Peer Revenue Returns subject to peer review

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Friday 22 May

Computation and Language

Naturalistic measure of social norms alignment

Yevhen Kostiuk, Kenneth Enevoldsen, Peter Bjerregaard Vahlstrup et al. (+2 more)

Social norms reflect shared expectations on acceptable behavior. Measuring social norms alignment remains challenging, with existing approaches typically relying on artificial closed-form evaluations such as multiple-choice questionnaires or measuring agreement with predefined statements. In the context of this work, social norms alignment refers to measuring an agreement between solutions with respect to the social problem or dilemma. We propose a framework for measuring social norm alignment in naturalistic, free-form settings through solution matching. The framework enables us to measure alignment between any two dilemma responses e.g., LLMs to a human, LLMs to LLMs, or human to human. We introduce two metrics: stated and explicit agreement accuracy, and construct a dataset of 3k non-trivial social dilemmas in Danish. All dilemmas are assigned reference solutions derived from three panelists, who serve as culturally grounded judges. We evaluate the agreement of several LLMs and human responses in an interaction setup that resembles natural user-model conversations. Our results show that the proposed metrics produce consistent model rankings and reveal variation in agreement across different types of dilemmas, with higher agreement observed for topics such as neighbor conflicts and shared living situations. Overall, our work introduces a dataset and evaluation framework for studying culturally grounded social reasoning in naturalistic open-ended conversations.
Read the full abstract

All authors Yevhen Kostiuk, Kenneth Enevoldsen, Peter Bjerregaard Vahlstrup, Márton Kardos, Kristoffer Nielbo

Computation and Language

Articulatory strategy as a source of variation in acoustic vowel dynamics

Patrycja Strycharczuk, Justin J. H. Lo and Sam Kirkham

Acoustic vowel dynamics have some speaker-identifying characteristics, which have been ascribed to individual properties of articulatory strategies: formant transitions have a particular shape because speakers move their articulators, using specific and practised movements. However, there is little existing evidence that different articulatory strategies systematically affect formant dynamics. The present study corroborates the link between the two. Ultrasound tongue imaging data from 36 speakers of Northern-Anglo English are used to identify distinct articulatory strategies for the production of palatal vowel /i/. Tongue shape in /i/ is found to be a significant predictor of formant dynamics in diphthongs with a palatal offglide. The observed relationships can be explained by the characteristics of articulatory movement conditioned by vocal tract shape. Greater articulatory displacement of tongue root and/or dorsum produces greater distortion from the mean tongue shape in palatal vowels, and it also requires higher articulatory velocities, resulting in relatively earlier and steeper formant transitions. The results contribute to the conceptual understanding of individuality in speech, by illuminating the regularising and individual aspects of articulatory compensation.
Read the full abstract

Computation and Language

EquiSumm : A Gender Bias-Aware Framework for Inclusive Tweet Summarization

Chaitanya Wanjari, Jessica Kamal, Riddhi Jain et al. (+2 more)

While social media platforms, such as Twitter, provide a medium for large-scale opinion sharing during news events, it is manually impossible for individuals or media agencies to process the vast volume of content to identify key viewpoints. In order to resolve this, several automatic summarization techniques have been proposed to condense large collections of tweets into concise and informative summaries. However, these algorithms do not explicitly consider demographic fairness. Several existing research works have developed automated summarization approaches that can provide a holistic overview of the key aspects and major opinions shared on social media platforms related to a news event. However, these approaches do not explicitly consider different forms of demographic representation, such as gender, which can lead to biased summary representation. In this paper, we propose EquiSumm, which considers the gender aspect of the shared opinion to generate a summary, and our experimental analysis on two major datasets indicates the performance effectiveness with respect to existing research works.
Read the full abstract

All authors Chaitanya Wanjari, Jessica Kamal, Riddhi Jain, Samruddhi Kurhe, Roshni Chakraborty

Computation and Language

Metacognition as Reward: Reinforcing LLM Reasoning via Knowledge and Regulation Signals

Sirui Chen, Lei Xu, Yuying Zhao et al. (+6 more)

Recent RL methods have substantially improved the reasoning abilities of LLMs. Existing reward designs mainly follow two paradigms: (1) Reinforcement learning with verifiable rewards (RLVR) derives outcome signals from executable checks or ground-truth answers, but provides limited guidance for intermediate reasoning behaviors. (2) Rubrics-as-reward (RaR) goes beyond final-answer checking by using natural-language rubrics to assess reasoning quality and task compliance, but often requires instance-specific rubrics and substantial design effort. To address these issues, we introduce Metacognition-as-Reward (MaR), a metacognition-inspired RL framework that guides LLM reasoning through two general process dimensions: i) metacognitive knowledge, which identifies task-relevant information without hand-crafted instance-specific rubrics, and ii) metacognitive regulation, which plans and adjusts the reasoning process to provide reward guidance beyond final-answer outcomes. MaR scaffolds model rollouts into explicit metacognitive components and optimizes them with a trajectory-level reward over task knowledge coverage, regulation fidelity, and final-answer correctness. In this way, MaR extends reward feedback to reasoning trajectories while grounding the reward signals in general metacognitive dimensions. Experiments on 22 benchmarks show that MaR consistently improves model performance, achieving up to a 7.7% gain over the base model and up to an 11.0% gain over vanilla DAPO. Notably, Qwen3.5-9B + MaR narrows the gap to frontier models, surpassing GPT-OSS-120B on overall average and outperforming stronger models on several individual benchmarks. Process-level analysis further shows substantial improvements in reasoning process quality. MaR also generalizes to out-of-domain datasets, where MaR-trained models improve over their corresponding base models on average.
Read the full abstract

