Peer Revenue Past citations are no guarantee of future returns

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Wednesday 11 February

Computation and Language

Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents

Yifei Li, Weidong Guo, Lingling Zhang et al. (+6 more)

Long-term conversational memory is a core capability for LLM-based dialogue systems, yet existing benchmarks and evaluation protocols primarily focus on surface-level factual recall. In realistic interactions, appropriate responses often depend on implicit constraints such as user state, goals, or values that are not explicitly queried later. To evaluate this setting, we introduce LoCoMo-Plus, a benchmark for assessing cognitive memory under cue–trigger semantic disconnect, where models must retain and apply latent constraints across long conversational contexts. We further show that conventional string-matching metrics and explicit task-type prompting are misaligned with such scenarios, and propose a unified evaluation framework based on constraint consistency. Experiments across diverse backbone models, retrieval-based methods, and memory systems demonstrate that cognitive memory remains challenging and reveals failures not captured by existing benchmarks. Our code and evaluation framework are publicly available at: https://github.com/xjtuleeyf/Locomo-Plus.
Read the full abstract

All authors Yifei Li, Weidong Guo, Lingling Zhang, Rongman Xu, Muye Huang, Hui Liu, Lijiao Xu, Yu Xu, Jun Liu

Computation and Language

Targeted Syntactic Evaluation of Language Models on Georgian Case Alignment

Daniel Gallagher and Gerhard Heyer

This paper evaluates the performance of transformer-based language models on split-ergative case alignment in Georgian, a particularly rare system for assigning grammatical cases to mark argument roles. We focus on subject and object marking determined through various permutations of nominative, ergative, and dative noun forms. A treebank-based approach for the generation of minimal pairs using the Grew query language is implemented. We create a dataset of 370 syntactic tests made up of seven tasks containing 50-70 samples each, where three noun forms are tested in any given sample. Five encoder- and two decoder-only models are evaluated with word- and/or sentence-level accuracy metrics. Regardless of the specific syntactic makeup, models performed worst in assigning the ergative case correctly and strongest in assigning the nominative case correctly. Performance correlated with the overall frequency distribution of the three forms (NOM > DAT > ERG). Though data scarcity is a known issue for low-resource languages, we show that the highly specific role of the ergative along with a lack of available training data likely contributes to poor performance on this case. The dataset is made publicly available and the methodology provides an interesting avenue for future syntactic evaluations of languages where benchmarks are limited.
Read the full abstract

Computation and Language

The Lexical Imprint of Pre-training Data

Jeonghoon Kim, Woojin Chung, Woomin Song et al. (+3 more)

How much of zero-shot benchmark accuracy is captured by word-level alignment between a pre-training corpus and the benchmark target? We study this question through word-level unigram cross-entropy (UCE), a tokenizer-agnostic, model-free measure of lexical alignment between a reference corpus and a target benchmark. Likelihood training shapes the model's predictive distribution from the pre-training corpus, and zero-shot scores under likelihood-based evaluation depend in part on the lexical alignment between that corpus and the benchmark target. Cleanly tracing how the lexical imprint of pre-training data scales with corpus–benchmark alignment requires controlled from-scratch pre-training that varies corpus identity while holding the training recipe fixed. Across this controlled sweep, spanning $10$ zero-shot benchmarks, $4$ pre-training corpora, and $3$ model scales, we find that corpus–benchmark lexical alignment, UCE, consistently tracks zero-shot accuracy on every benchmark, with corpora more aligned with the benchmark (lower UCE) achieving higher accuracy. The same signal extends from measurement to control at finer scopes. At the document level, strengthening the imprint by selecting training documents to maximize lexical alignment with the benchmark produces the accuracy gains predicted by the mechanism. At the prompt level, matching the imprint per prompt by routing each prompt to the drafter whose domain corpus is most lexically aligned with it produces measurable throughput gains over a generalist-drafter baseline. Taken together, pre-training corpora leave a measurable lexical imprint on base-model zero-shot behavior, and UCE provides a model-free probe of lexical alignment between reference and target data.
Read the full abstract

All authors Jeonghoon Kim, Woojin Chung, Woomin Song, Cheonbok Park, Nako Sung, Jinwoo Shin

Computation and Language

UMEM: Unified Memory Extraction and Management Framework for Generalizable Memory

Yongshi Ye, Hui Jiang, Feihu Jiang et al. (+7 more)

Self-evolving memory serves as the trainable parameters for Large Language Models (LLMs)-based agents, where extraction (distilling insights from experience) and management (updating the memory bank) must be tightly coordinated. Existing methods predominately optimize memory management while treating memory extraction as a static process, resulting in poor generalization, where agents accumulate instance-specific noise rather than robust memories. To address this, we propose Unified Memory Extraction and Management (UMEM), a self-evolving agent framework that jointly optimizes a Large Language Model to simultaneous extract and manage memories. To mitigate overfitting to specific instances, we introduce Semantic Neighborhood Modeling and optimize the model with a neighborhood-level marginal utility reward via GRPO. This approach ensures memory generalizability by evaluating memory utility across clusters of semantically related queries. Extensive experiments across five benchmarks demonstrate that UMEM significantly outperforms highly competitive baselines, achieving up to a 10.67% improvement in multi-turn interactive tasks. Futhermore, UMEM maintains a monotonic growth curve during continuous evolution. Codes and models will be publicly released.
Read the full abstract

All authors Yongshi Ye, Hui Jiang, Feihu Jiang, Tian Lan, Yichao Du, Biao Fu, Xiaodong Shi, Qianghuai Jia, Longyue Wang, Weihua Luo

