Peer Revenue Equity in science

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Sunday 22 March

Computation and Language · Cross-listed from eess.AS

SqueezeComposer: Temporal Speed-up is A Simple Trick for Long-form Music Composing

Jianyi Chen, Rongxiu Zhong, Shilei Zhang et al. (+4 more)

Composing coherent long-form music remains a significant challenge due to the complexity of modeling long-range dependencies and the prohibitive memory and computational requirements associated with lengthy audio representations. In this work, we propose a simple yet powerful trick: we assume that AI models can understand and generate time-accelerated (speeded-up) audio at rates such as 2x, 4x, or even 8x. By first generating a high-speed version of the music, we greatly reduce the temporal length and resource requirements, making it feasible to handle long-form music that would otherwise exceed memory or computational limits. The generated audio is then restored to its original speed, recovering the full temporal structure. This temporal speed-up and slow-down strategy naturally follows the principle of hierarchical generation from abstract to detailed content, and can be conveniently applied to existing music generation models to enable long-form music generation. We instantiate this idea in SqueezeComposer, a framework that employs diffusion models for generation in the accelerated domain and refinement in the restored domain. We validate the effectiveness of this approach on two tasks: long-form music generation, which evaluates temporal-wise control (including continuation, completion, and generation from scratch), and whole-song singing accompaniment generation, which evaluates track-wise control. Experimental results demonstrate that our simple temporal speed-up trick enables efficient, scalable, and high-quality long-form music generation. Audio samples are available at https://SqueezeComposer.github.io/.
Read the full abstract

All authors Jianyi Chen, Rongxiu Zhong, Shilei Zhang, Kun Qian, Jinglei Liu, Yike Guo, Wei Xue

Computation and Language · Cross-listed from cs.AI

LongCat-Flash-Prover: Advancing Native Formal Reasoning via Agentic Tool-Integrated Reinforcement Learning

Jianing Wang, Jianfei Zhang, Qi Guo et al. (+24 more)

We introduce LongCat-Flash-Prover, a flagship 560-billion-parameter open-source Mixture-of- Experts (MoE) model that advances Native Formal Reasoning in Lean4 through agentic tool-integrated reasoning (TIR). We decompose the native formal reasoning task into three independent formal capabilities, i.e., auto-formalization, sketching, and proving. To facilitate these capabilities, we propose a Hybrid-Experts Iteration Framework to expand high-quality task trajectories, including generating a formal statement based on a given informal problem, producing a whole-proof directly from the statement, or a lemma-style sketch. During agentic RL, we present a Hierarchical Importance Sampling Policy Optimization (HisPO) algorithm, which aims to stabilize the MoE model training on such long-horizon tasks. It employs a gradient masking strategy that accounts for the policy staleness and the inherent train-inference engine discrepancies at both sequence and token levels. Additionally, we also incorporate theorem consistency and legality detection mechanisms to eliminate reward hacking issues. Extensive evaluations show that our LongCat-Flash-Prover sets a new state-of-the-art for open-weights models in both auto-formalization and theorem proving. Demonstrating remarkable sample efficiency, it achieves a 97.1% pass rate on MiniF2F-Test using only 72 inference budget per problem. On more challenging benchmarks, it solves 70.8% of ProverBench and 41.5% of PutnamBench with no more than 220 attempts per problem, significantly outperforming existing open-weights baselines.
Read the full abstract

All authors Jianing Wang, Jianfei Zhang, Qi Guo, Linsen Guo, Rumei Li, Chao Zhang, Chong Peng, Cunguang Wang, Dengchang Zhao, Jiarong Shi, Jingang Wang, Liulin Feng, Mengxia Shen, Qi Li, Shengnan An, Shun Wang, Wei Shi, Xiangyu Xi, Xiaoyu Li, Xuezhi Cao, Yi Lu, Yunke Zhao, Zhengyu Chen, Zhimin Lin, Wei Wang, Peng Pei, Xunliang Cai

Computation and Language

Reading Between the Lines: How Electronic Nonverbal Cues shape Emotion Decoding

Taara Kumar and Kokil Jaidka

As text-based computer-mediated communication (CMC) increasingly structures everyday interaction, a central question re-emerges with new urgency: How do users reconstruct nonverbal expression in environments where embodied cues are absent? This paper provides a systematic, theory-driven account of electronic nonverbal cues (eNVCs) - textual analogues of kinesics, vocalics, and paralinguistics - in public microblog communication. Across three complementary studies, we advance conceptual, empirical, and methodological contributions. Study 1 develops a unified taxonomy of eNVCs grounded in foundational nonverbal communication theory and introduces a scalable Python toolkit for their automated detection. Study 2, a within-subject survey experiment, offers controlled causal evidence that eNVCs substantially improve emotional decoding accuracy and lower perceived ambiguity, while also identifying boundary conditions, such as sarcasm, under which these benefits weaken or disappear. Study 3, through focus group discussions, reveals the interpretive strategies users employ when reasoning about digital prosody, including drawing meaning from the absence of expected cues and defaulting toward negative interpretations in ambiguous contexts. Together, these studies establish eNVCs as a coherent and measurable class of digital behaviors, refine theoretical accounts of cue richness and interpretive effort, and provide practical tools for affective computing, user modeling, and emotion-aware interface design. The eNVC detection toolkit is available as a Python and R package at https://github.com/kokiljaidka/envc.
Read the full abstract

