Peer Revenue Past citations are no guarantee of future returns

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Saturday 7 February

Computation and Language

Advantages of Domain Knowledge Injection for Legal Document Summarization: A Case Study on Summarizing Indian Court Judgments in English and Hindi

Debtanu Datta, Rajdeep Mukherjee, Adrijit Goswami et al. (+1 more)

Summarizing Indian legal court judgments is a complex task not only due to the intricate language and unstructured nature of the legal texts, but also since a large section of the Indian population does not understand the complex English in which legal text is written, thus requiring summaries in Indian languages. In this study, we aim to improve the summarization of Indian legal text to generate summaries in both English and Hindi (the most widely spoken Indian language), by injecting domain knowledge into diverse summarization models. We propose a framework to enhance extractive neural summarization models by incorporating domain-specific pre-trained encoders tailored for legal texts. Further, we explore the injection of legal domain knowledge into generative models (including Large Language Models) through continual pre-training on large legal corpora in English and Hindi. Our proposed approaches achieve statistically significant improvements in both English-to-English and English-to-Hindi Indian legal document summarization, as measured by standard evaluation metrics, factual consistency metrics, and legal domain-specific metrics. Furthermore, these improvements are validated through domain experts, demonstrating the effectiveness of our approaches.
Read the full abstract

All authors Debtanu Datta, Rajdeep Mukherjee, Adrijit Goswami, Saptarshi Ghosh

Computation and Language

When the Model Said 'No Comment', We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified

Gautam Siddharth Kashyap, Mark Dras and Usman Naseem

Large Language Models (LLMs) need to be in accordance with human values-being helpful, harmless, and honest (HHH)-is important for safe deployment. Existing works use Supervised Fine-Tuning (SFT) and Mixture-of-Experts (MoE) to align LLMs. However, these works face challenges in multi-objective settings, such as SFT leading to interference between conflicting objectives, while MoEs suffer from miscalibrated routing. We term this failure mode Axis Collapse, marked by (1) disjoint feature spaces causing catastrophic forgetting, and (2) unreliable inference from misrouted experts. To resolve this, we propose AlignX, a two-stage framework. Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys a MoCaE module that calibrates expert routing using fractal and natural geometry, improving inference reliability. AlignX achieves significant gains on Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty), with +171.5% win rate, +110.1% in truthfulness-informativeness, and 4.3% fewer safety violations. It also reduces latency and memory usage by over 35% compared to prior MoEs. Results across four LLMs validate its generalizability.
Read the full abstract

Computation and Language

Do Large Language Models Reflect Demographic Pluralism in Safety?

Usman Naseem, Gautam Siddharth Kashyap, Sushant Kumar Ray et al. (+3 more)

Large Language Model (LLM) safety is inherently pluralistic, reflecting variations in moral norms, cultural expectations, and demographic contexts. Yet, existing alignment datasets such as ANTHROPIC-HH and DICES rely on demographically narrow annotator pools, overlooking variation in safety perception across communities. Demo-SafetyBench addresses this gap by modeling demographic pluralism directly at the prompt level, decoupling value framing from responses. In Stage I, prompts from DICES are reclassified into 14 safety domains (adapted from BEAVERTAILS) using Mistral 7B-Instruct-v0.3, retaining demographic metadata and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based deduplication, yielding 43,050 samples. In Stage II, pluralistic sensitivity is evaluated using LLMs-as-Raters-Gemma-7B, GPT-4o, and LLaMA-2-7B-under zero-shot inference. Balanced thresholds (delta = 0.5, tau = 10) achieve high reliability (ICC = 0.87) and low demographic sensitivity (DS = 0.12), confirming that pluralistic safety evaluation can be both scalable and demographically robust.
Read the full abstract

All authors Usman Naseem, Gautam Siddharth Kashyap, Sushant Kumar Ray, Rafiq Ali, Ebad Shabbir, Abdullah Mohammad

Computation and Language

Efficient Post-Training Pruning of Large Language Models with Statistical Correction

Peiqi Yu, Jinhao Wang, Xinyi Sui et al. (+3 more)