All authors Sirui Chen, Lei Xu, Yuying Zhao, Yutian Chen, Yu Wang, Beier Zhu, Hanwang Zhang, Shengjie Zhao, Chaochao Lu

Computation and Language

From Correctness to Preference: A Framework for Personalized Agentic Reinforcement Learning

Ranxu zhang, zeyang li, Jiacheng Huang et al. (+5 more)

Agentic reinforcement learning (Agentic RL) has achieved strong progress in tasks with clear success signals. However, many real-world agent applications require user-conditioned behavior: the same query may call for different planning strategies and tool-use decisions across users. This setting raises key challenges: generic rewards cannot capture heterogeneous user preferences, observed behaviors are entangled with conformity effects, and flat memories cannot support personalized skill retrieval. To this end, we propose a unified personalized Agentic RL framework that embeds personalization into training-time optimization. At its core is Personalized Anchor Reward-Decoupled Policy Optimization (PARPO), which decouples generic task-quality rewards from personalized preference rewards and uses user-specific anchors to stabilize learning under heterogeneous reward scales. We further introduce a two-stage preference-disentangled reward model and Preference-Aligned Skill Evolution Graph Memory (PSGM) for personalized supervision and preference-aligned skill retrieval. Together, they form a closed loop of preference identification, policy optimization, and structured skill accumulation. Experiments on ETAPP, ETAPP-Hard, and SJAgent show that our framework consistently outperforms strong memory and RL baselines. Code and data are included in the supplementary materials.
Read the full abstract

All authors Ranxu zhang, zeyang li, Jiacheng Huang, Rui Zhang, Xiaozhou Xu, sun zhe, Yanyong Zhang, Chao Wang

Computation and Language

Cultural Adaptation in Large Language Models for Political Discourse

Wajdi Zaghouani

The integration of large language models into political discourse analysis creates new opportunities for comparative research, policy analysis, and civic technology, while introducing material risks for democratic accountability. This paper argues that cultural adaptation is a prerequisite for trustworthy deployment of large language models in political communication across diverse linguistic and institutional contexts. Current systems remain shaped by English dominant data, uneven multilingual coverage, and assumptions grounded in a narrow range of political institutions and discourse conventions, producing systematic errors when applied across cultures. We formalize cultural adaptation across translation, discourse, and ontology levels, identify recurring cultural failure modes in political NLP, and propose an operational evaluation matrix grounded in cultural fidelity, calibration, and democratic safety. Building on political text analysis, sociotechnical auditing, and cross cultural pragmatics, we outline methodological pathways including participatory dataset development, culturally aware transfer learning, and benchmark design that makes cultural adaptation empirically measurable. We conclude by clarifying governance constraints and scope conditions under which culturally adaptive political NLP can support democratic legitimacy.
Read the full abstract

Computation and Language

Emotion Recognition in Sign Language Conversation

Yusong Wang, Keyu Mao, Takao Obi et al. (+2 more)

Emotion Recognition in Conversation is a core component of affective computing, while current sign language emotion datasets primarily focus on isolated sentences and lack conversational context. Models trained exclusively on these isolated utterances demonstrate degraded performance in real world scenarios because they cannot utilize historical dialogue flow. To address this structural limitation, we introduce the ERC task to sign language video analysis and propose the eJSL Dialog dataset. Constructed using the scripts from the STUDIES corpus, the dataset contains 1,920 video samples organized into 480 unique dialogues. We conduct systematic benchmarking on this dataset using models ranging from isolated visual networks to multimodal conversational architectures. The results suggest the feasibility of extending conversational ERC frameworks to sign-language dialogue under the current benchmark setting, while also revealing limitations in existing visual representations for capturing sign-specific affective cues, motivating future work on sign-specific visual modeling and larger sign-language conversational training resources.
Read the full abstract

All authors Yusong Wang, Keyu Mao, Takao Obi, Minghao Shao, Kotaro Funakoshi

Computation and Language

ClimateChat-300K: A Multi-Modal Facebook Dataset for Understanding Diverse Perspectives in Climate Communication

Wajdi Zaghouani, Md. Rafiul Biswas, Mabrouka Bessghaier et al. (+2 more)

We present ClimateChat-300K, a large-scale dataset of 299,329 public Facebook posts about climate change collected between May 2020 and May 2024 through the CrowdTangle platform. The dataset contains 41 metadata features including post content, engagement metrics, and page attributes, covering material from more than 26,000 global pages. Each post includes rich contextual information such as language, timestamp, page category, and interaction counts, enabling comprehensive analyses of public discourse around climate communication. Using topic modeling and sentiment analysis, we identify ten main themes grouped into five domains: policy, activism, cooperation, science, and conservation. The results reveal that emotional tone, post format, and page identity strongly influence audience engagement, with visually rich and emotionally charged content receiving the highest levels of interaction. The dataset also demonstrates how online discussions evolved in response to major events such as international climate summits and the COVID-19 pandemic period. ClimateChat-300K provides an open resource for reproducible and interdisciplinary research on polarization, misinformation, and the dynamics of digital climate discourse. By releasing this dataset, we aim to support transparent, data-driven research and contribute to a deeper un-derstanding of how public engagement with climate issues develops across time, geography, and institutional contexts.
Read the full abstract

All authors Wajdi Zaghouani, Md. Rafiul Biswas, Mabrouka Bessghaier, Shimaa Ibrahim, George Mikros