Computation and Language

Left Behind: Cross-Lingual Transfer as a Bridge for Low-Resource Languages in Large Language Models

Abdul-Salem Beibitkhan

We investigate how large language models perform on low-resource languages by benchmarking eight LLMs across five experimental conditions in English, Kazakh, and Mongolian. Using 50 hand-crafted questions spanning factual, reasoning, technical, and culturally grounded categories, we evaluate 2,000 responses on accuracy, fluency, and completeness. We find a consistent performance gap of 13.8-16.7 percentage points between English and low-resource language conditions, with models maintaining surface-level fluency while producing significantly less accurate content. Cross-lingual transfer-prompting models to reason in English before translating back-yields selective gains for bilingual architectures (+2.2pp to +4.3pp) but provides no benefit to English-dominant models. Our results demonstrate that current LLMs systematically underserve low-resource language communities, and that effective mitigation strategies are architecture-dependent rather than universal.
Read the full abstract

Computation and Language

Mitigating Selection Bias in Large Language Models via Permutation-Aware GRPO

Jinquan Zheng, Jia Yuan, Jiacheng Yao et al. (+3 more)

Large language models (LLMs) used for multiple-choice and pairwise evaluation tasks often exhibit selection bias due to non-semantic factors like option positions and label symbols. Existing inference-time debiasing is costly and may harm reasoning, while pointwise training ignores that the same question should yield consistent answers across permutations. To address this issue, we propose Permutation-Aware Group Relative Policy Optimization (PA-GRPO), which mitigates selection bias by enforcing permutation-consistent semantic reasoning. PA-GRPO constructs a permutation group for each instance by generating multiple candidate permutations, and optimizes the model using two complementary mechanisms: (1) cross-permutation advantage, which computes advantages relative to the mean reward over all permutations of the same instance, and (2) consistency-aware reward, which encourages the model to produce consistent decisions across different permutations. Experimental results demonstrate that PA-GRPO outperforms strong baselines across seven benchmarks, substantially reducing selection bias while maintaining high overall performance. The code is available on github (https://github.com/ECNU-Text-Computing/PA-GRPO).
Read the full abstract

All authors Jinquan Zheng, Jia Yuan, Jiacheng Yao, Chenyang Gu, Pujun Zheng, Guoxiu He

Computation and Language · Cross-listed from cs.LG

CLT-Forge: A Scalable Library for Cross-Layer Transcoders and Attribution Graphs

Florent Draye, Vedant Palit, Abir Harrasse et al. (+8 more)

Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at: https://github.com/LLM-Interp/CLT-Forge.
Read the full abstract

All authors Florent Draye, Vedant Palit, Abir Harrasse, Tung-Yu Wu, Jiarui Liu, Punya Syon Pandey, Roderick Wu, Chih-Hao Hsu, Terry Jingchen Zhang, Zhijing Jin, Bernhard Schölkopf

Computation and Language · Cross-listed from cs.LG

Structural Sensitivity in Compressed Transformers: Relative Error Propagation and Layer Removal

Abhinaba Basu, Kumkum Basu and Koushik Deb

Compressing transformer weights makes large language models cheaper to deploy. But each layer's compression introduces an error. These errors accumulate as the signal passes through later layers, and how they accumulate is not well understood. We measure this directly: at each layer, we take the ratio of output to input error, calling it rho. A value below one means the layer absorbs the error; above one means it grows. Computing rho on six transformers (117M to 8B parameters) yields three findings. (i) Errors at layer t scale downstream by the product of later rho values, predicting representation drift (Spearman r = -0.44, p < 10^-4). This explains why compressing early layers hurts more than late ones, and why depth-decreasing sparsity schedules outperform uniform ones. Across architecture families, however, model width and redundancy matter more than rho alone. (ii) Within a layer, naive pruning shows a ~600x spread in component sensitivity. Activation-aware pruning (Wanda) shrinks this to 3-7x; the ranking reverses across architectures, so fixed importance scores do not transfer. (iii) For depth pruning, ranking layers by how far rho is from one takes two forward passes. It beats ShortGPT's Block Influence with 1.6x lower perplexity at eight layers removed, and physical deletion delivers 1.22x wall-clock speed-up. A blend of the two criteria does best (perplexity 14.2, 60.0% downstream accuracy on LLaMA-2-7B). Twelve Lean 4 norm inequalities provide machine-checked per-matrix error bounds. The contraction profile thus gives a training-free instrument for two decisions: where to compress within layers, and which to remove.
Read the full abstract

Saturday 21 March

Computation and Language · Cross-listed from cs.CY

Clinical Note Bloat Reduction for Efficient LLM Use

Jordan L. Cahoon, Chloe Stanwyck, Asad Aali et al. (+6 more)