Post-training pruning is an effective approach for reducing the size and inference cost of large language models (LLMs), but existing methods often face a trade-off between pruning quality and computational efficiency. Heuristic pruning methods are efficient but sensitive to activation outliers, while reconstruction-based approaches improve fidelity at the cost of heavy computation. In this work, we propose a lightweight post-training pruning framework based on first-order statistical properties of model weights and activations. During pruning, channel-wise statistics are used to calibrate magnitude-based importance scores, reducing bias from activation-dominated channels. After pruning, we apply an analytic energy compensation to correct distributional distortions caused by weight removal. Both steps operate without retraining, gradients, or second-order information. Experiments across multiple LLM families, sparsity patterns, and evaluation tasks show that the proposed approach improves pruning performance while maintaining computational cost comparable to heuristic methods. The results suggest that simple statistical corrections can be effective for post-training pruning of LLMs.
Read the full abstract

All authors Peiqi Yu, Jinhao Wang, Xinyi Sui, Nam Ling, Wei Wang, Wei Jiang

Computation and Language

TernaryLM: Memory-Efficient Language Modeling via Native 1.5-Bit Quantization with Adaptive Layer-wise Scaling

Nisharg Nargund and Priyesh Shukla

Large language models (LLMs) achieve remarkable performance but demand substantial computational resources, limiting deployment on edge devices and resource-constrained environments. We present TernaryLM, a 132M-parameter transformer trained natively with ternary quantization {-1, 0, +1} (log2(3) ~ 1.58-bit effective precision), achieving significant memory reduction without sacrificing language modeling capability. Unlike post-training quantization approaches that quantize pre-trained full-precision models, TernaryLM learns quantization-aware representations from scratch using straight-through estimators and adaptive per-layer scaling factors. Our experiments demonstrate: (1) validation perplexity of 58.42 on TinyStories with a cross-seed standard deviation of +/- 0.17 PPL, confirming stable optimization; (2) strong downstream transfer with 82.47% F1 on MRPC, surpassing DistilBERT despite using 55x less pretraining data; (3) 2.4x memory reduction (498 MB vs 1,197 MB for an FP32 model of identical architecture) with latency parity; and (4) an implicit regularization effect whereby the ternary constraint yields a train/val ratio of 1.05x versus 3.51x for the FP32 baseline, demonstrating that discrete weights prevent overfitting on small corpora. We provide layer-wise sparsity analysis revealing that middle transformer layers (L5-L9) achieve 60-62% quantization sparsity versus 45-55% for boundary layers, establishing an actionable design principle for non-uniform precision allocation. Our implementation and trained models are publicly available at https://github.com/1nisharg/TernaryLM-Memory-Efficient-Language-Modeling.
Read the full abstract

Computation and Language

ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for Vietnamese Healthcare Regulations

Long S. T. Nguyen, Quan M. Bui, Tin T. Ngo et al. (+3 more)

Question Answering (QA) over regulatory documents is inherently challenging due to the need for multihop reasoning across legally interdependent texts, a requirement that is particularly pronounced in the healthcare domain where regulations are hierarchically structured and frequently revised through amendments and cross-references. Despite recent progress in retrieval-augmented and graph-based QA methods, systematic evaluation in this setting remains limited, especially for low-resource languages such as Vietnamese, due to the lack of benchmark datasets that explicitly support multihop reasoning over healthcare regulations. In this work, we introduce the Vietnamese Healthcare Regulations-Multihop Reasoning Dataset (ViHERMES), a benchmark designed for multihop QA over Vietnamese healthcare regulatory documents. ViHERMES consists of high-quality question-answer pairs that require reasoning across multiple regulations and capture diverse dependency patterns, including amendment tracing, cross-document comparison, and procedural synthesis. To construct the dataset, we propose a controlled multihop QA generation pipeline based on semantic clustering and graph-inspired data mining, followed by large language model-based generation with structured evidence and reasoning annotations. We further present a graph-aware retrieval framework that models formal legal relations at the level of legal units and supports principled context expansion for legally valid and coherent answers. Experimental results demonstrate that ViHERMES provides a challenging benchmark for evaluating multihop regulatory QA systems and that the proposed graph-aware approach consistently outperforms strong retrieval-based baselines. The ViHERMES dataset and system implementation are publicly available at https://github.com/ura-hcmut/ViHERMES.
Read the full abstract

