Peer Revenue Publish or profit

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Tuesday 12 May

Computation and Language

When Emotion Becomes Trigger: Emotion-style dynamic Backdoor Attack Parasitising Large Language Models

Ziyu Liu, Tao Li, Tao Yang et al. (+4 more)

Data-poisoning backdoors pose a practical threat to the fine-tuning of large language models (LLMs). Most existing attacks bind an attacker-selected behavior to fixed tokens, phrases, scenarios, or syntactic structures. These discrete triggers provide concrete handles for defenses based on local token anomalies, pattern matching, or trigger recovery. We found that, under semantics-preserving rewriting, emotionally styled inputs form representation clusters distinct from their neutral counterparts. Meanwhile, de-emotionalised controls move back towards the neutral distribution. This observation motivates our method Paraesthesia, a dynamic backdoor attack that encodes its triggering condition in an emotional style. Paraesthesia maps target emotions into a valence–arousal space, rewrites a small subset of clean samples, and retains semantically faithful rewrites for fine-tuning. Across instruction-following and classification tasks evaluated on four major LLMs, Paraesthesia achieves an attack success rate(ASR) above 98.25%, while introducing only negligible degradation to clean utility across the vast majority of model-task setups. Surface feature controls and paired de-emotionalization experiments demonstrate that no examined token-level cue can fully account for the triggered behavior. ASR remains high after word-level filtering, sample clustering, and subsequent clean-update procedures, whereas a white-box decoding defense with access to a task-aligned clean reference provides a distinct mitigation path. These findings identify emotional style as a concrete backdoor trigger surface beyond fixed lexical and syntactic patterns.\par\smallskip \noindent \textcolor{red}{WARNING: \textnormal{This paper contains risk-related textual content.}}
Read the full abstract

All authors Ziyu Liu, Tao Li, Tao Yang, Tianjie Ni, Xiaolong Lan, Wengang Ma, Junjiang He

Computation and Language · Cross-listed from cs.LG

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information

Guobin Shen, Xiang Cheng, Chenxiao Zhao et al. (+4 more)

On-policy self-distillation, where a student is pulled toward a copy of itself conditioned on privileged context (e.g., a verified solution or feedback), offers a promising direction for advancing reasoning capability without a stronger external teacher. Yet in math reasoning the gains are inconsistent, even when the same approach succeeds elsewhere. A pointwise mutual information analysis traces the failure to the privileged context itself: it inflates the teacher's confidence on tokens already implied by the solution (structural connectives, verifiable claims) and deflates it on deliberation tokens ("Wait", "Let", "Maybe") that drive multi-step search. We propose Anti-Self-Distillation (AntiSD), which ascends a divergence between student and teacher rather than descending it: this reverses the per-token sign and yields a naturally bounded advantage in one step. An entropy-triggered gate disables the term once the teacher entropy collapses, completing a drop-in replacement for default self-distillation. Across five models from 4B to 30B parameters on math reasoning benchmarks, AntiSD reaches the GRPO baseline's accuracy in 2 to 10x fewer training steps and improves final accuracy by up to 11.5 points. AntiSD opens a path to scalable self-improvement, where a language model bootstraps its own reasoning through its training signal.
Read the full abstract

All authors Guobin Shen, Xiang Cheng, Chenxiao Zhao, Lei Huang, Jindong Li, Dongcheng Zhao, Xing Yu

Computation and Language

PRISM: A Geometric Risk Bound for Decomposing Drift into Scale, Shape, and Head

Chieh-Yen Lin and Shao-Hua Sun

A single base LLM now comes with dozens of post-training variants, quantized, LoRA-adapted, or distilled, and each has to be checked before release. Existing evaluations provide only a partial picture: benchmark scores and likelihood screens say that a variant has degraded, similarity scores such as CKA and SVCCA say how its features moved, and nothing connects the two. We connect them with one structural fact and one design choice: the prediction head is linear, so feature geometry reaches the loss, and we compare the two feature sets through an orthogonal map, which leaves the geometry being measured unchanged. From these we derive PRISM, a closed-form upper bound on the cross-entropy risk gap between a target model and a proxy variant, and prove that it splits exactly into three measurable axes: scale, shape, and head. Each axis names a failure mode and where to intervene: low-bit quantization distorts shape and, at the lowest bit-widths, also inflates activation scale; quantizing the output embedding inflates the head term, which dominates the bound at high bit-widths. Because the shape term is differentiable, the same geometry becomes a regularizer that curbs catastrophic forgetting more than experience replay. Across two model families and five benchmarks, PRISM ranks quantized and fine-tuned variants from a single forward pass at mean Spearman above 0.8, and eight reference sequences already recover the ranking that the risk gap itself needs over a hundred to reach.
Read the full abstract

Computation and Language

DiffScore: Text Evaluation Beyond Autoregressive Likelihood

Wen Lai, Yingli Shen, Dingnan Jin et al. (+4 more)

Autoregressive language models are widely used for text evaluation, however, their left-to-right factorization introduces positional bias, i.e., early tokens are scored with only leftward context, conflating architectural asymmetry with true text quality. We propose masked reconstruction as an alternative paradigm, where every token is scored using full bidirectional context. We introduce DiffScore, an evaluation framework built on Masked Large Diffusion Language Models. By measuring text recoverability across continuous masking rates, DiffScore eliminates positional bias and naturally establishes an evaluation hierarchy from local fluency to global coherence. We further provide diagnostic tools unavailable to autoregressive frameworks: multi-timestep quality profiles that decompose scores across masking rates, and bidirectional PMI decomposition that disentangles fluency from faithfulness. Experiments across ten benchmarks show that DiffScore consistently outperforms autoregressive baselines in both zero-shot and fine-tuned settings. The code is released at: https://github.com/wenlai-lavine/DiffScore.
Read the full abstract