Computation and Language · Cross-listed from cs.AI

To Think or Not To Think, That is The Question for Large Reasoning Models in Theory of Mind Tasks

Nanxu Gong, Haotian Li, Sixun Dong et al. (+3 more)

Theory of Mind (ToM) assesses whether models can infer hidden mental states such as beliefs, desires, and intentions, which is essential for natural social interaction. Although recent progress in Large Reasoning Models (LRMs) has boosted step-by-step inference in mathematics and coding, it is still underexplored whether this benefit transfers to socio-cognitive skills. We present a systematic study of nine advanced Large Language Models (LLMs), comparing reasoning models with non-reasoning models on three representative ToM benchmarks. The results show that reasoning models do not consistently outperform non-reasoning models and sometimes perform worse. A fine-grained analysis reveals three insights. First, slow thinking collapses: accuracy significantly drops as responses grow longer, and larger reasoning budgets hurt performance. Second, moderate and adaptive reasoning benefits performance: constraining reasoning length mitigates failure, while distinct success patterns demonstrate the necessity of dynamic adaptation. Third, option matching shortcut: when multiple choice options are removed, reasoning models improve markedly, indicating reliance on option matching rather than genuine deduction. We also design two intervention approaches: Slow-to-Fast (S2F) adaptive reasoning and Think-to-Match (T2M) shortcut prevention to further verify and mitigate the problems. With all results, our study highlights the advancement of LRMs in formal reasoning (e.g., math, code) cannot be fully transferred to ToM, a typical task in social reasoning. We conclude that achieving robust ToM requires developing unique capabilities beyond existing reasoning methods.
Read the full abstract

All authors Nanxu Gong, Haotian Li, Sixun Dong, Jianxun Lian, Yanjie Fu, Xing Xie

Computation and Language

How Do Decoder-Only LLMs Perceive Users? Rethinking Attention Masking for User Representation Learning

Jiahao Yuan, Yike Xu, Jinyong Wen et al. (+8 more)

Decoder-only large language models are increasingly used as behavioral encoders for user representation learning, yet the impact of attention masking on the quality of user embeddings remains underexplored. In this work, we conduct a systematic study of causal, hybrid, and bidirectional attention masks within a unified contrastive learning framework trained on large-scale real-world Alipay data that integrates long-horizon heterogeneous user behaviors. To improve training dynamics when transitioning from causal to bidirectional attention, we propose Gradient-Guided Soft Masking, a gradient-based pre-warmup applied before a linear scheduler that gradually opens future attention during optimization. Evaluated on 9 industrial user cognition benchmarks covering prediction, preference, and marketing sensitivity tasks, our approach consistently yields more stable training and higher-quality bidirectional representations compared with causal, hybrid, and scheduler-only baselines, while remaining compatible with decoder pretraining. Overall, our findings highlight the importance of masking design and training transition in adapting decoder-only LLMs for effective user representation learning. Our code is available at https://github.com/JhCircle/Deepfind-GGSM.
Read the full abstract

All authors Jiahao Yuan, Yike Xu, Jinyong Wen, Baokun Wang, Yang Chen, Xiaotong Lin, Wuliang Huang, Ziyi Gao, Xing Fu, Yu Cheng, Weiqiang Wang

Computation and Language · Cross-listed from cs.SE

ISD-Agent-Bench: A Comprehensive Benchmark for Evaluating LLM-based Instructional Design Agents

YoungHoon Jeon, Suwan Kim, Haein Son et al. (+3 more)

Large Language Model (LLM) agents have shown promising potential in automating Instructional Systems Design (ISD), a systematic approach to developing educational programs. However, evaluating these agents remains challenging due to the lack of standardized benchmarks and the risk of LLM-as-judge bias. We present ISD-Agent-Bench, a comprehensive benchmark comprising 25,795 scenarios generated via a Context Matrix framework that combines 51 contextual variables across 5 categories with 33 ISD sub-steps derived from the ADDIE model. To ensure evaluation reliability, we employ a multi-judge protocol using diverse LLMs from different providers, achieving high inter-judge reliability. We compare existing ISD agents with novel agents grounded in classical ISD theories such as ADDIE, Dick & Carey, and Rapid Prototyping ISD. Experiments on 1,017 test scenarios demonstrate that integrating classical ISD frameworks with modern ReAct-style reasoning achieves the highest performance, outperforming both pure theory-based agents and technique-only approaches. Further analysis reveals that theoretical quality strongly correlates with benchmark performance, with theory-based agents showing significant advantages in problem-centered design and objective-assessment alignment. Our work provides a foundation for systematic LLM-based ISD research.
Read the full abstract

All authors YoungHoon Jeon, Suwan Kim, Haein Son, Sookbun Lee, Yeil Jeong, Unggi Lee

Computation and Language

Online Causal Kalman Filtering for Stable and Effective Policy Optimization

Shuo He, Lang Feng, Xin Cheng et al. (+2 more)