All authors Long S. T. Nguyen, Quan M. Bui, Tin T. Ngo, Quynh T. N. Vo, Dung N. H. Le, Tho T. Quan

Computation and Language

Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation

Geng Liu, Fei Zhu, Rong Feng et al. (+3 more)

Multi-turn conversation has emerged as a predominant interaction paradigm for Large Language Models (LLMs). Users often employ follow-up questions to refine their intent, expecting LLMs to adapt dynamically. However, recent research reveals that LLMs suffer a substantial performance drop in multi-turn settings compared to single-turn interactions with fully specified instructions, a phenomenon termed “Lost in Conversation” (LiC). While this prior work attributes LiC to model unreliability, we argue that the root cause lies in an intent alignment gap rather than intrinsic capability deficits. In this paper, we first demonstrate that LiC is not a failure of model capability but rather a breakdown in interaction between users and LLMs. We theoretically show that scaling model size or improving training alone cannot resolve this gap, as it arises from structural ambiguity in conversational context rather than representational limitations. To address this, we propose to decouple intent understanding from task execution through a Mediator-Assistant architecture. By utilizing an experience-driven Mediator to explicate user inputs into explicit, well-structured instructions based on historical interaction patterns, our approach effectively bridges the gap between vague user intent and model interpretation. Experimental results demonstrate that this method significantly mitigates performance degradation in multi-turn conversations across diverse LLMs.
Read the full abstract

All authors Geng Liu, Fei Zhu, Rong Feng, Changyi Ma, Shiqi Wang, Gaofeng Meng

Computation and Language · Cross-listed from cs.IR

High Fidelity Textual User Representation over Heterogeneous Sources via Reinforcement Learning

Rajat Arora, Ye Tao, Jianqiang Shen et al. (+7 more)

Effective personalization on large-scale job platforms requires modeling members based on heterogeneous textual sources, including profiles, professional data, and search activity logs. As recommender systems increasingly adopt Large Language Models (LLMs), creating unified, interpretable, and concise representations from heterogeneous sources becomes critical, especially for latency-sensitive online environments. In this work, we propose a novel Reinforcement Learning (RL) framework to synthesize a unified textual representation for each member. Our approach leverages implicit user engagement signals (e.g., clicks, applies) as the primary reward to distill salient information. Additionally, the framework is complemented by rule-based rewards that enforce formatting and length constraints. Extensive offline experiments across multiple LinkedIn products, one of the world's largest job platforms, demonstrate significant improvements in key downstream business metrics. This work provides a practical, labeling-free, and scalable solution for constructing interpretable user representations that are directly compatible with LLM-based systems.
Read the full abstract

All authors Rajat Arora, Ye Tao, Jianqiang Shen, Ping Liu, Muchen Wu, Qianqi Shen, Benjamin Le, Fedor Borisyuk, Jingwei Wu, Wenjing Zhang

Computation and Language

Beyond Accuracy: Risk-Sensitive Evaluation of Hallucinated Medical Advice

Savan Doshi

Large language models are increasingly being used in patient-facing medical question answering, where hallucinated outputs can vary widely in potential harm. However, existing hallucination standards and evaluation metrics focus primarily on factual correctness, treating all errors as equally severe. This obscures clinically relevant failure modes, particularly when models generate unsupported but actionable medical language. We propose a risk-sensitive evaluation framework that quantifies hallucinations through the presence of risk-bearing language, including treatment directives, contraindications, urgency cues, and mentions of high-risk medications. Rather than assessing clinical correctness, our approach evaluates the potential impact of hallucinated content if acted upon. We further combine risk scoring with a relevance measure to identify high-risk, low-grounding failures. We apply this framework to three instruction-tuned language models using controlled patient-facing prompts designed as safety stress tests. Our results show that models with similar surface-level behavior exhibit substantially different risk profiles and that standard evaluation metrics fail to capture these distinctions. These findings highlight the importance of incorporating risk sensitivity into hallucination evaluation and suggest that evaluation validity is critically dependent on task and prompt design.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs

Pengrui Han, Xueqiang Xu, Keyang Xuan et al. (+12 more)

Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.
Read the full abstract

All authors Pengrui Han, Xueqiang Xu, Keyang Xuan, Peiyang Song, Siru Ouyang, Runchu Tian, Yuqing Jiang, Cheng Qian, Pengcheng Jiang, Jiashuo Sun, Junxia Cui, Ming Zhong, Ge Liu, Jiawei Han, Jiaxuan You

Friday 6 February

Computation and Language · Cross-listed from cs.IR

UnWeaving the knots of GraphRAG – turns out VectorRAG is almost enough

Ryszard Tuora, Mateusz Galiński, Michał Godziszewski et al. (+4 more)

One of the key problems in Retrieval-augmented generation (RAG) systems is that chunk-based retrieval pipelines represent the source chunks as atomic objects, mixing the information contained within such a chunk into a single vector. These vector representations are then fundamentally treated as isolated, independent and self-sufficient, with no attempt to represent possible relations between them. Such an approach has no dedicated mechanisms for handling multi-hop questions. Graph-based RAG systems aimed to ameliorate this problem by modeling information as knowledge-graphs, with entities represented by nodes being connected by robust relations, and forming hierarchical communities. This approach however suffers from its own issues with some of them being: orders of magnitude increased componential complexity in order to create graph-based indices, and reliance on heuristics for performing retrieval. We propose UnWeaver, a novel RAG framework simplifying the idea of GraphRAG. UnWeaver disentangles the contents of the documents into entities which can occur across multiple chunks using an LLM. In the retrieval process entities are used as an intermediate way of recovering original text chunks hence preserving fidelity to the source material. We argue that entity-based decomposition yields a more distilled representation of original information, and additionally serves to reduce noise in the indexing, and generation process. Furthermore we experimentally show that on end to end QA evaluation VectorRAG performs better than standard GraphRAG and almost as good as current SOTA graph-based solutions, for a fraction of the cost.
Read the full abstract

All authors Ryszard Tuora, Mateusz Galiński, Michał Godziszewski, Michał Karpowicz, Mateusz Czyżnikiewicz, Adam Kozakiewicz, Tomasz Ziętkiewicz

Computation and Language

GhazalBench: Canonical Verse Access in LLMs across Persian Ghazals and Shakespearean Sonnets

Ghazal Kalhor and Yadollah Yaghoobzadeh

Persian poetry plays an active role in Iranian cultural practice, where verses by canonical poets such as Hafez and Saadi are frequently quoted, paraphrased, or completed from incomplete cues. Supporting such interactions requires language models to reliably access canonical verses from semantic and lexical information. We introduce GhazalBench, a benchmark for evaluating how large language models (LLMs) access the canonical surface forms of Persian ghazal under usage-grounded conditions. Unlike prior work that primarily treats memorization as a liability, GhazalBench studies settings in which access to exact wording is functionally important. The benchmark evaluates completion and recognition under varied semantic and lexical cues. Across proprietary and open-weight multilingual LLMs, exact completion remains challenging, while recognition is substantially stronger. Performance also varies across poets and model families. Parallel experiments on Shakespearean sonnets yield markedly higher completion for several models, consistent with differences in exposure to canonical texts. {Additional experiments suggest that post-training may reduce direct continuation-based access to canonical text.} Our findings motivate evaluation frameworks that distinguish production from recognition and assess access to culturally significant canonical texts. GhazalBench is available at https://github.com/kalhorghazal/GhazalBench.
Read the full abstract

Computation and Language

Attention's Gravitational Field:A Power-Law Interpretation of Positional Correlation

Edward Zhang

This paper explores the underlying principles of positional relationships and encodings within Large Language Models (LLMs) and introduces the concept of the Attention Gravitational Field (AGF). By decoupling positional encodings from semantic embeddings, we optimize the model architecture and achieve superior accuracy compared to prevailing encoding methods. Furthermore, we provide an in-depth analysis of AGF, demonstrating its intrinsic consistency with learning and stability curves, as well as its empirical alignment with Newton's Law of Universal Gravitation. By offering a rigorous theoretical exploration of these phenomena, this work represents a significant step toward interpreting the Attention mechanism and unlocks new possibilities for future research in model optimization and interpretability.
Read the full abstract