Background: Clinical notes contain extensive duplicated text from templates, copy-paste, and auto-populated fields ("note bloat"), diluting clinical signal, limiting longitudinal context, and increasing large language model (LLM) costs. Methods: TRACE removes note bloat using note-level EHR metadata to identify templated and copied content, with frequency-based de-duplication when metadata are unavailable. We evaluated TRACE using blinded physician span review and gold-standard templated-text annotations across four cohorts spanning liver transplant, obstetrics, and inpatient populations at multiple health systems (5.3M notes). We compared zero-shot LLMs and embedding-based classifiers using original and TRACE-processed notes for 20 information extraction tasks and prediction of 5-year survival, postpartum hemorrhage, and 30-day readmission. Results: Only 0.3-6.6% of removed text was flagged as author-generated; TRACE captured 86% of annotated templated characters. Information extraction F1 differences averaged by cohort ranged from -0.009 to +0.004; task-specific prediction F1 differences ranged from -0.011 to +0.018. Among 1,000 randomly sampled Stanford Health Care patients, TRACE reduced chart text by 47.3% (742.7M characters), averaging 220,167 fewer tokens per patient. Using 2024 encounter volumes at a large tertiary academic center and one query per encounter, projected three-year net savings ranged from $1.00M to $13.58M across evaluated model pricing schemes, including initial and annual TRACE processing costs. Conclusion: TRACE substantially reduces clinical note redundancy while preserving information extraction and prediction performance. Underused EHR metadata can reduce LLM inference costs, expand usable longitudinal context, and support scalable clinical AI.
Read the full abstract

All authors Jordan L. Cahoon, Chloe Stanwyck, Asad Aali, Rachel Madding, Sulaiman S. Somani, Emma Sun, Yixing Jiang, Renumathy Dhanasekaran, Emily Alsentzer

Computation and Language

IWLV-Ramayana: A Sarga-Aligned Parallel Corpus of Valmiki's Ramayana Across Indian Languages

Sumesh VP

The Ramayana is among the most influential literary traditions of South and Southeast Asia, transmitted across numerous linguistic and cultural contexts over two millennia. Despite extensive scholarship on regional Ramayana traditions, computational resources enabling systematic cross-linguistic analysis remain limited. This paper introduces the IWLV Ramayana Corpus, a structured parallel corpus aligning Valmiki's Ramayana across multiple Indian languages at the level of the sarga (chapter). The corpus currently includes complete English and Malayalam layers, with Hindi, Tamil, Kannada, and Telugu layers in active production. The dataset is distributed in structured JSONL format with explicit provenance metadata, enabling applications in comparative literature, corpus linguistics, digital humanities, and multilingual natural language processing. To our knowledge, this is the first sarga-aligned multilingual parallel corpus of the Valmiki Ramayana with explicit provenance metadata and machine-readable format.
Read the full abstract

Computation and Language

Can Large Language Models Reliably Extract Physiology Index Values from Coronary Angiography Reports?

Sofia Morgado, Filipa Valdeira, Niklas Sander et al. (+4 more)

Coronary angiography (CAG) reports contain clinically relevant physiological measurements, yet this information is typically in the form of unstructured natural language, limiting its use in research. We investigate the use of Large Language Models (LLMs) to automatically extract these values, along with their anatomical locations, from Portuguese CAG reports. To our knowledge, this study is the first addressing physiology indexes extraction from a large (1342 reports) corpus of CAG reports, and one of the few focusing on CAG or Portuguese clinical text. We explore local privacy-preserving general-purpose and medical LLMs under different settings. Prompting strategies included zero-shot, few-shot, and few-shot prompting with implausible examples. In addition, we apply constrained generation and introduce a post-processing step based on RegEx. Given the sparsity of measurements, we propose a multi-stage evaluation framework separating format validity, value detection, and value correctness, while accounting for asymmetric clinical error costs. This study demonstrates the potential of LLMs in for extracting physiological indices from Portuguese CAG reports. Non-medical models performed similarly, the best results were obtained with Llama with a zero-shot prompting, while GPT-OSS demonstrated the highest robustness to changes in the prompts. While MedGemma demonstrated similar results to non-medical models, MedLlama's results were out-of-format in the unconstrained setting, and had a significant lower performance in the constrained one. Changes in the prompt techinique and adding a RegEx layer showed no significant improvement across models, while using constrained generation decreased performance, although having the benefit of allowing the usage of specific models that are not able to conform with the templates.
Read the full abstract

All authors Sofia Morgado, Filipa Valdeira, Niklas Sander, Diogo Ferreira, Marta Vilela, Miguel Menezes, Cláudia Soares

Computation and Language

Alignment midtraining for animals

Jasmine Brazilek and Miles Tidmarsh

We investigate the robustness of value alignment via midtraining with synthetic documents, using animal compassion as a value that is both important in its own right and orthogonal to existing alignment efforts. To evaluate compassionate reasoning, we develop and publicly release Animal Norms In Moral Assessment (ANIMA), a 26-question evaluation spanning 13 ethical dimensions, publicly available as a dataset and Inspect evaluation. On ANIMA, training with 3000 documents achieves 77% compared to 40% for instruction-tuning approaches, with generalization to human compassion and no degradation in standard safety benchmarks or capabilities. However, subsequent unrelated instruction-tuning degrades the intervention, with the advantage disappearing after 5000 samples. Our exploratory results suggest document-based value interventions may require explicit preservation strategies to remain effective through typical training pipelines.
Read the full abstract