Reinforcement learning for large language models suffers from high-variance token-level importance sampling (IS) ratios, which would destabilize policy optimization at scale. To improve stability, recent methods typically use a fixed sequence-level IS ratio for all tokens in a sequence or adjust each token's IS ratio separately, thereby neglecting temporal off-policy derivation across tokens in a sequence. In this paper, we first empirically identify that local off-policy deviation is structurally inconsistent at the token level, which may distort policy-gradient updates across adjacent tokens and lead to training collapse. To address the issue, we propose Online Causal Kalman Filtering for stable and effective Policy Optimization (KPO). Concretely, we model the desired IS ratio as a latent state that evolves across tokens and apply a Kalman filter to update this state online and autoregressively based on the states of past tokens, regardless of future tokens. The resulting filtered IS ratios preserve token-wise local structure-aware variation while strongly smoothing noise spikes, yielding more stable and effective policy updates. Experimentally, KPO achieves superior results on challenging math reasoning datasets compared with state-of-the-art counterparts.
Read the full abstract

All authors Shuo He, Lang Feng, Xin Cheng, Lei Feng, Bo An

Computation and Language

Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters

Ailin Huang, Ang Li, Aobo Kong et al. (+213 more)

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.
Read the full abstract

All authors Ailin Huang, Ang Li, Aobo Kong, Bin Wang, Binxing Jiao, Bo Dong, Bojun Wang, Boyu Chen, Brian Li, Buyun Ma, Chang Su, Changxin Miao, Changyi Wan, Chao Lou, Chen Hu, Chen Xu, Chenfeng Yu, Chengting Feng, Chengyuan Yao, Chunrui Han, Dan Ma, Dapeng Shi, Daxin Jiang, Dehua Ma, Deshan Sun, Di Qi, Enle Liu, Fajie Zhang, Fanqi Wan, Guanzhe Huang, Gulin Yan, Guoliang Cao, Guopeng Li, Han Cheng, Hangyu Guo, Hanshan Zhang, Hao Nie, Haonan Jia, Haoran Lv, Hebin Zhou, Hekun Lv, Heng Wang, Heung-Yeung Shum, Hongbo Huang, Hongbo Peng, Hongyu Zhou, Hongyuan Wang, Houyong Chen, Huangxi Zhu, Huimin Wu, Huiyong Guo, Jia Wang, Jian Zhou, Jianjian Sun, Jiaoren Wu, Jiaran Zhang, Jiashu Lv, Jiashuo Liu, Jiayi Fu, Jiayu Liu, Jie Cheng, Jie Luo, Jie Yang, Jie Zhou, Jieyi Hou, Jing Bai, Jingcheng Hu, Jingjing Xie, Jingwei Wu, Jingyang Zhang, Jishi Zhou, Junfeng Liu, Junzhe Lin, Ka Man Lo, Kai Liang, Kaibo Liu, Kaijun Tan, Kaiwen Yan, Kaixiang Li, Kang An, Kangheng Lin, Lei Yang, Liang Lv, Liang Zhao, Liangyu Chen, Lieyu Shi, Liguo Tan, Lin Lin, Lina Chen, Luck Ma, Mengqiang Ren, Michael Li, Ming Li, Mingliang Li, Mingming Zhang, Mingrui Chen, Mitt Huang, Na Wang, Peng Liu, Qi Han, Qian Zhao, Qinglin He, Qinxin Du, Qiuping Wu, Quan Sun, Rongqiu Yang, Ruihang Miao, Ruixin Han, Ruosi Wan, Ruyan Guo, Shan Wang, Shaoliang Pang, Shaowen Yang, Shengjie Fan, Shijie Shang, Shiliang Yang, Shiwei Li, Shuangshuang Tian, Siqi Liu, Siye Wu, Siyu Chen, Song Yuan, Tiancheng Cao, Tianchi Yue, Tianhao Cheng, Tianning Li, Tingdan Luo, Wang You, Wei Ji, Wei Yuan, Wei Zhang, Weibo Wu, Weihao Xie, Wen Sun, Wenjin Deng, Wenzhen Zheng, Wuxun Xie, Xiangfeng Wang, Xiangwen Kong, Xiangyu Liu, Xiangyu Zhang, Xiaobo Yang, Xiaojia Liu, Xiaolan Yuan, Xiaoran Jiao, Xiaoxiao Ren, Xiaoyun Zhang, Xin Li, Xin Liu, Xin Wu, Xing Chen, Xingping Yang, Xinran Wang, Xu Zhao, Xuan He, Xuanti Feng, Xuedan Cai, Xuqiang Zhou, Yanbo Yu, Yang Li, Yang Xu, Yanlin Lai, Yanming Xu, Yaoyu Wang, Yeqing Shen, Yibo Zhu, Yichen Lv, Yicheng Cao, Yifeng Gong, Yijing Yang, Yikun Yang, Yin Zhao, Yingxiu Zhao, Yinmin Zhang, Yitong Zhang, Yixuan Zhang, Yiyang Chen, Yongchi Zhao, Yongshen Long, Yongyao Wang, Yousong Guan, Yu Zhou, Yuang Peng, Yuanhao Ding, Yuantao Fan, Yuanwei Lu, Yuanzhen Yang, Yuchu Luo, Yudi Zhao, Yue Peng, Yueqiang Lin, Yufan Lu, Yuling Zhao, Yunzhou Ju, Yurong Zhang, Yusheng Li, Yuxiang Yang, Yuyang Chen, Yuzhu Cai, Zejia Weng, Zetao Hong, Zexi Li, Zhe Xie, Zheng Ge, Zheng Gong, Zheng Zeng, Zhenyi Lu, Zhewei Huang, Zhichao Chang, Zhiguo Huang, Zhiheng Hu, Zidong Yang, Zili Wang, Ziqi Ren, Zixin Zhang, Zixuan Wang

Computation and Language

When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning

Leheng Sheng, Yongtao Zhang, Wenchang Ma et al. (+6 more)