All authors Wen Lai, Yingli Shen, Dingnan Jin, Qing Cui, Jun Zhou, Maosong Sun, Alexander Fraser

Computation and Language

Efficient LLM-based Advertising via Model Compression and Parallel Verification

Wenxin Dong, Chang Gao, Guanghui Yu et al. (+9 more)

Large language models (LLMs) have shown remarkable potential in advertising scenarios such as ad creative generation and targeted advertising. However, deploying LLMs in real-time advertising systems poses significant challenges due to their high inference latency and computational cost. In this paper, we propose an Efficient Generative Targeting framework that integrates adaptive group quantization, layer-adaptive hierarchical sparsification, and prefix-tree parallel verification to accelerate LLM inference while preserving generation quality. Extensive experiments on two real-world advertising scenarios demonstrate that our framework achieves significant speedup with acceptable quality degradation, making it operationally viable for practical deployments.
Read the full abstract

All authors Wenxin Dong, Chang Gao, Guanghui Yu, Xuewu Jiao, Mingqing Hu, Qiang Fu, Peng Xu, Penghui Wei, Hui Xu, Yue Xing, Shuanglong Li, Lin Liu

Computation and Language

Ada-MK: Adaptive MegaKernel Optimization via Automated DAG-based Search for LLM Inference

Wenxin Dong, Mingqing Hu, Guanghui Yu et al. (+7 more)

When large language models (LLMs) serve real-time inference in commercial online advertising systems, end-to-end latency must be strictly bounded to the millisecond range. Yet every token generated during the decode phase triggers thousands of kernel launches, and kernel launch overhead alone can account for 14.6% of end-to-end inference time. MegaKernel eliminates launch overhead and inter-operator HBM round-trips by fusing multiple operators into a single persistent kernel. However, existing MegaKernel implementations face a fundamental tension between portability and efficiency on resource-constrained GPUs such as NVIDIA Ada: hand-tuned solutions are tightly coupled to specific architectures and lack portability, while auto-compiled approaches introduce runtime dynamic scheduling whose branch penalties are unacceptable in latency-critical settings. We observe that under a fixed deployment configuration, the optimal execution path of a MegaKernel is uniquely determined, and runtime dynamic decision-making can be entirely hoisted to compile time. Building on this insight, we propose Ada-MK: (1) a three-dimensional shared-memory constraint model combined with K-dimension splitting that reduces peak shared memory usage by 50%; (2) MLIR-based fine-grained DAG offline search that solidifies the optimal execution path, completely eliminating runtime branching; and (3) a heterogeneous hybrid inference engine that embeds MegaKernel as a plugin into TensorRT-LLM, combining high-throughput Prefill with low-latency Decode. On an NVIDIA L20, Ada-MK improves single-batch throughput by up to 23.6% over vanilla TensorRT-LLM and 50.2% over vLLM, achieving positive gains across all tested scenarios–the first industrial deployment of MegaKernel in a commercial online advertising system.
Read the full abstract

All authors Wenxin Dong, Mingqing Hu, Guanghui Yu, Qiang Fu, Peng Xu, Hui Xu, Yue Xing, Xuewu Jiao, Shuanglong Li, Lin Liu

Computation and Language

BitLM: Unlocking Multi-Token Language Generation with Bitwise Continuous Diffusion

Shaobin Zhuang, Yuang Ai, Jiaming Han et al. (+7 more)

Autoregressive language models generate text one token at a time, yet natural language is inherently structured in multi-token units, including phrases, n-grams, and collocations that carry meaning jointly. This one-token bottleneck limits both the expressiveness of the model during pre-training and its throughput at inference time. Existing remedies such as speculative decoding or diffusion-based language models either leave the underlying bottleneck intact or sacrifice the causal structure essential to language modeling. We propose BitLM, a language model that represents each token as a fixed-length binary code and employs a lightweight diffusion head to denoise multiple tokens in parallel within each block. Crucially, BitLM preserves left-to-right causal attention across blocks while making joint lexical decisions within each block, combining the reliability of autoregressive modeling with the parallelism of iterative refinement. By replacing the large-vocabulary softmax with bitwise denoising, BitLM reframes token generation as iterative commitment in a compact binary space, enabling more efficient pre-training and substantially faster inference without altering the causal foundation that makes language models effective. Our results demonstrate that the one-token-at-a-time paradigm is not a fundamental requirement but an interface choice, and that changing it can yield a stronger and faster language model. We hope BitLM points toward a promising direction for next-generation language model architectures.
Read the full abstract

All authors Shaobin Zhuang, Yuang Ai, Jiaming Han, Xiaohui Li, Huaibo Huang, Xiangyu Yue, Xuefeng Hu, Kun Xu, Yali Wang, Hao Chen

Computation and Language

Three Regimes of Context-Parametric Conflict: A Predictive Framework and Empirical Validation