Computation and Language · Cross-listed from cs.CR

T-MAP: Red-Teaming LLM Agents with Trajectory-aware Evolutionary Search

Hyomin Lee, Sangwoo Park, Yumin Choi et al. (+3 more)

While prior red-teaming efforts have focused on eliciting harmful text outputs from large language models (LLMs), such approaches fail to capture agent-specific vulnerabilities that emerge through multi-step tool execution, particularly in rapidly growing ecosystems such as the Model Context Protocol (MCP). To address this gap, we propose a trajectory-aware evolutionary search method, T-MAP, which leverages execution trajectories to guide the discovery of adversarial prompts. Our approach enables the automatic generation of attacks that not only bypass safety guardrails but also reliably realize harmful objectives through actual tool interactions. Empirical evaluations across diverse MCP environments demonstrate that T-MAP substantially outperforms baselines in attack realization rate (ARR) and remains effective against frontier models, including GPT-5.2, Gemini-3-Pro, Qwen3.5, and GLM-5, thereby revealing previously underexplored vulnerabilities in autonomous LLM agents. Code is available at https://github.com/pwnhyo/T-MAP.
Read the full abstract

All authors Hyomin Lee, Sangwoo Park, Yumin Choi, Sohyun An, Seanie Lee, Sung Ju Hwang

Computation and Language · Cross-listed from cs.LG

Problems with Chinchilla Approach 2: Systematic Biases in IsoFLOP Parabola Fits

Eric Czech, Zhiwei Xu, Yael Elmatad et al. (+2 more)

Chinchilla Approach 2 is among the most widely used methods for fitting neural scaling laws. Its parabolic approximation introduces systematic biases in compute-optimal allocation estimates, even on noise-free synthetic data. Applied to published Llama 3 IsoFLOP data at open frontier compute scales, these biases imply a parameter underallocation corresponding to 6.5% of the $3.8\times10^{25}$ FLOP training budget and \$1.4M (90% CI: \$412K-\$2.9M) in unnecessary compute at 50% H100 MFU. Simulated multimodal model misallocations show even greater opportunity costs due to higher loss surface asymmetry. Three sources of this error are examined: IsoFLOP sampling grid width (Taylor approximation accuracy), uncentered IsoFLOP sampling, and loss surface asymmetry ($\alpha \neq \beta$). Chinchilla Approach 3 largely eliminates these biases but is often regarded as less data-efficient, numerically unstable, prone to local minima, and harder to implement. Each concern is shown to be unfounded or addressable, especially when the partially linear structure of the objective is exploited via Variable Projection, enabling unbiased inference on all five loss surface parameters through a two-dimensional optimization that is well-conditioned, analytically differentiable, and amenable to dense, or even exhaustive, grid search. It may serve as a more convenient replacement for Approach 2 or a more scalable alternative for adaptations of Approach 3 to richer scaling law formulations. See https://github.com/Open-Athena/vpnls for details and https://openathena.ai/scaling-law-analysis for other results from this study.
Read the full abstract

All authors Eric Czech, Zhiwei Xu, Yael Elmatad, Yixin Wang, William Held

Computation and Language

DiscoUQ: Structured Disagreement Analysis for Uncertainty Quantification in LLM Agent Ensembles

Bo Jiang

Multi-agent LLM systems, where multiple prompted instances of a language model independently answer questions, are increasingly used for complex reasoning tasks. However, existing methods for quantifying the uncertainty of their collective outputs rely on shallow voting statistics that discard the rich semantic information in agents' reasoning. We introduce DiscoUQ, a framework that extracts and leverages the structure of inter-agent disagreement – both linguistic properties (evidence overlap, argument strength, divergence depth) and embedding geometry (cluster distances, dispersion, cohesion) – to produce well-calibrated confidence estimates. We propose three methods of increasing complexity: DiscoUQ-LLM (logistic regression on LLM-extracted structure features), DiscoUQ-Embed (logistic regression on embedding geometry), and DiscoUQ-Learn (a neural network combining all features). Evaluated on four diverse benchmarks (StrategyQA, MMLU, TruthfulQA, ARC-Challenge) with a 5-agent system using Qwen3.5-27B, DiscoUQ-LLM achieves an average AUROC of 0.802, outperforming the best baseline (LLM Aggregator, 0.791) while being substantially better calibrated (ECE 0.036 vs. 0.098). The learned features generalize across benchmarks with near-zero performance degradation and provide the largest improvements where they are most needed: in the ambiguous "weak disagreement" tier where simple vote counting fails.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Understanding Contextual Recall in Transformers: How Finetuning Enables In-Context Reasoning over Pretraining Knowledge

Bhavya Vasudeva, Puneesh Deora, Alberto Bietti et al. (+2 more)