While reasoning over long context is crucial for various real-world applications, it remains challenging for large language models (LLMs) as they suffer from performance degradation as the context length grows. Recent work MemAgent has tried to tackle this by processing context chunk-by-chunk in an RNN-like loop and updating a textual memory for final answering. However, this naive recurrent memory update faces two crucial drawbacks: (i) memory can quickly explode because it can update indiscriminately, even on evidence-free chunks; and (ii) the loop lacks an exit mechanism, leading to unnecessary computation after even sufficient evidence is collected. To address these issues, we propose GRU-Mem, which incorporates two text-controlled gates for more stable and efficient long-context reasoning. Specifically, in GRU-Mem, the memory only updates when the update gate is open and the recurrent loop will exit immediately once the exit gate is open. To endow the model with such capabilities, we introduce two reward signals $r^{\text{update}}$ and $r^{\text{exit}}$ within end-to-end RL, rewarding the correct updating and exiting behaviors respectively. Experiments on various long-context reasoning tasks demonstrate the effectiveness and efficiency of GRU-Mem, which generally outperforms the vanilla MemAgent with up to 400% times inference speed acceleration.
Read the full abstract

All authors Leheng Sheng, Yongtao Zhang, Wenchang Ma, Yaorui Shi, Ting Huang, Xiang Wang, An Zhang, Ke Shen, Tat-Seng Chua

Computation and Language

LHAW: Controllable Underspecification for Long-Horizon Tasks

George Pu, Michael S. Lee, Udari Madhushani Sehwag et al. (+6 more)

Long-horizon workflow agents that operate effectively over extended periods are essential for truly autonomous systems. Their reliable execution critically depends on the ability to reason through ambiguous situations in which clarification seeking is necessary to ensure correct task execution. However, progress is limited by the lack of scalable, task-agnostic frameworks for systematically curating and measuring the impact of ambiguity across custom workflows. We address this gap by introducing LHAW (Long-Horizon Augmented Workflows), a modular, dataset-agnostic synthetic pipeline that transforms any well-specified task into controllable underspecified variants by systematically removing information across four dimensions - Goals, Constraints, Inputs, and Context - at configurable severity levels. Unlike approaches that rely on LLM predictions of ambiguity, LHAW validates variants through empirical agent trials, classifying them as outcome-critical, divergent, or benign based on observed terminal state divergence. We release 285 task variants from TheAgentCompany, SWE-Bench Pro and MCP-Atlas according to our taxonomy alongside formal analysis measuring how current agents detect, reason about, and resolve underspecification across ambiguous settings. LHAW provides the first systematic framework for cost-sensitive evaluation of agent clarification behavior in long-horizon settings, enabling development of reliable autonomous systems.
Read the full abstract

All authors George Pu, Michael S. Lee, Udari Madhushani Sehwag, David J. Lee, Bryan Zhu, Yash Maurya, Mohit Raghavendra, Yuan Xue, Samuel Marc Denton

Computation and Language

On the Robustness of Knowledge Editing for Detoxification

Ming Dong, Shiyi Tang, Ziyan Peng et al. (+2 more)

Knowledge-Editing-based (KE-based) detoxification has emerged as a promising approach for mitigating harmful behaviours in Large Language Models. Existing evaluations, however, largely rely on automatic toxicity classifiers, implicitly assuming that reduced toxicity scores reflect genuine behavioural suppression. In this work, we propose a robustness-oriented evaluation framework for KE-based detoxification that examines its reliability beyond standard classifier-based metrics along three dimensions: optimisation robustness, compositional robustness, and cross-lingual robustness. We identify pseudo-detoxification as a common failure mode, where apparent toxicity reductions arise from degenerate generation behaviours rather than meaningful suppression of unsafe content. We further show that detoxification effectiveness degrades when multiple unsafe behaviours are edited jointly, and that both monolingual and cross-lingual detoxification remain effective only under specific model-method combinations. Overall, our results indicate that KE-based detoxification is robust only for certain models, limited numbers of detoxification objectives, and a subset of languages.
Read the full abstract

All authors Ming Dong, Shiyi Tang, Ziyan Peng, Guanyi Chen, Tingting He

Computation and Language

Canvas-of-Thought: Grounding Reasoning via Mutable Structured States

Lingzhuang Sun, Yuxia Zhu, Ruitong Liu et al. (+10 more)

While Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), relying solely on linear text sequences remains a bottleneck for complex tasks. We observe that even when auxiliary visual elements are interleaved, they are often treated as static snapshots within a one-dimensional, unstructured reasoning chain. We argue that such approaches treat reasoning history as an immutable stream: correcting a local error necessitates either generating verbose downstream corrections or regenerating the entire context. This forces the model to implicitly maintain and track state updates, significantly increasing token consumption and cognitive load. This limitation is particularly acute in high-dimensional domains, such as geometry and SVG design, where the textual expression of CoT lacks explicit visual guidance, further constraining the model's reasoning precision. To bridge this gap, we introduce Canvas-of-Thought (Canvas-CoT). By leveraging a HTML Canvas as an external reasoning substrate, Canvas-CoT empowers the model to perform atomic, DOM-based CRUD operations. This architecture enables in-place state revisions without disrupting the surrounding context, allowing the model to explicitly maintain the "ground truth". Furthermore, we integrate a rendering-based critique loop that serves as a hard constraint validator, providing explicit visual feedback to resolve complex tasks that are difficult to articulate through text alone. Extensive experiments on VCode, RBench-V, and MathVista demonstrate that Canvas-CoT significantly outperforms existing baselines, establishing a new paradigm for context-efficient multimodal reasoning.
Read the full abstract