Computation and Language

AraHopeCorpus: Annotation Guidelines and Dataset for Hope Speech in Arabic Social Media Crisis Discourse

Esra'a Sharqawi and Wajdi Zaghouani

Social media has become a crucial arena for shaping public narratives during armed conflicts, providing space for both harmful and constructive communication. While hate speech and misinformation have been widely studied, expressions that promote resilience, solidarity, and optimism remain underexplored, particularly in Arabic contexts. This paper introduces AraHopeCorpus, the first annotated dataset of Arabic hope speech collected from ten thousand YouTube comments related to the war on Gaza between 2023 and 2024. Using a detailed annotation framework, comments were classified into three categories: hope speech, no hope speech, and neutral or unclear discourse. The dataset shows that hopeful language dominates, accounting for more than sixty four percent of all comments. These expressions of hope appear mainly as religious encouragement, collective solidarity, and optimism for endurance and justice. No hope speech, representing about thirteen percent, reflects despair and disillusionment, while the rest of the comments contain neutral or mixed content. Inter-Annotator Agreement reached substantial levels (Cohen's Kappa equals 0.71), though dialectal variation, sarcasm, and implicit meaning posed annotation challenges. A comparative analysis between human annotators and ChatGPT revealed that large language models can support annotation but remain limited in handling dialectal and culturally embedded expressions. AraHopeCorpus will be released for research purposes under an open and non commercial license. It provides a valuable resource for studying constructive digital discourse, enabling further research on hope speech detection, crisis communication, and resilience in Arabic social media.
Read the full abstract

Computation and Language

Convergence Without Understanding: When Language Models Agree on Representations but Disagree on Reasoning

Muhammad Usama and Dong Eui Chang

Large language models trained under diverse objectives and architectures have been shown to develop increasingly similar internal representations, an observation formalized as the Platonic Representation Hypothesis. Whether this representational convergence extends to the reasoning processes that operate over shared representations remains untested. We evaluate representational similarity across 16 language models from 8 families (1.5B to 72B parameters) on 800 reasoning problems spanning mathematics, science, commonsense, and truthfulness, stratifying by problem difficulty, computational stage, and causal relevance. Our analysis reveals three dissociations: a difficulty inversion, where models converge more on problems they collectively fail (Centered Kernel Alignment [CKA] = 0.897) than on those they solve (CKA = 0.830); a generation gap, where pre-decision representations align (CKA = 0.875) while post-decision representations diverge (CKA = 0.274); and epiphenomenal correctness, where shared information is decodable across models (66% transfer accuracy) but exerts minimal causal influence on predictions (1.5% to 5.5% flip rate across ablation protocols). These results indicate that representational convergence in language models reflects shared input processing constraints rather than shared reasoning strategies, with direct implications for ensemble design, interpretability transfer, and evaluations of model similarity. Code is available at https://github.com/Usama1002/convergence-without-understanding.
Read the full abstract

Computation and Language

When Is Next-Token Prediction Useful? Marginalization, Ergodicity, Mixture Identifiability, Local Sufficiency, RAG, Tools, and Programming

Francesco Corielli

Language models trained on observed sequences are often described as learning the conditional distribution of the next token given previous tokens. This description is only conditionally correct. A model trained on realized token trajectories does not observe full conditional laws; it receives sampled continuations. Moreover, real language generation is conditioned not only on previous words but also on non-textual circumstances: facts, events, intentions, goals, beliefs, social context, and task-specific constraints. This paper distinguishes three objects that are often conflated: the full conditional language process conditioned on latent circumstances, the marginal text-only process obtained by integrating those circumstances out, and the model-induced distribution learned from finite observed corpora. The paper argues that interpreting model training as estimating the marginal text-only law requires strong assumptions of stationarity, representativeness, and ergodicity, assumptions that are standard in statistical estimation but problematic when applied to heterogeneous language corpora. Even if these assumptions hold, the marginal text-only law is useful only when the observed prefix is an approximately sufficient statistic for the latent circumstances relevant to continuation. In information-theoretic terms, usefulness requires that the residual conditional mutual information between the next token and the omitted circumstances, given the observed text, be small. The paper then extends this argument to heterogeneous training corpora. Finally, the paper interprets Retrieval Augmented Generation (RAG) and tool use as conditional sufficiency devices.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Multi-Gate Residuals

Zhizhan Zheng, Feiyun Zhang, Shuchun Liu et al. (+4 more)

While Attention Residuals has shown some effectiveness in addressing the widespread issue of unbounded activation growth across deep residual layers, it inevitably incurs significant communication overhead. To circumvent this bottleneck, we propose Multi-Gate Residuals (MGR), which stabilizes activation scales without additional communication burden. It utilizes a straightforward scoring and gating mechanism to maintain multi-stream context, coupled with Attention Pooling to extract hidden states from the stream states. Empirical experiments demonstrate that MGR is practical for large-scale training and deployment, offering tangible performance improvements over existing architectures.
Read the full abstract

All authors Zhizhan Zheng, Feiyun Zhang, Shuchun Liu, Tian Xia, Xi Liu, Dasheng Hu, Hongquan Zhou

Computation and Language · Cross-listed from cs.LG

FastKernels: Benchmarking GPU Kernel Generation in Production

Gabriele Oliaro, Jaeseong Lee, Yichao Fu et al. (+8 more)