Pruthvinath Jeripity Venkata

The literature on how large language models handle conflict between their training knowledge and a contradicting document presents a persistent empirical contradiction: some studies find models stubbornly retain their trained answers, ignoring provided documents nearly half the time, while others find models readily defer to the document, following context approximately 96% of the time. We argue these contradictions dissolve once one recognises that prior experiments have studied three qualitatively distinct processing situations without distinguishing them. We propose a three-regime framework: Regime 1 (single-source updating, dominant predictor: evidence coherence), Regime 2 (competitive integration, dominant predictor: parametric certainty), and Regime 3 (task-appropriate selection, dominant predictor: task knowledge requirement). We formalise a distinction between parametric strength (exposure frequency) and parametric uniqueness (encoding consistency), showing empirically that these are orthogonal dimensions (r = -0.002, p = .97) with strength as the operative predictor in stable factual domains. We validate the framework across Claude Sonnet 4.6, GPT-5.5, Gemini 2.5 Flash, Llama 4 Maverick, and DeepSeek V3 using 9,970 API calls in three experimental phases. GEE logistic regression confirms the predicted Regime 2 certainty gradient for all five models (beta = -0.38 to -0.50, all p <= .013, BH-FDR corrected). A Regime 3 ablation shows task framing alone flips context-following from near-100% (contextual knowledge condition) to 6-71% (parametric knowledge condition), with all five models significant (p < .001). The certainty gradient is robust to multinomial outcome modeling, sensitivity analyses for hedging responses, and FDR correction.
Read the full abstract

Computation and Language

Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting

Cheng Wang, Qin Liu, Wenxuan Zhou et al. (+1 more)

Group Relative Policy Optimization (GRPO) has emerged as a promising approach for improving the reasoning capabilities of large language models. However, it struggles to effectively balance the tradeoff between exploration and exploitation during training, often resulting in suboptimal performance. Motivated by the theoretical insight that changes in entropy are governed by the covariance between token probabilities and their corresponding advantages, we propose a hyperparameter-free, covariance-weighted optimization method that dynamically down-weights extreme token-level updates via a Gaussian kernel. This approach automatically reduces the instability caused by exploration-exploitation trade-off while preserving informative learning signals. Extensive empirical evaluations show that our approach improves downstream performance across reasoning benchmarks compared with GRPO, and effectively stablizes entropy as training progresses.
Read the full abstract

All authors Cheng Wang, Qin Liu, Wenxuan Zhou, Muhao Chen

Computation and Language

Checkup2Action: A Multimodal Clinical Check-up Report Dataset for Patient-Oriented Action Card Generation

Sike Xiang, Shuang Chen, Kevin Qinghong Lin et al. (+4 more)

Routine clinical check-up reports combine laboratory measurements, physiological assessments, imaging findings and visually structured information, but rarely tell patients what to do next. Translating them into follow-up actions requires models to connect evidence across pages, tables and modalities, identify clinically relevant issues and communicate next steps without unsupported diagnostic or treatment claims. Yet this report-to-action capability remains poorly benchmarked. We introduce C2A, a dataset and benchmark for generating structured Action Cards from multimodal check-up reports, together with Checkup2Action, a constrained workflow for the task. C2A contains 2,000 de-identified real-world reports covering physical examinations, laboratory tests, cardiovascular assessments and imaging evidence. Each card specifies one issue, its priority, recommended department, follow-up window, patient-facing explanation and questions for clinicians. Our evaluation measures issue coverage and precision, priority consistency, department and timing accuracy, action complexity, usefulness, readability and safety. Experiments across general-purpose and medical large language models show that no model performs best on every dimension. Clinical experts judged 84% of evaluated outputs fully reasonable and effective. Removing safety constraints increased problem recall from 0.527 to 0.825, but produced more diagnostic overstatement. C2A therefore supports systematic study of the balance between coverage, actionable guidance and safe patient communication.
Read the full abstract

All authors Sike Xiang, Shuang Chen, Kevin Qinghong Lin, Jialin Yu, Yijia Sun, Philip Torr, Amir Atapour-Abarghouei

Computation and Language · Cross-listed from cs.AI

Controllable User Simulation

Guy Tennenholtz, Ofer Meshi, Amir Globerson et al. (+3 more)

Using offline datasets to evaluate conversational agents often fails to cover rare scenarios or to support testing new policies. This has motivated the use of controllable user simulators for targeted, counterfactual evaluation, typically implemented by prompting or fine-tuning large language models. In this work, we formalize controllable simulation as a causal inference problem. By bridging natural language evaluation with off-policy evaluation methodology, we show that the standard practice of training simulators via supervised fine-tuning on post-hoc trajectory labels yields a structurally biased model. Specifically, these labels are inextricably coupled to the data-generating behavior policy, injecting a look-ahead bias that breaks causal consistency. Furthermore, we prove that under policy shift this failure causes the variance of evaluation metrics to explode geometrically, a phenomenon we term controllability collapse. To restore causal consistency, we establish theoretical conditions for accurate simulation and propose practical training mitigations: a priori controls, step-wise dynamic controls, and direct policy-conditioned learning. Empirical evaluation confirms that while standard global controls distort conversational distributions and collapse behavioral diversity, our causally grounded simulators eliminate look-ahead bias, preserve natural variance, and exhibit robust zero-shot generalization to unseen agent behaviors.
Read the full abstract

All authors Guy Tennenholtz, Ofer Meshi, Amir Globerson, Uri Shalit, Jihwan Jeong, Craig Boutilier

Computation and Language · Cross-listed from cs.AI

AutoLLMResearch: Training Research Agents for Automating LLM Experiment Configuration - Learning from Cheap, Optimizing Expensive

Taicheng Guo, Nitesh V. Chawla, Olaf Wiest et al. (+1 more)