All authors Lingzhuang Sun, Yuxia Zhu, Ruitong Liu, Hao Liang, Zheng Sun, Caijun Jia, Honghao He, Yuchen Wu, Siyuan Li, Jingxuan Wei, Xiangxiang Zhang, Bihui Yu, Wentao Zhang

Computation and Language

Neuro-Symbolic Synergy for World Modeling

Hongyu Zhao, Siyu Zhou, Haolin Yang et al. (+2 more)

Large language models (LLMs) exhibit strong general-purpose reasoning capabilities, yet they frequently hallucinate when used as world models (WMs), where strict compliance with deterministic transition rules–particularly in corner cases–is essential. In contrast, Symbolic WMs provide logical consistency but lack semantic expressivity. To bridge this gap, we propose Neuro-Symbolic Synergy (NeSyS), a framework that integrates the probabilistic semantic priors of LLMs with executable symbolic rules to achieve both expressivity and robustness. NeSyS alternates training between the two models using trajectories inadequately explained by the other. Unlike rule-based prompting, the symbolic WM contributes candidate-level scores through log-linear reranking, without requiring the LLM to interpret rule text. Rule-guided sampling prioritizes transitions that are weakly covered by symbolic rules, using 35–60% of the training pairs while outperforming full-data supervised fine-tuning in five of six settings. Experiments on ScienceWorld, WebShop, and PlanCraft demonstrate consistent gains in WM prediction accuracy and data efficiency; one-step lookahead on open-ended WebShop also improves agent reward. Our models, rules, and code are available at https://github.com/tianyi-lab/NeSyS.
Read the full abstract

All authors Hongyu Zhao, Siyu Zhou, Haolin Yang, Zengyi Qin, Tianyi Zhou

Computation and Language · Cross-listed from cs.SE

TestExplora: Benchmarking LLMs for Proactive Bug Discovery via Repository-Level Test Generation

Steven Liu, Jane Luo, Xin Zhang et al. (+7 more)

Given that Large Language Models (LLMs) are increasingly applied to automate software development, comprehensive software assurance spans three distinct goals: regression prevention, reactive reproduction, and proactive discovery. Current evaluations systematically overlook the third goal. Specifically, they either treat existing code as ground truth (a compliance trap) for regression prevention, or depend on post-failure artifacts (e.g., issue reports) for bug reproduction-so they rarely surface defects before failures. To bridge this gap, we present TestExplora, a benchmark designed to evaluate LLMs as proactive testers within full-scale, realistic repository environments. TestExplora contains 2,389 tasks from 482 repositories and hides all defect-related signals. Models must proactively find bugs by comparing implementations against documentation-derived intent, using documentation as the oracle. Furthermore, to keep evaluation sustainable and reduce leakage, we propose continuous, time-aware data collection. Our evaluation reveals a significant capability gap: state-of-the-art models achieve a maximum Fail-to-Pass (F2P) rate of only 16.06%. Further analysis indicates that navigating complex cross-module interactions and leveraging agentic exploration are critical to advancing LLMs toward autonomous software quality assurance. Consistent with this, SWEAgent instantiated with GPT-5-mini achieves an F2P of 17.27% and an F2P@5 of 29.7%, highlighting the effectiveness and promise of agentic exploration in proactive bug discovery tasks.
Read the full abstract

All authors Steven Liu, Jane Luo, Xin Zhang, Aofan Liu, Hao Liu, Jie Wu, Ziyang Huang, Yangyu Huang, Yu Kang, Scarlett Li

Computation and Language

LATA: A Tool for LLM-Assisted Translation Annotation

Baorong Huang and Ali Asiri

The construction of high-quality parallel corpora for translation research has increasingly evolved from simple sentence alignment to complex, multi-layered annotation tasks. This methodological shift presents significant challenges for structurally divergent language pairs, such as Arabic–English, where standard automated tools frequently fail to capture deep linguistic shifts or semantic nuances. This paper introduces a novel, LLM-assisted interactive tool designed to reduce the gap between scalable automation and the rigorous precision required for expert human judgment. Unlike traditional statistical aligners, our system employs a template-based Prompt Manager that leverages large language models (LLMs) for sentence segmentation and alignment under strict JSON output constraints. In this tool, automated preprocessing integrates into a human-in-the-loop workflow, allowing researchers to refine alignments and apply custom translation technique annotations through a stand-off architecture. By leveraging LLM-assisted processing, the tool balances annotation efficiency with the linguistic precision required to analyze complex translation phenomena in specialized domains.
Read the full abstract

Computation and Language · Cross-listed from cs.CR

The Landscape of Prompt Injection Threats in LLM Agents: From Taxonomy to Analysis

Peiran Wang, Xinfeng Li, Chong Xiang et al. (+5 more)

The evolution of Large Language Models (LLMs) has resulted in a paradigm shift towards autonomous agents, necessitating robust security against Prompt Injection (PI) vulnerabilities where untrusted inputs hijack agent behaviors. This SoK presents a comprehensive overview of the PI landscape, covering attacks, defenses, and their evaluation practices. Through a systematic literature review and quantitative analysis, we establish taxonomies that categorize PI attacks by payload generation strategies (heuristic vs. optimization) and defenses by intervention stages (text, model, and execution levels). Our analysis reveals a key limitation shared by many existing defenses and benchmarks: they largely overlook context-dependent tasks, in which agents are authorized to rely on runtime environmental observations to determine actions. To address this gap, we introduce AgentPI, a new benchmark designed to systematically evaluate agent behavior under context-dependent interaction settings. Using AgentPI, we empirically evaluate representative defenses and show that no single approach can simultaneously achieve high trustworthiness, high utility, and low latency. Moreover, we show that many defenses appear effective under existing benchmarks by suppressing contextual inputs, yet fail to generalize to realistic agent settings where context-dependent reasoning is essential. This SoK distills key takeaways and open research problems, offering structured guidance for future research and practical deployment of secure LLM agents.
Read the full abstract