Transformer-based language models excel at in-context learning (ICL), where they can adapt to new tasks based on contextual examples, without parameter updates. In a specific form of ICL, which we refer to as contextual recall, models pretrained on open-ended text leverage pairwise examples to recall specific facts in novel prompt formats. We investigate whether contextual recall emerges from pretraining alone, what finetuning is required, and what mechanisms drive the necessary representations. For this, we introduce a controlled synthetic framework where pretraining sequences consist of subject-grammar-attribute tuples, with attribute types tied to grammar statistics. We demonstrate that while such pretraining successfully yields factual knowledge, it is insufficient for contextual recall: models fail to implicitly infer attribute types when the grammar statistics are removed in ICL prompts. However, we show that finetuning on tasks requiring implicit inference, distinct from the ICL evaluation, using a subset of subjects, triggers the emergence of contextual recall across all subjects. This transition is accompanied by the formation of low-dimensional latent encodings of the shared attribute type. For mechanistic insight, we derive a construction for an attention-only transformer that replicates the transition from factual to contextual recall, corroborated by empirical validation.
Read the full abstract

All authors Bhavya Vasudeva, Puneesh Deora, Alberto Bietti, Vatsal Sharan, Christos Thrampoulidis

Computation and Language

Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg et al. (+1 more)

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami's novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields far little extraction, indicating that finetuning on individual authors' works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions ($r \ge 0.90$), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors' works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.
Read the full abstract

All authors Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg, Tuhin Chakrabarty

Computation and Language

User Preference Modeling for Conversational LLM Agents: Weak Rewards from Retrieval-Augmented Interaction

Yuren Hao, Shuhaib Mehri, ChengXiang Zhai et al. (+1 more)

Large language models are increasingly used as personal assistants, yet most lack a persistent user model, forcing users to repeatedly restate preferences across sessions. We propose Vector-Adapted Retrieval Scoring (VARS), a pipeline-agnostic, frozen-backbone framework that represents each user with long-term and short-term vectors in a shared preference space and uses these vectors to bias retrieval scoring over structured preference memory. The vectors are updated online from weak scalar rewards from users' feedback, enabling personalization without per-user fine-tuning. We evaluate on MultiSessionCollab, an online multi-session collaboration benchmark with rich user preference profiles, across math and code tasks. Under frozen backbones, the main benefit of user-aware retrieval is improved interaction efficiency rather than large gains in raw task accuracy: our full VARS agent achieves the strongest overall performance, matches a strong Reflection baseline in task success, and reduces timeout rate and user effort. The learned long-term vectors also align with cross-user preference overlap, while short-term vectors capture session-specific adaptation, supporting the interpretability of the dual-vector design. Code, model, and data are available at https://github.com/YurenHao0426/VARS.
Read the full abstract

All authors Yuren Hao, Shuhaib Mehri, ChengXiang Zhai, Dilek Hakkani-Tür

Computation and Language

The Hidden Puppet Master: Predicting Human Belief Change in Manipulative LLM Dialogues

Jocelyn Shen, Amina Luvsanchultem, Jessica Kim et al. (+6 more)

As users increasingly turn to LLMs for practical and personal advice, they become vulnerable to subtle steering toward hidden incentives misaligned with their own interests. While existing NLP research has benchmarked manipulation detection, these efforts often rely on simulated debates and remain fundamentally decoupled from actual human belief shifts in real-world scenarios. We introduce PUPPET, a theoretical taxonomy and resource that bridges this gap by focusing on the moral direction of hidden incentives in everyday, advice-giving contexts. We provide an evaluation dataset of N=1,035 human-LLM interactions, where we measure users' belief shifts. Our analysis reveals a critical disconnect in current safety paradigms: while models can be trained to detect manipulative strategies, they do not correlate with the magnitude of resulting belief change. As such, we define the task of human belief shift prediction and show that while state-of-the-art LLMs achieve moderate correlation (r=0.3-0.5), they exhibit systematic directional biases, with certain models over or under-predicting the magnitude of human belief change. This work establishes a theoretically grounded and behaviorally validated foundation for AI social safety efforts by studying incentive-driven manipulation in LLMs during everyday, practical user queries.
Read the full abstract

All authors Jocelyn Shen, Amina Luvsanchultem, Jessica Kim, Kynnedy Smith, Valdemar Danry, Kantwon Rogers, Hae Won Park, Maarten Sap, Cynthia Breazeal

Computation and Language

Mitigating Shortcut Reasoning in Language Models: A Gradient-Aware Training Approach

Hongyu Cao, Kunpeng Liu, Dongjie Wang et al. (+1 more)

Large language models exhibit strong reasoning capabilities, yet often rely on shortcuts such as surface pattern matching and answer memorization rather than genuine logical inference. We propose Shortcut-Aware Reasoning Training (SART), a gradient-aware framework that detects and mitigates shortcut-promoting samples via ShortcutScore and gradient surgery. Our method identifies shortcut signals through gradient misalignment with validation objectives and answer-token concentration, and modifies training dynamics accordingly. Experiments on controlled reasoning benchmarks show that SART achieves +16.5% accuracy and +40.2% robustness over the strongest baseline, significantly improving generalization under distribution shifts. Code is available at: https://github.com/fuyanjie/short-cut-aware-data-centric-reasoning.
Read the full abstract

All authors Hongyu Cao, Kunpeng Liu, Dongjie Wang, Yanjie Fu

Computation and Language

LLM Router: Rethinking Routing with Prefill Activations

Tanay Varshney, Annie Surla, Michelle Xu et al. (+5 more)