Effectively configuring scalable large language model (LLM) experiments, spanning architecture design, hyperparameter tuning, and beyond, is crucial for advancing LLM research, as poor configuration choices can waste substantial computational resources and prevent models from realizing their full potential. Prior automated methods are designed for low-cost settings where repeated trial and error is feasible, but scalable LLM experiments are too expensive for such extensive iteration. To our knowledge, no work has addressed the automation of high-cost LLM experiment configurations, leaving this problem labor-intensive and dependent on expert intuition. Motivated by this gap, we propose AutoLLMResearch, an agentic framework that mimics how human researchers learn generalizable principles from low-fidelity experiments and extrapolate to efficiently identify promising configurations in expensive LLM settings. The core challenge is how to enable an agent to learn, through interaction with a multi-fidelity experimental environment that captures the structure of the LLM configuration landscape. To achieve this, we propose a systematic framework with two key components: 1) LLMConfig-Gym, a multi-fidelity environment encompassing four critical LLM experiment tasks, supported by over one million GPU hours of verifiable experiment outcomes; 2) A structured training pipeline that formulates configuration research as a long-horizon Markov Decision Process and accordingly incentivizes cross-fidelity extrapolation reasoning. Extensive evaluation against diverse strong baselines on held-out experiments demonstrates the effectiveness, generalization, and interpretability of our framework, supporting its potential as a practical and general solution for scalable real-world LLM experiment automation.
Read the full abstract

All authors Taicheng Guo, Nitesh V. Chawla, Olaf Wiest, Xiangliang Zhang

Computation and Language

A Study on Hidden Layer Distillation for Large Language Model Pre-Training

Maxime Guigon, Lucas Dixon and Michaël E. Sander

Knowledge Distillation (KD) is a critical tool for training Large Language Models (LLMs), yet the majority of research focuses on approaches that rely solely on output logits, neglecting semantic information in the teacher's intermediate representations. While Hidden Layer Distillation (HLD) showed potential for encoder architectures, its application to decoder-only pre-training at scale remains largely unexplored. Through compute-controlled experiments, we benchmark HLD against logit-based KD and self-supervised baselines with Gemma3 3.4B as teacher and 123M and 735M students trained on up to 168B tokens from the C4 dataset. Our experiments show that HLD does not consistently outperform standard KD on downstream evaluation tasks. Nevertheless, we show that HLD can yield a systematic perplexity gain over KD across all shared-hyperparameter configurations, suggesting that a latent signal can be extracted, but a breakthrough may be needed for it to play a more significant role in LLM pre-training.
Read the full abstract

Computation and Language

Robust Biomedical Publication Type and Study Design Classification with Knowledge-Guided Perturbations

Shufan Ming, Joe D. Menke, Neil R. Smalheiser et al. (+1 more)

Accurately and consistently indexing biomedical literature by publication type and study design is essential for supporting evidence synthesis and knowledge discovery. Prior work on automated publication type and study design indexing has primarily focused on expanding label coverage, enriching feature representations, and improving in-domain accuracy, with evaluation typically conducted on data drawn from the same distribution as training. Although pretrained biomedical language models achieve strong performance under these settings, models optimized for in-domain accuracy may rely on superficial lexical or dataset-specific cues, resulting in reduced robustness under distributional shift. In this study, we introduce an evaluation framework based on controlled semantic perturbations to assess the robustness of a publication type classifier and investigate robustness-oriented training strategies that combine entity masking and domain-adversarial training to mitigate reliance on spurious topical correlations. Our results show that the commonly observed trade-off between robustness and in-domain accuracy can be mitigated when robustness objectives are designed to selectively suppress non-task-defining features while preserving salient methodological signals. We find that these improvements arise from two complementary mechanisms: (1) increased reliance on explicit methodological cues when such cues are present in the input, and (2) reduced reliance on spurious domain-specific topical features. These findings highlight the importance of feature-level robustness analysis for publication type and study design classification and suggest that refining masking and adversarial objectives to more selectively suppress topical information may further improve robustness. Data, code, and models are available at: https://github.com/ScienceNLP-Lab/MultiTagger-v2/tree/main/ICHI
Read the full abstract

All authors Shufan Ming, Joe D. Menke, Neil R. Smalheiser, Halil Kilicoglu

Computation and Language

StoicLLM: Preference Optimization for Philosophical Alignment in Small Language Models

Ishmam Khan, Sindhuja Thogarrati and Shuo Zhang

While large language models excel at factual adaptation, their ability to internalize nuanced philosophical frameworks under severe data constraints remains underexplored. We investigate this by specializing small LLMs on micro-datasets of foundational Stoic texts using preference optimization (ORPO, AlphaPO). Evaluated via a multi-model critic bank, our results show that just 300 high-fidelity examples can induce strong alignment with inward-facing Stoic virtues, closely approaching few-shot prompting while freeing the context window. Critically, however, all models, including few-shot baselines, exhibit a persistent failure on Stoicism's outward-facing cosmopolitan duties, pointing to a representational limitation of small models that micro-dataset adaptation alone cannot overcome.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning

Zihao Han, Tiangang Zhang, Huaibin Wang et al. (+1 more)

On-policy self-distillation has become a strong recipe for LLM reasoning, where a privileged teacher supervises the student's own rollouts while conditioning on the reference solution. A design choice shared by nearly all such methods, however, has gone unquestioned: the teacher always sees the full reference reasoning. We argue that this default itself is part of the problem and identify a teacher-side exposure mismatch: when the teacher conditions on reasoning far beyond the student's current competence, the resulting token targets become too strong to absorb. A controlled fixed-exposure sweep makes this concrete on two fronts: 1) full exposure is not reliably the best choice, and 2) student-teacher mismatch grows monotonically as the teacher sees more privileged reasoning. This motivates treating teacher exposure not as a fixed hyperparameter but as a learnable training-time control variable. We therefore propose Adaptive Teacher Exposure for Self-Distillation (ATESD). ATESD models the reveal ratio with a lightweight Beta-policy controller conditioned on compact training-state statistics, and uses one sampled exposure for a short hold window of student updates. To make this exposure controller learnable, we optimize it with a discounted learning-progress reward that scores each held decision by its effect on the student's future improvement rather than its immediate loss change, addressing the delayed credit assignment induced by on-policy distillation. Experiments on AIME 24, AIME 25, and HMMT 25 across Qwen3-{1.7B, 4B, 8B} show that ATESD consistently outperforms competitive self-distillation and RL baselines, improving over OPSD by +0.95, +2.05, and +2.33 Average@12 points respectively, and establishing adaptive teacher exposure as an effective new axis for reasoning self-distillation.
Read the full abstract