LLM-based agents for GPU kernel generation are advancing rapidly, but the benchmarks they optimize against evaluate kernels in isolation, with synthetic inputs and weak baselines, rewarding sandbox speedups that break or vanish in real inference systems. We introduce FastKernels, a benchmark of 384 tasks drawn from 47 representative architectures across 8 categories, whose kernels suffice to reimplement 94.6% (472/499) of HuggingFace Transformers architectures with outputs matching the native implementations. Each task mirrors the interface of the corresponding production module and is scored against the kernels production frameworks ship, and tasks form a compositional hierarchy, from primitives to full models, in which higher-level modules import lower-level ones. Candidates are scored at the kernel level and end to end inside the models they come from, on the production execution path, and MacroEval aggregates calibrated correctness, coverage, and speedup into a leaderboard. Seeding it with five representative agents (6,900 agent-hours), we find that kernel-level speedups of $1.6$-$6.6\times$ shrink to at most $1.25\times$ end to end, only 20% of winning kernel sets run correctly as-is, and kernel-level scores mis-rank agents: Claude Code matches or beats KDA at every level in isolation, yet KDA scores $3\times$ higher end to end. Code is available at https://github.com/Snowflake-AI-Research/fastkernels.
Read the full abstract

All authors Gabriele Oliaro, Jaeseong Lee, Yichao Fu, Zhaoyuan Su, Owen Lu, Sreeram Vennam, May Jiang, Junli Wang, Hao Zhang, Zhihao Jia, Samyam Rajbhandari

Computation and Language · Cross-listed from cs.HC

CultivAgents: Cultivating Relationship-Centered Multi-Agent Systems for Personalized Gardening

Yiyang Wang, Moeiini Reilly, Britney Johnson et al. (+3 more)

Gardening is critical to support well-being, cultural continuity, and food autonomy, yet existing digital tools often provide generic advice that overlooks gardeners' skills, local ecologies, seasons, and cultural contexts. We introduce CultivAgents, a relationship-centered multi-agent system for personalized, socio-culturally grounded gardening support. Grounded in ethics of care, CultivAgents coordinates multiple specialized agents: an Experience Agent that adapts guidance to users' skill levels, an Environmental Agent that grounds advice in local and seasonal conditions, and an Ethnobotanical Agent that connects plants to cultural knowledge and histories. We evaluated CultivAgents through a three-phase mixed-methods study with domain experts (n=3), HCI researchers (n=7), and community gardeners (n=5), analyzing expert feedback, pre/post surveys, and participatory design activities. Results suggest that CultivAgents helped gardeners translate interest into situated action: community gardeners reported increased confidence (3.00 to 3.60), motivation (4.00 to 4.40), and trust in acting on AI advice (3.20 to 4.00). Participants valued hyperlocal ecological guidance and complementary agent perspectives, while also identifying limits in cultural specificity, ecological grounding, and agent coordination. The work advances relationship-centered AI, offering design implications for multi-agent systems that support food sovereignty, community resilience, and cultural preservation.
Read the full abstract

All authors Yiyang Wang, Moeiini Reilly, Britney Johnson, Kefei Yan, Alex Cabral, Josiah Hester

Computation and Language

Hidden Human-Like Nature of Machine-Generated Texts: Theory and Detection Enhancement

Chenwang Wu, Yiu-ming Cheung, Bo Han et al. (+1 more)

Machine-generated texts (MGTs) produced by large language models (LLMs) are increasingly prevalent across various applications, while their potential misuse in fake news propagation and phishing has raised serious concerns, highlighting the need for MGT detection. Existing paragraph-level detection methods commonly treat MGTs as entirely machine-like, overlooking the hidden human-like nature of machine-generated texts: even fully machine-generated texts may contain spans that are highly consistent with human writing. To this end, we first reveal the existence of such hidden human-like spans, and then theoretically analyze their impact on detection. Our analysis shows that these spans increase the sentence complexity for detection, thereby making MGT detection intrinsically harder. Based on this finding, we propose a model-agnostic stacked enhancement framework that improves existing detectors by reducing the influence of hidden human-like spans. Specifically, we model span-level retention decisions as a latent-variable problem and instantiate the optimization with a hard-EM-inspired procedure, where the detector iteratively filters confidently human-like subsequences and refines itself on the remaining text. Extensive experiments across various LLMs and practical scenarios demonstrate that the proposed framework consistently enhances existing detectors. Notably, the framework can also work in a training-free manner, offering flexibility and scalability for practical deployment.
Read the full abstract

All authors Chenwang Wu, Yiu-ming Cheung, Bo Han, Defu Lian

Computation and Language

Self-Improving In-Context Learning

Baturay Saglam and Dionysis Kalogerias

We propose to improve in-context learning (ICL) by optimizing the continuous embeddings of a fixed few-shot prompt at test time. The key observation is that the log-probabilities a model assigns to its demonstrated outputs$\unicode{x2013}$available from a single forward pass without generating any tokens$\unicode{x2013}$provide a meaningful signal for how well the model has inferred the task from its demonstrations. We formalize this signal as a bounded, self-supervised confidence proxy and maximize it via zeroth-order optimization over the prompt embeddings, yielding a test-time calibration procedure. The approach requires no finetuning, no token generation, no predefined label set, and no external data, making it equally applicable to both classification and free-form generation tasks. Across a comprehensive suite of ICL tasks, the proposed calibration consistently matches or improves upon the base model and outperforms classification-specific baselines on most tasks. The statistically significant correlation between proxy improvement and downstream accuracy gain confirms that the proposed proxy encodes a reliable optimization signal for in-context learning.
Read the full abstract

