challenging round-trip translation
benchmark spanning widely spoken
languages worldwide, for realistic evaluation of
multilingual frontier models.
Read the full abstract ▾
Computation and Language
Growing Pains: Extensible and Efficient LLM Benchmarking Via Fixed Parameter Calibration
Eliya Habba, Itay Itzhak, Asaf Yehudai et al. (+5 more)
The rapid release of both
language models and
benchmarks makes it increasingly costly to evaluate every model on every
dataset. In practice, models are often evaluated on different samples, making scores difficult to compare across studies. To address this, we propose a framework based on multidimensional Item Response Theory (IRT) that uses anchor items to calibrate new
benchmarks to the evaluation suite while holding previously calibrated item parameters fixed. Our approach supports a realistic evaluation setting in which
datasets are introduced over time and models are evaluated only on the
datasets available at the time of evaluation, while a fixed anchor set for each
dataset is used so that results from different evaluation periods can be compared directly. In large-scale experiments on more than 400 models, our framework predicts full-evaluation performance within 2-3 percentage points using only 100 anchor questions per
dataset, with Spearman $\rho \geq 0.9$ for ranking preservation preservation. This shows that
benchmark suites can grow over time while preserving score comparability, since adding a new
dataset requires running existing models only on that
dataset's anchors. Code and data are available at: https://eliyahabba.github.io/growing-pains/
Read the full abstract ▾
All authors Eliya Habba, Itay Itzhak, Asaf Yehudai, Yotam Perlitz, Elron Bandel, Michal Shmueli-Scheuer, Leshem Choshen, Gabriel Stanovsky
Computation and Language · Cross-listed from cs.AI
RePAIR: Interactive Machine Unlearning through Prompt-Aware Model Repair
Jagadeesh Rachapudi, Pranav Singh, Ritali Vatsi et al. (+2 more)
Large
language models (LLMs) inherently absorb harmful knowledge, misinformation, and personal data during pretraining on large-scale web corpora, with no native mechanism for selective removal. While machine unlearning offers a principled solution, existing approaches are provider-centric, requiring retraining pipelines, curated retain
datasets, and direct intervention by model service providers (MSPs), thereby excluding end users from controlling their own data. We introduce Interactive Machine Unlearning (IMU), a new paradigm in which users can instruct LLMs to forget targeted knowledge through natural
language at inference time. To realize IMU, we propose RePAIR, a
prompt-aware model repair framework comprising (i) a watchdog model for unlearning intent detection, (ii) a surgeon model for generating repair procedures, and (iii) a patient model whose parameters are updated autonomously. At the core of RePAIR, we develop Steering Through Activation Manipulation with PseudoInverse (STAMP), a training-free, single-sample unlearning method that redirects MLP activations toward a refusal subspace via closed-form pseudoinverse updates. Its low-rank variant reduces computational complexity from O(d^3) to O(r^3 + r^2 * d), enabling efficient on-device unlearning with up to ~3x speedup over training-based baselines. Extensive experiments across harmful knowledge suppression, misinformation correction, and personal data erasure demonstrate that RePAIR achieves near-zero forget scores (Acc_f = 0.00, F-RL = 0.00) while preserving model utility (Acc_r up to 84.47, R-RL up to 0.88), outperforming six state-of-the-art baselines. These results establish RePAIR as an effective and practical framework for user-driven model editing, advancing transparent and on-device control over learned knowledge, with potential extensions to
multimodal foundation models.
Read the full abstract ▾
All authors Jagadeesh Rachapudi, Pranav Singh, Ritali Vatsi, Praful Hambarde, Amit Shukla
Computation and Language
The role of System 1 and System 2 semantic memory structure in human and LLM biases
Katherine Abramski, Giulio Rossetti and Massimo Stella
Implicit biases in both humans and large
language models (LLMs) pose significant societal risks. Dual process theories propose that biases arise primarily from associative System 1 thinking, while deliberative System 2 thinking mitigates bias, but the cognitive mechanisms that give rise to this phenomenon remain poorly understood. To better understand what underlies this duality in humans, and possibly in LLMs, we model System 1 and System 2 thinking as semantic
memory networks with distinct structures, built from comparable
datasets generated by both humans and LLMs. We then investigate how these distinct semantic
memory structures relate to implicit gender bias using network-based evaluation metrics. We find that semantic
memory structures are irreducible only in humans, suggesting that LLMs lack certain types of human-like conceptual knowledge. Moreover, semantic
memory structure relates consistently to implicit bias only in humans, with lower levels of bias in System~2 structures. These findings suggest that certain types of conceptual knowledge contribute to bias regulation in humans, but not in LLMs, highlighting fundamental differences between human and machine cognition.
Read the full abstract ▾