All authors Zihao Han, Tiangang Zhang, Huaibin Wang, Yilun Sun

Computation and Language · Cross-listed from cs.CR

Mobius Injection: From One-Shot Prompt Injection to Sustained Resource Exhaustion in AI Agent Infrastructure

Zi Liang, Ronghua Li, Yanyun Wang et al. (+2 more)

Large Language Model (LLM) agents have emerged as key intermediaries, orchestrating complex interactions between human users and a wide range of digital services and LLM infrastructures. While prior research has extensively examined the security of LLMs and agents in isolation, the systemic risk of the agent acting as a disruptive "hub" within the user-agent-service chain remains largely overlooked. In this work, we expose a novel threat paradigm by introducing Mobius Injection, a sophisticated attack that weaponizes autonomous agents into zombie nodes to launch what we define as Agent-based and -Oriented DDoS (AbO-DDoS) attacks. Mobius Injection composes known indirect prompt injection and persistent agent components through a distinct structural mechanism named Semantic Closure, which induces sustained recursive execution of agent components through a single injection event. We demonstrate that this attack is lightweight for the adversary, can evade traditional volumetric DDoS monitors and input filters calibrated for low false-positive rates, and is highly configurable, allowing for surgical targeting of specific environments or model providers. To evaluate the real-world impact, we conduct extensive experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs. Our results demonstrate that Mobius Injection achieves substantial attack success across diverse tasks, driving single-node call amplification up to 51.0x. Under coordinated multi-node pressure, each infected node still drains about 123k–127k tokens on a shared five-GPU backend; benign request latency rises by abou 23x–24x, and coordinated infections eventually saturate shared capacity until benign traffic is effectively denied service.
Read the full abstract

All authors Zi Liang, Ronghua Li, Yanyun Wang, Qingqing Ye, Haibo Hu

Computation and Language

Agent-BRACE: Decoupling Beliefs from Actions in Long-Horizon Tasks via Verbalized State Uncertainty

Joykirat Singh, Zaid Khan, Archiki Prasad et al. (+5 more)

Large language models (LLMs) are increasingly deployed on long-horizon tasks in partially observable environments, where they must act while inferring and tracking a complex environment state over many steps. This leads to two challenges: partial observability requires maintaining uncertainty over unobserved world attributes, and long interaction history causes context to grow without bound, diluting task-relevant information. A principled solution to both challenges is a belief state: a posterior distribution over environment states given past observations and actions, which compactly encodes history for decision making regardless of episode length. In LLM agents, however, the open-ended nature of text makes it unclear how to represent such a distribution. Therefore, we introduce Agent-BRACE: Agent Belief state Representation via Abstraction and Confidence Estimation, a method that decouples an LLM agent into a belief state model and a policy model, jointly optimized via reinforcement learning. The belief state model produces a structured approximation of the belief distribution: a set of atomic natural language claims about the environment, each annotated with an ordinal verbalized certainty label ranging from certain to unknown. The policy model conditions on this compact, structured approximate belief rather than the full history, learning to select actions under explicit uncertainty. Across long-horizon, partially observable embodied language environments, Agent-BRACE achieves an average absolute improvement of +14.5% (Qwen2.5-3B-Instruct) and +5.3% (Qwen3-4B-Instruct), outperforming strong RL baselines while maintaining a near-constant context window independent of episode length. Further analysis shows that the learned belief becomes increasingly calibrated over the course of an episode as evidence accumulates.
Read the full abstract

All authors Joykirat Singh, Zaid Khan, Archiki Prasad, Justin Chih-Yao Chen, Akshay Nambi, Hyunji Lee, Elias Stengel-Eskin, Mohit Bansal

Computation and Language

Freeze Deep, Train Shallow: Interpretable Layer Allocation for Continued Pre-Training

Yu-Hang Wu, Qin-Yuan Liu, Qiu-Yang Zhao et al. (+3 more)

Selective layer-wise updates are essential for low-cost continued pre-training of Large Language Models (LLMs), yet determining which layers to freeze or train remains an empirical black-box problem due to the lack of interpretable guidance. To address this issue, we propose LayerTracer, an architecture-agnostic diagnostic framework that reveals the evolution patterns of layer-wise representations and stability by locating task execution positions and quantifying layer sensitivity. Analysis results reveal that deep layers act as critical regions for task execution and maintain high stability against disruptive updates. Guided by this finding, we conduct three controlled continued pre-training trials to compare diverse freeze-train strategies, demonstrating that training shallow layers while freezing deep layers consistently outperforms full-parameter fine-tuning and the opposite allocation on both C-Eval and CMMLU benchmarks. We further present a hybrid model case study, which validates that placing high-quality pre-trained modules in deep layers effectively preserves inherent knowledge of the model. This work delivers a low-cost and interpretable solution for resource-constrained teams, offering actionable guidance for layer-wise parameter allocation in continued pre-training and hybrid model construction.
Read the full abstract