Computation and Language · Cross-listed from cs.CR

Robust LLM Watermarking with Minimal Semantic Distortion for IP Protection

Kieu Dang, Phung Lai, NhatHai Phan et al. (+2 more)

Proprietary large language models (LLMs) face risks of intellectual property (IP) violation, as adversaries can replicate an LLM by collecting input-output pairs to train a surrogate model, causing financial setbacks. Watermarks offer a promising defense to verify ownership, but existing methods often struggle with semantic distortion, factual inconsistency, and adversarial attacks. In addition, key-conditioned watermarks for provider-specific detection, especially in cross-provider and multi-user scenarios, remain largely underexplored. To address these challenges, we propose SAFESEAL, a novel key-conditioned watermarking framework that achieves strong detectability with minimal impact on model utility, effectively balancing detectability, utility, and robustness. SAFESEAL preserves named entities while substituting linguistic terms with context-aware synonyms through a key-conditioned Tournament sampling mechanism, maintaining semantic fidelity and factual consistency. For detection, we introduce a key-conditioned contrastive detector that jointly encodes the text and key, enabling provider-specific and robust watermark verification. We derive theoretical bounds on the utility-detectability trade-off and significantly reduce latency through lightweight models, batching, and parallelism. Extensive experiments show that SAFESEAL outperforms baselines in utility, detectability, and robustness, achieving a BERTScore of 0.983, entity similarity of 0.963, a 98.2% detection rate, and the highest human ratings for text quality and content preservation, with latency comparable to the fastest baseline. To promote transparency and community-driven progress, we release the first public watermark leaderboard and an interactive demo.
Read the full abstract

All authors Kieu Dang, Phung Lai, NhatHai Phan, Yelong Shen, Ruoming Jin

Computation and Language

Positional Failures in Long-Context LLMs: A Blind Spot in Reasoning Benchmarks

Chuyifei Zhang, Hongyu Cui, Xiaowen Huang et al. (+1 more)

Position-controlled evaluation is standard for retrieval tasks such as Needle-in-a-Haystack and RULER, but mainstream reasoning benchmarks do not control positional placement of target tasks in long contexts. We audit 11 long-context benchmarks and find none jointly controls task position, filler content, and context length for reasoning. An audit of four flagship long-context releases finds no main result-table entry for NIAH, RULER, or LongBench-family benchmarks, while agentic and coding benchmarks appear in main result-tables across all four. We propose Context Rot Evaluation (CRE), a controlled framework varying all three factors, and evaluate nine LLMs on GSM8K and ARC-Challenge across two rounds: an initial five-model set and four newer vendor releases. Models can drop sharply when the target task moves from end to middle, and the drop grows worse with context length for vulnerable models. MiMo-v2-Flash drops 88pp at 64K under with_solutions filler (middle accuracy 8%). Newer releases show smaller drops: at 64K, three of four stay within +/-6pp of end-position accuracy; MiMo-V2.5-Pro narrows the MiMo-v2-Flash 88pp drop to 32pp. Under questions_only_v2 filler, middle-position drops persist across all four (range -16pp to -56pp across 8K, 32K, 64K). At 8K, a diagnostic probe adding a target-task copy at the end brings middle accuracy within +/-4pp of end baseline across all nine models, consistent with a positional explanation. In the initial five-model set, 76% of middle-position errors match surrounding filler text versus 22% at the end position, consistent with filler-answer interference as a dominant error mode. These results expose a structural evaluation gap in current reasoning benchmark design and vendor evaluation practice: positional vulnerabilities that grow with context length cannot be measured when task position is not controlled.
Read the full abstract

All authors Chuyifei Zhang, Hongyu Cui, Xiaowen Huang, Jitao Sang

Computation and Language · Cross-listed from cs.RO

Autonomous Frontier-Based Exploration with VLM Guidance

Aarush Aitha and Avideh Zakhor

Autonomous robotic exploration of unknown and hazardous environments, a long-standing challenge, can be significantly improved by leveraging the advanced reasoning of Vision-Language Models (VLMs). We introduce a novel exploration pipeline where a VLM performs high-level strategic decision-making, guiding a conventional low-level robotics control stack. At decision points, the robot generates a multimodal prompt with its current map and visual imagery of potential paths, or frontiers. The VLM analyzes this prompt to select the most promising frontier, replacing simple geometric heuristics with contextual spatial reasoning. This approach, validated in simulation across six indoor environments, improves map coverage by up to 24% over existing methods. Our pipeline is lightweight, training-free, and easily transferable to any robot with standard sensors and an internet connection.
Read the full abstract

Computation and Language

Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving

Kewei Zhang, Jin Wang, Sensen Gao et al. (+9 more)

End-to-end autonomous driving via Vision-Language-Action (VLA) models demands a precarious balance between high-fidelity trajectory planning and efficient inference. Existing paradigms typically fall short: autoregressive (AR) VLAs are memory-bandwidth-bound on edge hardware and prone to exposure-bias drift, while full-sequence diffusion models preclude KV-cache reuse and suffer from "logical leakage" that violates the fundamental perceive-then-plan causality. We present Fast-dDrive, a block-diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across them. Leveraging the observation that driving VLAs often emit structured JSON-like outputs, Fast-dDrive freezes structural tokens into a section scaffold and employs a section-aware training recipe that prioritizes safety-critical planning. We further introduce Scaffold Speculative Decoding to achieve AR-equivalent quality at significantly higher throughput. Finally, we propose a low-overhead test-time scaling scheme: by forking $N$ stochastic trajectory rollouts from a single shared-prefix KV cache and averaging them, we effectively suppress prediction variance at a fractional computational cost. Empirical results demonstrate that Fast-dDrive redefines the speed-accuracy frontier for driving agents. On the WOD-E2E test set, Fast-dDrive achieves SOTA ADE@3s and ADE@5s, alongside the highest RFS among diffusion-based VLAs; on nuScenes, it reduces average L2 error to $0.32$m (a $22\%$ improvement). When integrated with SGLang, our framework delivers $12\times$ throughput speedup over the AR baseline, narrowing the gap between high-capacity VLAs and the efficiency demands of real-time on-vehicle deployment.
Read the full abstract