Existing routers rely on semantic query features or handcrafted features, which often fail to capture model-specific failures or intrinsic task difficulty. We instead route using internal LLM activations, specifically the residual stream. Our key idea, Encoder-Target Decoupling, separates the model that produces the predictive signal (the Encoder) from the model whose correctness is being estimated (the Target), allowing open-weight encoders to predict the performance of closed-source target models. We evaluate layerwise geometric probes, finding that Fisher Separability ($J$) effectively identifies informative layers, supported by Effective Dimensionality ($d_{\mathrm{eff}}$) diagnostics. We then utilize a SharedTrunkNet, a joint multi-output MLP that predicts simultaneous correctness probabilities across candidate models using concatenated prefill features. In our experiments, SharedTrunkNet consistently outperforms semantic baselines. At its best, SharedTrunkNet closes 45.58% of the gap between the strongest standalone model and the oracle while achieving 74.31% cost savings relative to the most expensive model. These results demonstrate that prefill activations provide a robust routing signal, establishing activation-based routing as a high-performance alternative to purely semantic selection.
Read the full abstract

All authors Tanay Varshney, Annie Surla, Michelle Xu, Gomathy Venkata Krishnan, Maximilian Jeblick, David Austin, Neal Vaidya, Davide Onofrio

Computation and Language

MemoNoveltyAgent: A Historical Research Memory-Aware Agent Workflow for Paper Novelty Assessment

Jiajun Hou, Hexuan Deng, Wenxiang Jiao et al. (+4 more)

To alleviate the heavy burden of paper screening, researchers increasingly rely on existing AI agents, such as AI reviewers or DeepResearch, for paper evaluation and novelty assessment. However, lacking specialized mechanisms for processing scholarly literature, their analyses often produce superficial results with noticeable deficiencies in quality. To bridge this gap, we introduce MemoNoveltyAgent, a multi-agent system designed to generate comprehensive and faithful novelty reports. Beyond retrieving concrete prior-paper evidence via RAG, our system incorporates a high-level abstract memory constructed from large-scale scholarly corpora. This memory organizes research into hierarchical trees to distill field-specific evolutionary trajectories, thereby providing a broader historical context. Furthermore, we decompose papers into discrete novelty points for fine-grained analysis and retrieval, while employing a self-validation mechanism to improve report faithfulness. Finally, to address the evaluation challenges of such open-ended generation tasks, we propose a RAG-augmented checklist evaluation method that enables reliable and evidence-grounded assessments. Extensive experiments demonstrate that MemoNoveltyAgent outperforms GPT-5 DeepResearch by 13.69%. Code and demo are available at https://github.com/SStan1/MemoNoveltyAgent
Read the full abstract

All authors Jiajun Hou, Hexuan Deng, Wenxiang Jiao, Xuebo Liu, Xiaopeng Ke, Derek F. Wong, Min Zhang

Computation and Language · Cross-listed from cs.IR

RubricRAG: Towards Interpretable and Reliable LLM Evaluation via Domain Knowledge Retrieval for Rubric Generation

Kaustubh D. Dhole and Eugene Agichtein

Large language models (LLMs) are increasingly evaluated and sometimes trained using automated graders such as LLM-as-judges that output scalar scores or preferences. While convenient, these approaches are often opaque: a single score rarely explains why an answer is good or bad, which requirements were missed, or how a system should be improved. This lack of interpretability limits their usefulness for model development, dataset curation, and high-stakes deployment. Query-specific rubric-based evaluation offers a more transparent alternative by decomposing quality into explicit, checkable criteria. However, manually designing high-quality, query-specific rubrics is labor-intensive and cognitively demanding and not feasible for deployment. While previous approaches have focused on generating intermediate rubrics for automated downstream evaluation, it is unclear if these rubrics are both interpretable and effective for human users. In this work, we investigate whether LLMs can generate useful, instance-specific rubrics as compared to human-authored rubrics, while also improving effectiveness for identifying good responses. Through our systematic study on two rubric benchmarks, and on multiple few-shot and post-training strategies, we find that off-the-shelf LLMs produce rubrics that are poorly aligned with human-authored ones. We introduce a simple strategy, RubricRAG, which retrieves domain knowledge via rubrics at inference time from related queries. We demonstrate that RubricRAG can generate more interpretable rubrics both for similarity to human-authored rubrics, and for improved downstream evaluation effectiveness. Our results highlight both the challenges and a promising approach of scalable, interpretable evaluation through automated rubric generation.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Semantic Sections: An Atlas-Native Feature Ontology for Obstructed Representation Spaces