All authors Yu-Hang Wu, Qin-Yuan Liu, Qiu-Yang Zhao, Bo Jiang, Jiang-Feng Yang, Qing-Wei Cong

Computation and Language · Cross-listed from cs.LG

MaskTab: Scalable Masked Tabular Pretraining with Scaling Laws and Distillation for Industrial Classification

Bo Zheng, Yudong Chen, Zihua Xiong et al. (+4 more)

Tabular data forms the backbone of high-stakes decision systems in finance, healthcare, and beyond. Yet industrial tabular datasets are inherently difficult: high-dimensional, riddled with missing entries, and rarely labeled at scale. While foundation models have revolutionized vision and language, tabular learning still leans on handcrafted features and lacks a general self-supervised framework. We present MaskTab, a unified pre-training framework designed specifically for industrial-scale tabular data. MaskTab encodes missing values via dedicated learnable tokens, enabling the model to distinguish structural absence from random dropout. It jointly optimizes a hybrid supervised pre-training scheme–utilizing a twin-path architecture to reconcile masked reconstruction with task-specific supervision–and an MoE-augmented loss that adaptively routes features through specialized subnetworks. On industrial-scale benchmarks, it achieves +5.04% AUC and +8.28% KS over prior art under rigorous scaling. Moreover, its representations distill effectively into lightweight models, yielding +2.55% AUC and +4.85% KS under strict latency and interpretability constraints, while improving robustness to distribution shifts. Our work demonstrates that tabular data admits a foundation-model treatment–when its structural idiosyncrasies are respected.
Read the full abstract

All authors Bo Zheng, Yudong Chen, Zihua Xiong, Shuai Fang, Peidong He, Yang Yang, Sheng Guo

Computation and Language · Cross-listed from cs.LG

fg-expo: Frontier-guided exploration-prioritized policy optimization via adaptive kl and gaussian curriculum

Mingxiong Lin, Zhangquan Gong, Maowen Tang et al. (+6 more)

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, with Group Relative Policy Optimization (GRPO) serving as the dominant algorithm. We identify two overlooked inefficiencies inherent in GRPO. First, a fixed KL coefficient overly restricts policy exploration at moments when the model needs to diverge significantly from the reference policy. Second, uniform question sampling overlooks that moderately difficult problems produce the most informative gradient signals. We propose FG-ExPO, short for Frontier-Guided Exploration-Prioritized Policy Optimization, which integrates two lightweight components. Accuracy-Conditioned KL Scaling (AKL) adjusts the KL penalty strength through a smooth nonlinear function of batch average accuracy, loosening the constraint when the model performs poorly and strengthening it when the model achieves satisfactory results. Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at a moderate accuracy level around 0.5, focusing model training on its learning frontier. We conduct evaluations on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base across six mainstream mathematical reasoning benchmarks. Experimental results demonstrate that FG-ExPO consistently outperforms vanilla GRPO. It delivers an absolute improvement of 13.34 on the AIME 2025 pass@32 metric, rising from 63.33 percent to 76.67 percent, and obtains an average pass@32 gain of 2.66 on the 8B model. The substantially larger performance gains observed on pass@32 compared to pass@1 verify that FG-ExPO enlarges the model's effective exploration space under a fixed inference budget.
Read the full abstract

All authors Mingxiong Lin, Zhangquan Gong, Maowen Tang, Qian Li, Chuangchuang Wang, Jian Ma, Sutian Huang, Kai Tang, Haonan Lu

Computation and Language · Cross-listed from cs.AI

AcuityBench: Evaluating Clinical Acuity Identification and Uncertainty Alignment

Robin Linzmayer, Georgianna Lin, Di Coneybeare et al. (+22 more)

We introduce AcuityBench, a benchmark for evaluating whether language models identify the appropriate urgency of care from user medical presentations. Existing health benchmarks emphasize medical question answering, broad health interactions, or narrow workflow-specific triage tasks, but they do not offer a unified evaluation of acuity identification across these settings. AcuityBench addresses this gap by harmonizing five public datasets spanning user conversations, online forum posts, clinical vignettes, and patient portal messages under a shared four-level acuity framework ranging from home monitoring to immediate emergency care. The benchmark contains 914 cases, including 697 consensus cases for standard accuracy evaluation and 217 physician-confirmed ambiguous cases for uncertainty-aware evaluation. It supports two complementary task formats: explicit four-way classification in a QA setting, and free-form conversational responses evaluated with a rubric-based judge anchored to the same framework. Across 12 frontier proprietary and open-weight models, we find substantial variation in clear-case acuity accuracy and error direction. Comparing task formats reveals a systematic tradeoff: conversational responses reduce over-triage but increase under-triage relative to QA, especially in higher-acuity cases. In ambiguous cases, no model closely matches the distribution of physician judgments, and model predictions are more concentrated than expert clinical uncertainty. We also compare expert and model adjudication on a subset of maximally ambiguous cases, using those cases to examine the role of clinical uncertainty in label disagreement. Together, these results position acuity identification as a distinct safety-critical capability and show that AcuityBench enables systematic comparison and stress-testing of how well models guide users to the right level of care in real-world health use.
Read the full abstract

All authors Robin Linzmayer, Georgianna Lin, Di Coneybeare, Jason Chu, Trudi Cloyd, Manish Garg, Miles Gordon, Elizabeth Hartofilis, Benjamin Hong, Ashraf Hussain, Eugene Y. Kim, Oluchi Iheagwara King, Ross McCormack, Erica Olsen, John K. Riggins Jr, Mustafa N. Rasheed, Dana L. Sacco, Vinay Saggar, Osman R. Sayan, Amit Shembekar, Janice Shin-Kim, Wendy W. Sun, Bernard P. Chang, David Kessler, Noémie Elhadad