All authors Kewei Zhang, Jin Wang, Sensen Gao, Chengyue Wu, Yulong Cao, Songyang Han, Boris Ivanovic, Langechuan Liu, Marco Pavone, Song Han, Daquan Zhou, Enze Xie

Computation and Language · Cross-listed from cs.CR

What Does the Server See? Understanding Privacy Leakage from Large Language Models in Split Inference

Mingyuan Fan, Yu Liu, Fuyi Wang et al. (+1 more)

The deployment of large language models (LLMs) on resource-constrained devices remains challenging, spurring interest in split inference, where models are partitioned between client and server to reduce computational burden and enhance privacy by transmitting only intermediate activations. However, the privacy-preserving capabilities of split inference, particularly in the context of LLMs, have not been exhaustively investigated. To fill this gap, we introduce ActInv, which solves an intermediate activation matching problem to reconstruct the client's input. Extensive evaluations demonstrate that ActInv achieves high-fidelity reconstructions, even in the presence of common perturbation-based defenses such as Gaussian noise injection and activation sparsification. To systematically understand this vulnerability, we develop Perturbation Amplification Factor (PAF), a metric for quantifying a layer's inherent resistance to reconstruction. Our analysis reveals that privacy vulnerability is not uniform across layers, with some layers being highly susceptible to leakage while others offer natural resistance. Furthermore, we demonstrate that defense effectiveness can be significantly improved by calibrating perturbation directions to maximize reconstruction error during backpropagation. Building on these insights, we design PriPert and conduct comprehensive evaluations, covering privacy, utility, and computational overhead, to demonstrate its effectiveness.
Read the full abstract

All authors Mingyuan Fan, Yu Liu, Fuyi Wang, Cen Chen

Computation and Language

Same Model, Different Weakness: How Language and Modality Reshape the Jailbreak Attack Surface in Frontier MLLMs

Casey Ford, Madison Van Doren, Sicheng Jin et al. (+1 more)

The attack surface of a multimodal large language model (MLLM) is language-dependent in ways that reveal the mechanistic structure of alignment failures. We present the first systematic cross-lingual, multimodal red-teaming study comparing jailbreak vulnerability in US English (en-US) and Mexican Spanish (es-MX) across four frontier MLLMs: Claude Sonnet 4.5, GPT-5, Pixtral Large, and Qwen Omni. Using a fixed adversarial benchmark of 363 diverse prompt scenarios administered in text-only and multimodal conditions, we collected 52,272 harm ratings and binary attack success judgements from matched panels of nine native-speaker annotators per language group. Our central finding is that language does not scale vulnerability uniformly. Bayesian mixed-effects analyses reveal that linguistic framing attacks such as role-play become substantially less effective under Spanish prompting, while visually explicit multimodal attacks become more effective, which directly implicates the prompt-language interface rather than global annotator leniency. This dissociation indicates that linguistic and visual alignment failures operate through distinct mechanisms, and that switching language is sufficient to expose that separation. The practical consequence is that safety rankings are not preserved across languages. Qwen Omni overtakes Pixtral Large as the most vulnerable model among es-MX participants, a rank reversal no scalar correction of English-condition scores could recover, and absolute attack success rates have declined across model generations without closing the gaps between them. These findings demonstrate that safety evaluation frameworks treating language and modality as independent dimensions fundamentally misspecify the attack surface of globally deployed MLLMs, and must be redesigned accordingly.
Read the full abstract

All authors Casey Ford, Madison Van Doren, Sicheng Jin, Emily Dix

Computation and Language

When Symptoms Are Not Enough: Evidence-Weighting Patterns in Large Language Model Psychiatric Screening

Jianfeng Zhu, Megan Korhummel, Ruoming Jin et al. (+1 more)

As demand for mental health care outpaces clinician-delivered assessment, scalable screening tools are increasingly needed. Large language models (LLMs) may identify psychiatric risk from patient narratives, but their reliability across diagnoses, demographic subgroups, and evidence-use patterns remains uncertain. We introduce a SCID-anchored benchmark of 555 semi-structured experiential interviews paired with diagnostic reference labels for anxiety disorder, major depressive disorder, post-traumatic stress disorder, and any current mental health disorder. Using zero-shot task-specific prompting, we evaluated five state-of-the-art LLMs and examined whether false-negative errors reflected missed psychiatric evidence or differential weighting of symptom, functional-impairment, and protective-context cues. Performance varied across tasks and models, with accuracy ranging from 0.49 to 0.86 and Matthews correlation coefficients from 0.16 to 0.38. GPT-4.1 Mini and GPT-5 Mini showed the most consistent disorder-specific accuracy. Subgroup analyses found higher depression-classification accuracy among male than female participants, no consistent age-related pattern, and modest non-uniform variation across race strata. Evidence-integration analyses showed that false-negative anxiety and PTSD classifications often contained explicit symptom evidence but were accompanied by preserved functioning, coping ability, or social support. Functional-impairment evidence shifted model outputs toward positive classifications, whereas protective-context evidence shifted outputs away. These findings suggest that LLMs may support scalable psychiatric screening, but their tendency to discount symptom evidence in the presence of preserved functioning or protective context requires careful validation before clinical deployment.
Read the full abstract