Hossein Javidnia

Recent interpretability work often treats a feature as a single global direction, dictionary atom, or latent coordinate shared across contexts. We argue that this ontology can fail in obstructed representation spaces, where locally coherent meanings need not assemble into one globally consistent feature. We introduce an atlas-native replacement object, the semantic section: a transport-compatible family of local feature representatives defined over a context atlas. We formalize semantic sections, prove that tree-supported propagation is always pathwise realizable, and show that cycle consistency is the key criterion for genuine globalization. This yields a distinction between tree-local, globalizable, and twisted sections, with twisted sections capturing locally coherent but holonomy-obstructed meanings. We then develop a discovery-and-certification pipeline based on seeded propagation, synchronization across overlaps, defect-based pruning, cycle-aware taxonomy, and deduplication. Across layer-16 atlases for Llama 3.2 3B Instruct, Qwen 2.5 3B Instruct, and Gemma 2 2B IT, we find nontrivial populations of semantic sections, including cycle-supported globalizable and twisted regimes after deduplication. Most importantly, semantic identity is not recovered by raw global-vector similarity. Even certified globalizable sections show low cross-chart signed cosine similarity, and raw similarity baselines recover only a small fraction of true within-section pairs, often collapsing at moderate thresholds. By contrast, section-based identity recovery is perfect on certified supports. These results support semantic sections as a better feature ontology in obstructed regimes.
Read the full abstract

Computation and Language

SozKZ: Training Efficient Small Language Models for Kazakh from Scratch

Saken Tukenov

Kazakh, a Turkic language spoken by over 22 million people, remains underserved by existing multilingual language models, which allocate minimal capacity to low-resource languages and employ tokenizers ill-suited to agglutinative morphology. We present SozKZ, a family of Llama-architecture language models (50M-600M parameters) trained entirely from scratch on 9 billion tokens of Kazakh text with a dedicated 50K BPE tokenizer. We evaluate all models on three Kazakh benchmarks – multiple-choice cultural QA, reading comprehension (Belebele), and topic classification (SIB-200) – alongside five multilingual baselines ranging from 500M to 3B parameters. Our 600M model achieves 30.3% accuracy on Kazakh cultural QA, approaching the 32.0% of Llama-3.2-1B (2x larger), and 25.5% on SIB-200 topic classification, surpassing all evaluated multilingual models up to 2B parameters. We observe consistent scaling from 50M to 600M, with MC QA accuracy rising from 22.8% to 30.3%, suggesting that further scaling remains beneficial. These results demonstrate that small, dedicated models trained from scratch with a language-appropriate tokenizer offer a viable path for low-resource language technology, achieving competitive performance at a fraction of the computational cost. All models and the tokenizer are released under open licenses.
Read the full abstract

Computation and Language

Can ChatGPT Really Understand Modern Chinese Poetry?

Shanshan Wang, Derek F. Wong, Jingming Yao et al. (+1 more)

ChatGPT has demonstrated remarkable capabilities on both poetry generation and translation, yet its ability to truly understand poetry remains unexplored. Previous poetry-related work merely analyzed experimental outcomes without addressing fundamental issues of comprehension. This paper introduces a comprehensive framework for evaluating ChatGPT's understanding of modern poetry. We collaborated with professional poets to evaluate ChatGPT's interpretation of modern Chinese poems by different poets along multiple dimensions. Evaluation results show that ChatGPT's interpretations align with the original poets' intents in over 73% of the cases. However, its understanding in certain dimensions, particularly in capturing poeticity, proved to be less satisfactory. These findings highlight the effectiveness and necessity of our proposed framework. This study not only evaluates ChatGPT's ability to understand modern poetry but also establishes a solid foundation for future research on LLMs and their application to poetry-related tasks.
Read the full abstract

All authors Shanshan Wang, Derek F. Wong, Jingming Yao, Lidia S. Chao

Computation and Language

HiCI: Hierarchical Construction-Integration for Long-Context Attention

Xiangyu Zeng, Qi Xu, Yunke Wang et al. (+1 more)

Long-context language modeling is commonly framed as a scalability challenge of token-level attention, yet local-to-global information structuring remains largely implicit in existing approaches. Drawing on cognitive theories of discourse comprehension, we propose HiCI (Hierarchical Construction–Integration), a hierarchical attention module that constructs segment-level representations, integrates them into a shared global context, and broadcasts both to condition segment-level attention. We validate HiCI through parameter-efficient adaptation of LLaMA-2 with only <5.5% additional parameters, extending context from 4K to 100K tokens (7B) and 64K tokens (13B). Across language modeling, retrieval, and instruction-following benchmarks, HiCI yields consistent improvements over strong baselines, including matching proprietary models on topic retrieval and surpassing GPT-3.5-Turbo-16K on code comprehension. These results demonstrate the effectiveness of explicit hierarchical structuring as an inductive bias for long-context modeling.
Read the full abstract

All authors Xiangyu Zeng, Qi Xu, Yunke Wang, Chang Xu

Computation and Language

BenchBench: Benchmarking Automated Benchmark Generation

Yandan Zheng, Haoran Luo, Zhenghong Lin et al. (+2 more)

Benchmarks are the de facto standard for tracking progress in large language models (LLMs), yet static test sets can rapidly saturate, become vulnerable to contamination, and are costly to refresh. Scalable evaluation of open-ended items often relies on LLM judges, introducing additional sources of bias and prompt sensitivity. We argue that evaluation must extend beyond how well models answer benchmarks to how well models design them. We introduce BenchBench, a three-stage pipeline and dataset for benchmarking automated benchmark generation: (i) extract structured domain cards from seed benchmarks, (ii) prompt multiple designer LLMs to generate quota-controlled suites, and (iii) validate items with a multi-model answerer panel using exact/numeric/symbolic verifiers when possible and rubric-guided judging otherwise, yielding designer–answerer matrices with item-level quality flags and psychometric diagnostics. Across nine variants spanning computer science, mathematics, medicine, and theory-of-mind reasoning (including multilingual and multimodal settings), we generate 16.7K items, retain ~15K core items post-filtering, and produce ~152K graded model–item responses. BenchBench shows that benchmark-design ability is only moderately correlated with answer-time strength (Spearman rho ~0.37), invalidity is negatively associated with discrimination (Pearson r~0.62), and the resulting designer–answerer matrices enable scalable audits of format/modality/language fidelity and suite-dependent self/family interactions. The project is available at: https://github.com/koanatakiyo/BenchBench.
Read the full abstract