All authors Peiran Wang, Xinfeng Li, Chong Xiang, Jinghuai Zhang, Ying Li, Lixia Zhang, Xiaofeng Wang, Yuan Tian

Computation and Language · Cross-listed from cs.LG

Control Reinforcement Learning: Interpretable Token-Level Steering of LLMs via Sparse Autoencoder Features

Seonglae Cho, Zekun Wu and Adriano Koshiyama

Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified. We introduce Control Reinforcement Learning (CRL), which trains a policy to select SAE features for steering at each token, producing interpretable intervention logs: the learned policy identifies features that change model outputs when amplified. Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability. The framework yields new analysis capabilities: branch point tracking locates tokens where feature choice determines output correctness; critic trajectory analysis separates policy limitations from value estimation errors; layer-wise comparison reveals syntactic features in early layers and semantic features in later layers. On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs. These results establish learned feature steering as a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes
Read the full abstract

Computation and Language · Cross-listed from cs.LG

AI-rithmetic

Alex Bie, Travis Dick, Alex Kulesza et al. (+3 more)

Modern AI systems have been successfully deployed to win medals at international math competitions, assist with research workflows, and prove novel technical lemmas. However, despite their progress at advanced levels of mathematics, they remain stubbornly bad at basic arithmetic, consistently failing on the simple task of adding two numbers. We present a systematic investigation of this phenomenon. We demonstrate empirically that all frontier models suffer significantly degraded accuracy for integer addition as the number of digits increases. Furthermore, we show that most errors made by these models are highly interpretable and can be attributed to either operand misalignment or a failure to correctly carry; these two error classes explain 87.9%, 62.9%, and 92.4% of Claude Opus 4.1, GPT-5, and Gemini 2.5 Pro errors, respectively. Finally, we show that misalignment errors are frequently related to tokenization, and that carrying errors appear largely as independent random failures.
Read the full abstract

All authors Alex Bie, Travis Dick, Alex Kulesza, Prabhakar Raghavan, Vinod Raman, Sergei Vassilvitskii

Computation and Language

EVOKE: Emotion Vocabulary Of Korean and English

Yoonwon Jung, Hagyeong Shin and Benjamin K. Bergen

This paper introduces EVOKE (Emotion Vocabulary of Korean and English), a Korean-English parallel dataset of emotion words. The dataset offers comprehensive coverage of emotion words in each language, in addition to many-to-many translations between words in the two languages and identification of language-specific emotion words. The dataset contains 1,426 Korean words and 1,397 English words, and we systematically annotate 819 Korean and 924 English adjectives and verbs. We also annotate multiple meanings of each word and their relationships, identifying polysemous emotion words and emotion-related metaphors. The dataset is, to our knowledge, the most systematic and theory-agnostic dataset of emotion words in both Korean and English to date. It can serve as a practical tool for emotion science, psycholinguistics, computational linguistics, and natural language processing, allowing researchers to adopt different views on the resource reflecting their needs and theoretical perspectives. The dataset is publicly available at https://github.com/yoonwonj/EVOKE.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Gated Normalization Removal and Scale Anchoring in Pre-Norm Transformers

Andrei Kanavalau, Carmen Amo Alonso and Sanjay Lall

Normalization layers are standard in transformers, but it is not clear whether their sample-dependent computations are necessary throughout both training and inference. This work develops a gated normalization-removal approach for pre-norm transformers. The approach is implemented using TaperNorm, which starts from standard RMSNorm/LayerNorm and gradually tapers to learned sample-independent linear or affine maps. Once the gate reaches zero, per-token statistics are no longer computed in the tapered layers and the resulting maps can be folded into adjacent linear projections. The results indicate that internal normalization can be tapered in the tested pre-training and fine-tuning settings with small validation-loss increases. Our approach helps reveal a distinct role for final normalization, namely that it anchors the scale of the pre-logit representation. With this anchor present, radial changes in the last hidden state do not directly reduce the loss; when it is removed, reducing cross-entropy can be achieved by increasing logit magnitudes. A fixed-target scale loss provides an explicit alternative anchor and enables fully norm-free ablations in the tested regimes. Finally, in a KV-cached autoregressive decoding benchmark, tapering internal norms gives up to $1.14\times$ higher throughput with explicit scaling operations and up to $1.18\times$ after folding.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Modular Multi-Task Learning for Chemical Reaction Prediction

Jiayun Pang, Ahmed M. Zaitoun, Xacobe Couso Cambeiro et al. (+1 more)