All authors Jianfeng Zhu, Megan Korhummel, Ruoming Jin, Karin G. Coifman

Computation and Language

As X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs

Eric Xu

Role prompts of the form As X, do Y admit a clean linear decomposition at one specific site in the residual stream: the prompt-to-answer transition – the last prompt token together with the first two generated tokens – in an early/mid layer band. There, persona and task contribute through partially orthogonal additive directions. Forming a pure persona effect $\Delta_X$, a pure task effect $\Delta_Y$, and substituting $h_{BB} + \Delta_X + \Delta_Y$ for the clean residual yields downstream output within a small KL of clean on Gemma-2-2B-IT and Qwen-2.5-\{1.5B, 3B\}-Instruct, across a 12-cell short grid and a 48-cell long-persona grid, with persona-specific behavioral markers preserved. The natural inference from this additive structure is that the role prompt can be compressed into a single cached residual vector. We show it cannot. Injecting the cached additive prediction – or even the oracle clean residual $h_{XY}$ – into a baseline host prompt with the persona text removed does not approach the clean long-persona target, at one site or at many layers. Persona-conditioned multi-token generation flows through attention back to the persona-text positions throughout the prompt, which no residual at one site reproduces. Local additivity in the residual stream does not imply prompt compressibility. The additive structure at the prompt-to-answer transition supports interpretability and fine-grained steering of persona or task contributions; persona-conditioned behavior across the full continuation depends on a distributed prompt/KV mechanism that local activation arithmetic does not displace.
Read the full abstract

Thursday 21 May

Computation and Language

RIG-RoPE: Relation-Stratified Multimodal Attention with Instance-Local Rotary Geometry and Representation-Aware Traversal Coordinates

Donggen Li

Multimodal rotary positional encodings apply temporal, height, and width phases to interleaved text, image, and video tokens. This creates two ambiguities: cross-instance spatial displacement depends on preprocessing chart choices unless registration is declared, and scalar advance across visual blocks is often inherited from coordinate extrema rather than defined at the representation level. We introduce RIG-RoPE, combining instance-local rotary geometry, relation-stratified attention, and representation-aware traversal coordinates. RIG-RoPE normalizes relation-homogeneous scores separately, allocates mass with a common H/W-neutral LogSumExp statistic, and uses traversal extent that is additive over ordered slices and sublinear over parallel spatial scale. Text advances by unit increments, image patches are simultaneous, and video accumulates over tokenizer temporal tokens. In a matched, inference-only Qwen2-VL-2B checkpoint experiment, native and RIG text-only paths were exactly equal. RIG was exactly invariant to a whole-chart single-image translation and to translating only the second instance of an unregistered image pair. Native attention remained sensitive to the latter, while an H/W-collapse control confirmed that RIG retained same-instance spatial effects; visual embeddings and all parameters were unchanged. Across three seeds of a frozen tiny task, RIG also had zero clean-to-Gauge logit change, whereas the raw-H/W baseline changed in every seed. Gauge-accuracy differences were +2/72, 0, and 0, failing the preregistered stability gate. These results support the specified activation and Gauge mechanisms, not stable task improvement, universality, or empirical superiority.
Read the full abstract

2608.05154 · 21 May · 26 pages, 2 figures, 4 tables. This revision adds matched, inference-only Qwen2-VL-2B native/RIG L1 mechanism evidence: exact text-only equivalence, exact RIG Gauge invariance, native Gauge sensitivity, and retained same-instance spatial effects. Task-level Qwen benchmarks, training gains, stable task improvement, universality, and empirical superiority remain unestablished

Abstract PDF $ATTN ×3$MMOD ×2$TOKN ×2$VID ×2 +1

Computation and Language

OncoTriad-QA: A Patient-Level Radiology-Pathology-Genomics Benchmark for Pan-Cancer Reasoning

Ahnaf Munir, Dannong Wang, Michael W. McDonald et al. (+3 more)

Cancer diagnosis and characterization require integrating complementary evidence from radiology, pathology, genomics, and clinical metadata. However, most medical large language model (LLM) and vision-language model (VLM) benchmarks focus on isolated modalities or narrow image-text tasks, leaving patient-level oncology assessment across multiple evidence streams largely untested. We introduce OncoTriad-QA, a patient-level radiology-pathology-genomics benchmark for pan-cancer question answering. OncoTriad-QA contains 86.1k semantic questions across 9,281 TCGA patient cases from 32 cancer cohorts, aligning CT/MRI radiology, whole-slide histopathology, somatic mutations, copy-number alterations, DNA methylation, bulk RNA-seq, and clinical metadata. Case-specific annotations are constructed through a source-grounded LLM-assisted pipeline using curated labels, diagnostic reports, molecular profiles, and modality-derived evidence as primary sources of truth, with automated consistency checks and clinician review. We also introduce OncoVLM, a reference multimodal model that maps modality-native radiology, pathology, DNA methylation, and RNA-seq evidence into an LLM interface through learned projectors. Experiments show that existing general-purpose and medical LLMs remain limited on comprehensive pan-cancer QA, especially when questions require integrating imaging findings, tumor morphology, and molecular evidence. After fine-tuning on OncoTriad-QA, OncoVLM exceeds MedGemma-4B by an average of 10.7 points when using MCQ accuracy and BERTScore-F1, with consistent gains across multiple-choice and open-ended questions under radiology-only, pathology-only, and all-available settings. These results demonstrate the benchmark's value for training and evaluating models for integrated cancer question answering.
Read the full abstract