All authors Yandan Zheng, Haoran Luo, Zhenghong Lin, Wenjin Liu, Luu Anh Tuan

Computation and Language

RLVR Training of LLMs Does Not Improve Thinking Ability for General QA: Evaluation Method and a Simple Solution

Kaiyuan Li, Jing-Cheng Pang and Yang Yu

Reinforcement learning from verifiable rewards (RLVR) stimulates the thinking processes of large language models (LLMs), substantially enhancing their reasoning abilities on verifiable tasks. It is often assumed that similar gains should transfer to general question answering (GQA), but this assumption has not been thoroughly validated. To assess whether RLVR automatically improves LLM performance on GQA, we propose a Cross-Generation evaluation framework that measures the quality of intermediate reasoning by feeding the generated thinking context into LLMs of varying capabilities. Our evaluation leads to a discouraging finding: the efficacy of the thinking process on GQA tasks is markedly lower than on verifiable tasks, suggesting that explicit training on GQA remains necessary in addition to training on verifiable tasks. We further observe that direct RL training on GQA is less effective than RLVR. Our hypothesis is that, whereas verifiable tasks demand robust logical chains to obtain high rewards, GQA tasks often admit shortcuts to high rewards without cultivating high-quality thinking. To avoid possible shortcuts, we introduce a simple method, Separated Thinking And Response Training (START), which first trains only the thinking process, using rewards defined on the final answer. We show that START improves both the quality of thinking and the final answer across several GQA benchmarks and RL algorithms.
Read the full abstract

Computation and Language

The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing

Yuan Cao, Mingyang Wang and Hinrich Schütze

Large language models (LLMs) are increasingly used as knowledge bases, but keeping them up to date requires targeted knowledge editing (KE). However, it remains unclear how edits are implemented inside the model once applied. In this work, we take a mechanistic view of KE using neuron-level knowledge attribution (NLKA). Unlike prior work that focuses on pre-edit causal tracing and localization, we use post-edit attribution – contrasting successful and failed edits – to isolate the computations that shift when an edit succeeds. Across representative KE methods, we find a consistent pattern: mid-to-late attention predominantly promotes the new target, while attention and FFN modules cooperate to suppress the original fact. Motivated by these findings, we propose MEGA, a MEchanism-Guided Activation steering method that performs attention-residual interventions in attribution-aligned regions without modifying model weights. On CounterFact and Popular, MEGA achieves strong editing performance across KE metrics on GPT2-XL and LLaMA2-7B. Overall, our results elevate post-edit attribution from analysis to engineering signal: by pinpointing where and how edits take hold, it powers MEGA to deliver reliable, architecture-agnostic knowledge edits.
Read the full abstract

Computation and Language

Code-MIE: A Code-style Model for Multimodal Information Extraction with Scene Graph and Entity Attribute Knowledge Enhancement

Jiang Liu, Ge Qiu, Hao Fei et al. (+5 more)

With the rapid development of large language models (LLMs), more and more researchers have paid attention to information extraction based on LLMs. However, there are still some spaces to improve in the existing related methods. First, existing multimodal information extraction (MIE) methods usually employ natural language templates as the input and output of LLMs, which mismatch with the characteristics of information tasks that mostly include structured information such as entities and relations. Second, although a few methods have adopted structured and more IE-friendly code-style templates, they just explored their methods on text-only IE rather than multimodal IE. Moreover, their methods are more complex in design, requiring separate templates to be designed for each task. In this paper, we propose a Code-style Multimodal Information Extraction framework (Code-MIE) which formalizes MIE as unified code understanding and generation. Code-MIE has the following novel designs: (1) Entity attributes such as gender, affiliation are extracted from the text to guide the model to understand the context and role of entities. (2) Images are converted into scene graphs and visual features to incorporate rich visual information into the model. (3) The input template is constructed as a Python function, where entity attributes, scene graphs and raw text compose of the function parameters. In contrast, the output template is formalized as Python dictionaries containing all extraction results such as entities, relations, etc. To evaluate Code-MIE, we conducted extensive experiments on the M$^3$D, Twitter-15, Twitter-17, and MNRE datasets. The results show that our method achieves state-of-the-art performance compared to six competing baseline models, with 61.03% and 60.49% on the English and Chinese datasets of M$^3$D, and 76.04%, 88.07%, and 73.94% on the other three datasets.
Read the full abstract

All authors Jiang Liu, Ge Qiu, Hao Fei, Dongdong Xie, Jinbo Li, Fei Li, Chong Teng, Donghong Ji