Adapting large language models (LLMs) trained on broad organic chemistry to smaller, domain-specific reaction datasets is a key challenge in chemical and pharmaceutical R&D. Effective specialisation requires learning new reaction knowledge while preserving general chemical understanding across related tasks. Here, we evaluate Low-Rank Adaptation (LoRA) as a parameter-efficient alternative to full fine-tuning for organic reaction prediction on limited, complex datasets. Using USPTO reaction classes and challenging C-H functionalisation reactions, we benchmark forward reaction prediction, retrosynthesis and reagent prediction. LoRA achieves accuracy comparable to full fine-tuning while effectively mitigating catastrophic forgetting and better preserving multi-task performance. Both fine-tuning approaches generalise beyond training distributions, producing plausible alternative solvent predictions. Notably, C-H functionalisation fine-tuning reveals that LoRA and full fine-tuning encode subtly different reactivity patterns, suggesting more effective reaction-specific adaptation with LoRA. As LLMs continue to scale, our results highlight the practicality of modular, parameter-efficient fine-tuning strategies for their flexible deployment for chemistry applications.
Read the full abstract

All authors Jiayun Pang, Ahmed M. Zaitoun, Xacobe Couso Cambeiro, Ivan Vulić

Computation and Language

When are We Worried? Temporal Trends of Anxiety and What They Reveal about Us

Saif M. Mohammad

In this short paper, we make use of a recently created lexicon of word-anxiety associations to analyze large amounts of US and Canadian social media data (tweets) to explore *when* we are anxious and what insights that reveals about us. We show that our levels of anxiety on social media exhibit systematic patterns of rise and fall during the day – highest at 8am (in-line with when we have high cortisol levels in the body) and lowest around noon. Anxiety is lowest on weekends and highest mid-week. We also examine anxiety in past, present, and future tense sentences to show that anxiety is highest in past tense and lowest in future tense. Finally, we examine the use of anxiety and calmness words in posts that contain pronouns to show: more anxiety in 3rd person pronouns (he, they) posts than 1st and 2nd person pronouns and higher anxiety in posts with subject pronouns (I, he, she, they) than object pronouns (me, him, her, them). Overall, these trends provide valuable insights on not just when we are anxious, but also how different types of focus (future, past, self, outward, etc.) are related to anxiety.
Read the full abstract

Computation and Language

Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders

Zhongzhi Li, Xuansheng Wu, Yijiang Li et al. (+2 more)

The diversity of post-training data is critical for effective downstream performance in large language models (LLMs). Many existing approaches to constructing post-training data quantify diversity using text-based metrics that capture linguistic variation, but such metrics provide only weak signals for the task-relevant features that determine downstream performance. In this work, we introduce Feature Activation Coverage (FAC) which measures data diversity in an interpretable feature space. Building upon this metric, we further propose a diversity-driven data synthesis framework, named FAC Synthesis, that first uses a sparse autoencoder to identify missing features from a seed dataset, and then generates synthetic samples that explicitly reflect these features. Experiments show that our approach consistently improves both data diversity and downstream performance on various tasks, including instruction following, toxicity detection, reward modeling, and behavior steering. Interestingly, we identify a shared, interpretable feature space across model families (i.e., LLaMA, Mistral, and Qwen), enabling cross-model knowledge transfer. Our work provides a solid and practical methodology for exploring data-centric optimization of LLMs.
Read the full abstract

All authors Zhongzhi Li, Xuansheng Wu, Yijiang Li, Lijie Hu, Ninghao Liu

Computation and Language

When Tables Go Crazy: Evaluating Multimodal Models on French Financial Documents

Virginie Mouilleron, Théo Lasnier, Anna Mosolova et al. (+1 more)

Vision-language models (VLMs) perform well on many document understanding tasks, yet their reliability in specialized, non-English domains remains underexplored. This gap is especially critical in finance, where documents mix dense regulatory text, numerical tables, and visual charts, and where extraction errors can have real-world consequences. We introduce Scribe Finance, the first multimodal benchmark for evaluating French financial document understanding. The dataset contains 1,204 expert-validated questions spanning text extraction, table comprehension, chart interpretation, and multi-turn conversational reasoning, drawn from real investment prospectuses, KIDs, and PRIIPs. We evaluate six open-weight VLMs (8B-124B parameters) using an LLM-as-judge protocol. While models achieve strong performance on text and table tasks (85-90% accuracy), they struggle with chart interpretation (34-62%). Most notably, multi-turn dialogue reveals a sharp failure mode: early mistakes propagate across turns, driving accuracy down to roughly 50% regardless of model size. These results show that current VLMs are effective for well-defined extraction tasks but remain brittle in interactive, multi-step financial analysis. Scribe Finance offers a challenging benchmark to measure and drive progress in this high-stakes setting.
Read the full abstract

All authors Virginie Mouilleron, Théo Lasnier, Anna Mosolova, Djamé Seddah

Computation and Language

Language Triggers Hijack Language Circuits: A Mechanistic Analysis of Backdoor Behaviors in Large Language Models

Théo Lasnier, Wissam Antoun, Francis Kulumba et al. (+2 more)

Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of trigger-induced language-switching backdoors injected during pre-training, studying the Gaperon model family (1B, 8B and 24B). Using activation patching, we localize trigger formation and identify which attention heads process trigger and natural language information. Our central finding is that trigger heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.43 over the top 10 heads identified. This suggests that backdoor triggers do not form new circuits but instead co-opt the model's existing language components and representations. These findings have implications for backdoor defense as detection methods and mitigation strategies could leverage this entanglement between triggers and natural behaviors. More broadly, our work represents a first step toward a more realistic mechanistic understanding of pre-training-injected backdoors in LLMs, paving the way for principled, interpretability-driven defenses.
Read the full abstract

All authors Théo Lasnier, Wissam Antoun, Francis Kulumba, Benoît Sagot, Djamé Seddah