Computation and Language

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

Dean Light, Michael Theologitis, Kshitish Ghate et al. (+7 more)

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning – an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.
Read the full abstract

All authors Dean Light, Michael Theologitis, Kshitish Ghate, Shuyue Stella Li, Benjamin Newman, Chirag Shah, Aylin Caliskan, Pang Wei Koh, Dan Suciu, Yulia Tsvetkov

Computation and Language

An Empirical Study of Automating Agent Evaluation

Kang Zhou, Sangmin Woo, Haibo Ding et al. (+15 more)

Agent evaluation requires assessing complex multi-step behaviors involving tool use and intermediate reasoning, making it costly and expertise-intensive. A natural question arises: can frontier coding assistants reliably automate this evaluation process? Our study shows that simply prompting coding assistants is insufficient for this task. Without domain-specific evaluation knowledge, frontier coding assistants achieve only a 30% execution success rate and produce over-engineered evaluations averaging 12+ metrics per agent, indicating that strong coding ability does not automatically translate to reliable agent evaluation. We introduce EvalAgent, an AI assistant that automates the end-to-end agent evaluation pipeline. EvalAgent encodes evaluation domain expertise as evaluation skills (procedural instructions, reusable code and templates, and dynamically retrieved API documentation) that compose into a trace-based pipeline producing complete evaluation artifacts including metrics, executable code, and reports. To systematically assess generated evaluations, we introduce a meta-evaluation framework alongside AgentEvalBench, a benchmark comprising 20 agents, each paired with evaluation requirements and test scenarios. We further propose the Eval@1 metric to measure whether generated evaluation code both executes and yields meaningful results on the first run. Our experiments show that EvalAgent produces focused evaluations, improving Eval@1 from 17.5% to 65%, and achieving 79.5% human expert preference over baseline approaches. Further ablation studies show that evaluation skills are critical for handling complex evaluation: removing them causes Eval@1 to drop significantly from 65% to 30%.
Read the full abstract

All authors Kang Zhou, Sangmin Woo, Haibo Ding, Kiran Ramnath, Subramanian Chidambaram, Aosong Feng, Vinayak Arannil, Muhyun Kim, Ishan Singh, Darren Wang, Zhichao Xu, Megha Gandhi, Nirmal Prabhu, Soumya Smruti Mishra, Smeet Dhakecha, Vivek Singh, Gouri Pandeshwar, Lin Lee Cheong

Computation and Language · Cross-listed from cs.LG

Test-Time Compute for Frozen Embedding Models through Agentic Program Search

Han Xiao

Test-time compute is widely believed to benefit only large reasoning models, leaving small models with nothing to gain. We argue the opposite for dense retrieval, since modern small embedding models are distilled or adapted from large language model backbones and can inherit their latent test-time-compute potential. We ask how much retrieval quality a frozen embedding model gains at inference alone, with no auxiliary model and no parameters trained at deployment. An agentic loop in which a large language model writes programs over a frozen encoder API explores 144 candidates and yields twelve Pareto-optimal programs that trade inference compute for quality across cost ratios from $c{=}1.2$ to $14.7$, every one improving nDCG@10 on all 14 discovery tasks. The programs use no trainable parameters and recover classical retrieval primitives, among them reciprocal rank fusion, the Fisher linear discriminant, Rocchio pseudo-relevance feedback, and sentence-level MaxSim. Applied unmodified to nineteen held-out tasks and three unseen encoder families, a single fixed program improves the majority of tasks, with a positive median $\Delta$nDCG@10 and a 54 to 57% win-rate at $c{\ge}4$, and the gains are largest on encoder families never seen during discovery. A matched-budget learned projection head trained on the same tasks does not transfer this way, improving in-domain retrieval by $+0.20$ to $+0.25$ nDCG@10 yet falling below baseline on every held-out encoder. Small embedding models therefore inherit usable test-time-compute potential, and a frozen encoder converts inference compute into retrieval gains that transfer to new corpora and encoders with no per-domain labels.
Read the full abstract

Computation and Language · Cross-listed from cs.CV

PresentAgent-2: Towards Generalist Multimodal Presentation Agents

Wei Wu, Ziyang Xu, Zeyu Zhang et al. (+2 more)

Presentation generation is moving beyond static slide creation toward end-to-end presentation video generation with research grounding, multimodal media, and interactive delivery. We introduce PresentAgent-2, an agentic framework for generating presentation videos from user queries. Given an open-ended user query and a selected presentation mode, PresentAgent-2 first summarizes the query into a focused topic and performs deep research over presentation-friendly sources to collect multimodal resources, including relevant text, images, GIFs, and videos. It then constructs presentation slides, generates mode-specific scripts, and composes slides, audio, and dynamic media into a complete presentation video. PresentAgent-2 supports three independent presentation modes within a unified framework: Single Presentation, which generates a single-speaker narrated presentation video; Discussion, which creates a multi-speaker presentation with structured speaker roles, such as for asking guiding questions, explaining concepts, clarifying details, and summarizing key points; and Interaction, which independently supports answering audience questions grounded in the generated slides, scripts, retrieved evidence, and presentation context. To evaluate these capabilities, we build a multimodal presentation benchmark covering single presentation, discussion, and interaction scenarios, with task-specific evaluation criteria for content quality, media relevance, dynamic media use, dialogue naturalness, and interaction grounding. Overall, PresentAgent-2 extends presentation generation from document-dependent slide creation to query-driven, research-grounded presentation video generation with multimodal media, dialogue, and interaction. Code: https://github.com/AIGeeksGroup/PresentAgent-2. Website: https://aigeeksgroup.github.io/PresentAgent-2.
Read the full abstract