All authors Ahnaf Munir, Dannong Wang, Michael W. McDonald, Mubarak Shah, Pegah Khosravi, Yu Tian

Computation and Language · Cross-listed from cs.IR

VLD-RAG: Agentic Vision-Language Retrieval-Augmented Generation for Long, Visually-Rich Multi-Page Documents

Seonok Kim

Visually-rich documents such as reports, slides, and manuals often distribute the evidence needed to answer a question across multiple pages, mixing text with layout cues, tables, charts, and figures. This work studies multimodal retrieval-augmented generation for question answering over such visually-rich long documents, where retrieval must select evidence pages that include both textual and visual signals. We present VLD-RAG, an agentic multimodal RAG framework for multi-page evidence retrieval and cross-page reasoning over long documents. VLD-RAG builds a page-preserving multimodal index that stores parsed text, page-level metadata, and dense visual representations, and uses a hybrid retrieval strategy that combines keyword-based sparse search with dense semantic queries to identify candidate sources and evidence pages. A verifier-guided agent workflow coordinates a Retrieval Agent, Answer Agent, and Validation Agent to broaden evidence coverage, detect missing citations, and refine retrieval requests when needed. We evaluate retrieval with Top-1 and Top-5 evidence-page accuracy and generation with generalized accuracy, and show that VLD-RAG improves both evidence-page retrieval and end-task question answering on visually-rich long-document benchmarks, including LongDocURL and MMLongBench-Doc, outperforming previous vision-based retrieval baselines. These findings highlight that coordinated agent verification and multimodal hybrid retrieval are crucial for reliable grounding when correct answers depend on evidence scattered across pages.
Read the full abstract

Computation and Language

From Lexicon to AI: A Structured-Data Pipeline for Specialized Conversational Systems in Low-Resource Languages

Siddhant Hitesh Mantri, Dhara Gorasiya, Malhar Kulkarni et al. (+1 more)

Low-resource languages face a critical challenge in AI development: creating specialized conversational systems without access to massive training corpora. We present a systematic methodology for transforming structured linguistic resources into specialized AI systems, demonstrating that expert-curated lexical databases can serve as effective foundations for conversational AI development. Our approach converts Hindi WordNet into 1.25 million diverse instruction-response pairs, fine-tunes a 12B-parameter language model using resource-efficient LoRA with 4-bit quantization. Evaluation through a Hindi language learning chatbot demonstrates that structured-knowledge-based systems achieve superior pedagogical effectiveness (91.0 vs. 79.4-83.6 for general-purpose models) while maintaining competitive semantic performance and exceptional consistency. The complete pipeline demonstrates a proof-of-concept methodology using Hindi for developing specialized AI systems for any languages with WordNet resources. This work addresses the critical gap in AI accessibility for low-resource languages, offering a practical alternative to corpus-intensive approaches and potentially enabling specialized AI development for the hundreds of languages with existing WordNet resources.
Read the full abstract

All authors Siddhant Hitesh Mantri, Dhara Gorasiya, Malhar Kulkarni, Pushpak Bhattacharya

Computation and Language · Cross-listed from cs.SI

Simulating Hate Speech Cascades with Multi-LLM Agents: Empirical Grounding, Modeling Fidelity, and Intervention Strategies

Fan Huang

Faithful modeling of hateful content propagation on online platforms remains an open problem for moderation research. Classical cascade models that do not explicitly represent the profile, community, and content factors associated with hateful-content propagation may yield moderation strategies that behave less effectively when deployed in real-world scenarios. Multi-agent large language model (LLM) systems can, in principle, make each reshare decision depend on the user's profile, the surrounding community, and the post's content, but it remains unclear whether this added flexibility actually reproduces real hateful cascades more faithfully than classical baselines. We study three hateful Bluesky cascades and a size-matched benign control. In the empirical Bluesky data, we found that: 97.4–99.7% of reposters take a hostile stance; toxicity-engagement homophily is higher on the diffusion tree than on the follower graph for hateful cascades; topology is star-like for the hateful cascades (most reposts come directly from the root) versus tree-like for the benign cascade (reposts propagate through multi-hop chains). In simulation, a multi-LLM-agent simulator reproduces the stance monoculture and the toxicity-delta direction. A structured ablation identifies agent heterogeneity as the leading fidelity factor, and amplifier targeting on dense networks yields 7.5–12.9% reduction at 5.7% benign collateral.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?

Jeanmely Rojas Nunez, Viraj Sawant, Nathan Allen et al. (+4 more)

Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit degrades under fine-tuning, and use it to compare RL and SFT on Qwen2.5-3B-Instruct adapted to scientific question-answering. We find a clear mechanistic trade-off: SFT adapts more rapidly to the target task but produces substantially greater circuit disruption and forgetting of prior capabilities, whereas RL preserves a larger fraction of the base circuit at the cost of slower task adaptation. These findings suggest that circuit preservation may help explain why RL is more robust to catastrophic forgetting. We released our code here: https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability.
Read the full abstract

All authors Jeanmely Rojas Nunez, Viraj Sawant, Nathan Allen, Nomgondalai Amgalanbaatar, Yannis Zongo, Vasu Sharma, Maheep Chaudhary