Computation and Language

The Alignment Bottleneck in Decomposition-Based Claim Verification

Mahmud Elahi Akhter, Federico Ruggeri, Iman Munire Bilal et al. (+2 more)

Structured claim decomposition is often proposed as a solution for verifying complex, multi-faceted claims, yet empirical results have been inconsistent. We argue that these inconsistencies stem from two overlooked bottlenecks: evidence alignment and sub-claim error profiles. To better understand these factors, we introduce a new dataset of real-world complex claims, featuring temporally bounded evidence and human-annotated sub-claim evidence spans. We evaluate decomposition under two evidence alignment setups: Sub-claim Aligned Evidence (SAE) and Repeated Claim-level Evidence (SRE). Our results reveal that decomposition brings significant performance improvement only when evidence is granular and strictly aligned. By contrast, standard setups that rely on repeated claim-level evidence (SRE) fail to improve and often degrade performance as shown across different datasets and domains (PHEMEPlus, MMM-Fact, COVID-Fact). Furthermore, we demonstrate that in the presence of noisy sub-claim labels, the nature of the error ends up determining downstream robustness. We find that conservative "abstention" significantly reduces error propagation compared to aggressive but incorrect predictions. These findings suggest that future claim decomposition frameworks must prioritize precise evidence synthesis and calibrate the label bias of sub-claim verification models.
Read the full abstract

All authors Mahmud Elahi Akhter, Federico Ruggeri, Iman Munire Bilal, Rob Procter, Maria Liakata

Tuesday 10 February

Computation and Language

AraModernBERT: Transtokenized Initialization and Long-Context Encoder Modeling for Arabic

Omar Elshehy, Omer Nacar, Abdelbasset Djamai et al. (+3 more)

Encoder-only transformer models remain widely used for discriminative NLP tasks, yet recent architectural advances have largely focused on English. In this work, we present AraModernBERT, an adaptation of the ModernBERT encoder architecture to Arabic, and study the impact of transtokenized embedding initialization and native long-context modeling up to 8,192 tokens. We show that transtokenization is essential for Arabic language modeling, yielding dramatic improvements in masked language modeling performance compared to non-transtokenized initialization. We further demonstrate that AraModernBERT supports stable and effective long-context modeling, achieving improved intrinsic language modeling performance at extended sequence lengths. Downstream evaluations on Arabic natural language understanding tasks, including inference, offensive language detection, question-question similarity, and named entity recognition, confirm strong transfer to discriminative and sequence labeling settings. Our results highlight practical considerations for adapting modern encoder architectures to Arabic and other languages written in Arabic-derived scripts.
Read the full abstract

All authors Omar Elshehy, Omer Nacar, Abdelbasset Djamai, Muhammed Ragab, Khloud Al Jallad, Mona Abdelazim

Computation and Language

PulseLM: A Foundation Dataset and Benchmark for PPG-Text Learning

Hung Manh Pham, Jinyang Wu, Xiao Ma et al. (+6 more)

Photoplethysmography (PPG) is a widely used non-invasive sensing modality for continuous cardiovascular and physiological monitoring across clinical, laboratory, and wearable settings. While existing PPG datasets support a broad range of downstream tasks, they typically provide supervision in the form of numerical measurements or task-specific labels, limiting their compatibility with language-based interfaces and multimodal foundation models. In this work, we introduce PulseLM, a large-scale PPG-text question-answering dataset that bridges raw PPG waveforms and natural language through a unified question-answering (QA) formulation. PulseLM aggregates PPG recordings from sixteen publicly available sources and harmonizes heterogeneous annotations into 12 downstream tasks. The dataset comprises over 1 million standardized 10-second PPG segments, associated with nearly 2.5 million question-answer pairs. We further define reproducible data pipeline, training, and evaluation protocols and establish baseline benchmarks using multimodal PPG-aware large language models. PulseLM provides a standardized foundation for studying language-grounded physiological inference, cross-dataset generalization, and scalable benchmarking of PPG-based multimodal models. We publicly release the dataset and code at https://huggingface.co/datasets/Manhph2211/PulseLM and https://github.com/manhph2211/PULSE-LM, respectively.
Read the full abstract

All authors Hung Manh Pham, Jinyang Wu, Xiao Ma, Yiming Zhang, Yixin Xu, Aaqib Saeed, Bin Zhu, Zhou Pan, Dong Ma

Computation and Language

Certainty robustness: Evaluating LLM stability under self-challenging prompts

Mohammadreza Saadat and Steve Nemzer

Large language models (LLMs) often present answers with high apparent confidence despite lacking an explicit mechanism for reasoning about certainty or truth. While existing benchmarks primarily evaluate single-turn accuracy, truthfulness or confidence calibration, they do not capture how models behave when their responses are challenged in interactive settings. We introduce the Certainty Robustness Benchmark, a two-turn evaluation framework that measures how LLMs balance stability and adaptability under self-challenging prompts such as uncertainty ("Are you sure?") and explicit contradiction ("You are wrong!"), alongside numeric confidence elicitation. Using 200 reasoning and mathematics questions from LiveBench, we evaluate four state-of-the-art LLMs and distinguish between justified self-corrections and unjustified answer changes. Our results reveal substantial differences in interactive reliability that are not explained by baseline accuracy alone: some models abandon correct answers under conversational pressure, while others demonstrate strong resistance to challenge and better alignment between confidence and correctness. These findings identify certainty robustness as a distinct and critical dimension of LLM evaluation, with important implications for alignment, trustworthiness and real-world deployment.
Read the full abstract