All authors Wei Wu, Ziyang Xu, Zeyu Zhang, Yang Zhao, Hao Tang

Computation and Language

Large Language Models for Causal Relations Extraction in Social Media: A Validation Framework for Disaster Intelligence

Ujun Jeong, Saketh Vishnubhatla, Bohan Jiang et al. (+3 more)

During disasters, extracting causal relations from social media can strengthen situational awareness by identifying factors linked to casualties, physical damage, infrastructure disruption, and cascading impacts. However, disaster-related posts are often informal, fragmented, and context-dependent, and they may describe personal experiences rather than explicit causal relations. In this work, we examine whether Large Language Models (LLMs) can effectively extract causal relations from disaster-related social media posts. To this end, we (1) propose an expert-grounded evaluation framework that compares LLM-generated causal graphs with reference graphs derived from disaster-specific reports and (2) assess whether the extracted relations are supported by post-event evidence or instead reflect model priors. Our findings highlight both the potential and risks of using LLMs for causal relation extraction in disaster decision-support systems.
Read the full abstract

All authors Ujun Jeong, Saketh Vishnubhatla, Bohan Jiang, Andre Harrison, Adrienne Raglin, Huan Liu

Monday 11 May

Computation and Language

TopoGuard: Graph Theory Based Defenses Against Split-Knowledge Attacks on RAG

Chahana Dahal and Zuobin Xiong

Production Retrieval Augmented Generation (RAG) systems rely on aggregating multiple external documents to answer complex queries. However, the retrieved documents introduce a new threat surface that can be exploited to launch split-knowledge attacks. In this attack, the adversary injects documents that are individually benign but create false associations when combined and fed to language models. This paper shows that the new attack is structurally invisible to existing per-document filters, like LlamaGuard. To address this issue in RAG, this work introduces TopoGuard, a family of graph theory-based methods specifically targeting the split-knowledge attacks by building a semantic similarity graph from retrieved documents and detecting contexts with malicious topology. Grounded on the theoretical analysis, the TopoGuard family has been proven to be effective and robust even with noisy inputs. Extensive experiments are conducted on two retrieval datasets and compared with multiple baseline methods. Specifically, the TopoGuard-$\lambda_2$+Entity catches 21$\times$ more attacks than LlamaGuard-2-8B at 1% FPR (32.6% vs 1.5% recall) on the HotpotQA dataset. Compared with production RAG detection systems using large language models, the proposed TopoGuard variants run efficiently at sub-millisecond latency and stay robust under adaptive adversaries and benign cross-domain queries.
Read the full abstract

Computation and Language

Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models

Phongsakon Mark Konrad, Toygar Tanyel and Serkan Ayvaz

Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption. A checkpoint can appear fixed under evaluation-style prompts while the same behavior persists under ordinary-use prompts. Output scores reveal this mismatch but do not locate it. We investigate whether the distinction is encoded in a stable internal site and introduce an audit that fits a paired activation contrast at a path-patching-informed mid-depth window, then modifies the resulting coordinate on held-out prompts. The intervention reduces the evaluation-to-deployment gap by more than $0.12$ in ten of twelve model–behavior settings and closes it (post-intervention gap below $0.06$) in four, across four full-matrix instruction-tuned model instances; among the eight settings with $n{\geq}120$ paired questions, six reduce and one closes. A fifth model supports localization and edit-provenance checks, and in the twelve settings deployment-framed rates change by at most $6.1$pp. On retrained checkpoints, random directions at the same dose do not reproduce the effect, and off-window controls do so only in part. The strongest literal evidence for an evaluation-versus-deployment distinction comes from sandbagging, whose prompt framing is explicitly evaluative; sycophancy and refusal test framing-contingent task distinctions. The two weak cells, both sycophancy, show where a single-coordinate audit at a fixed window is not sufficient. The audit is a non-adaptive, white-box diagnostic for controlled fine-tuned checkpoints, not a training-time defense, a guarantee of deployment safety, or evidence about naturally arising failures.
Read the full abstract

Computation and Language

Making Open-Source Text LLM Watermarks Durable Against Merging

Luisa Scharff, Thibaud Gloaguen, Robin Staab et al. (+1 more)

Open-source LLMs (OSMs)arereaching near state-of-the-art performance, prompting prior works to trace the text they generate by embedding text watermarking algorithms directly into their weights. Yet, OSMs are subject to post-training modifications, which has been shown to remove the watermark. Model merging in particular, a prominent method used for combining expert knowledge and preventing catastrophic forgetting, strongly removes such OSM watermarks. A key question is how to enable OSM watermarks that survive subsequent merging. In this work, we show for the first time how to design an OSM watermark that is durable against model merging. We propose Merge-Adversarial Training, an adversarial training algorithm to distill text watermarks into model weights while being robust to subsequent model merging. Our approach consistently outperforms all baselines (e.g. with SLERP up to +51 percentage points (pp) TPR@1%FPR with +25 pp on average) while preserving downstream capabilities. We also for the first time evaluate OSM watermarks against realistic merge scenarios, representing common use-cases such as combining expert capabilities or preventing catastrophic forgetting, and with 3 prominent merging algorithms. More broadly, our findings suggest that adversarial training is a reliable approach for increasing OSM watermark durability against post-training modifications.
Read the full abstract

All authors Luisa Scharff, Thibaud Gloaguen, Robin Staab, Martin Vechev