Computation and Language

TATRA: Training-Free Instance-Adaptive Prompting Through Rephrasing and Aggregation

Bartosz Dziuba, Kacper Kuchta, Paweł Batorski et al. (+2 more)

Large Language Models (LLMs) have improved substantially alignment, yet their behavior remains highly sensitive to prompt phrasing. This brittleness has motivated automated prompt engineering, but most existing methods (i) require a task-specific training set, (ii) rely on expensive iterative optimization to produce a single dataset-level prompt, and (iii) must be rerun from scratch for each new task. We introduce TATRA, a dataset-free prompting method that constructs instance-specific few-shot prompts by synthesizing on-the-fly examples to accompany a user-provided instruction. TATRA requires no labeled training data and avoids task-specific optimization loops, while retaining the benefits of demonstration-based prompting. Across standard text classification benchmarks, TATRA matches or improves over strong prompt-optimization baselines that depend on training data and extensive search. On mathematical reasoning benchmarks, TATRA achieves state-of-the-art performance on GSM8K and DeepMath, outperforming methods that explicitly optimize prompts on those tasks. Our results suggest that per-instance construction of effective in-context examples is more important than running long, expensive optimization loops to produce a single prompt per task. We will make all code publicly available upon acceptance of the paper. Code is available at https://github.com/BMD223/TATRA
Read the full abstract

All authors Bartosz Dziuba, Kacper Kuchta, Paweł Batorski, Przemysław Spurek, Paul Swoboda

Computation and Language

TTSR: Test-Time Self-Evolving via Reflection

Haoyang He, Zihua Rong, Yunjia Zhao et al. (+3 more)

Test-time training (TTT) adapts large language models (LLMs) during inference using only unlabeled test inputs. Existing methods, however, face two major bottlenecks on hard reasoning tasks: (1) lack of learnable samples, as self-generated pseudo-labels on difficult questions are often noisy and yield unstable rewards; and (2) inefficient exploration, as performance gains depend on repeatedly sampling many rollouts without explicit diagnosis of why previous attempts fail. We propose TTSR (Test-Time Self-Reflection), a self-evolving framework based on a reflect-then-synthesize paradigm. A single pretrained model alternates between a Student role and a Teacher role: the Student solves test questions and updates, while the Teacher analyzes failed trajectories and synthesizes targeted variant questions closer to the Student's capability frontier. TTSR further maintains a cross-iteration weakness memory and compiles persistent weaknesses into a lightweight strategy note prepended to subsequent Student inputs, so diagnostic knowledge can guide exploration and gradually fade as weaknesses are resolved. Experiments on challenging mathematical reasoning benchmarks show consistent test-time improvements, strong cross-backbone generalization, and transfer to general-domain reasoning tasks.
Read the full abstract

All authors Haoyang He, Zihua Rong, Yunjia Zhao, Lan Yang, Jian Chang, Honggang Zhang

Computation and Language

PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents

Ke Yang, Zixi Chen, Xuan He et al. (+6 more)

Long-term memory is essential for large language model (LLM) agents operating in complex environments, yet existing memory designs are either task-specific and non-transferable, or task-agnostic but less effective due to low task-relevance and context explosion from raw memory retrieval. We propose PlugMem, a task-agnostic plugin memory module that can be attached to arbitrary LLM agents without task-specific redesign. Motivated by the fact that decision-relevant information is concentrated as abstract knowledge rather than raw experience, we draw on cognitive science to structure episodic memories into a compact, extensible knowledge-centric memory graph that explicitly represents propositional and prescriptive knowledge. This representation enables efficient memory retrieval and reasoning over task-relevant knowledge, rather than verbose raw trajectories, and departs from other graph-based methods like GraphRAG by treating knowledge as the unit of memory access and organization instead of entities or text chunks. We evaluate PlugMem unchanged across three heterogeneous benchmarks (long-horizon conversational question answering, multi-hop knowledge retrieval, and web agent tasks). The results show that PlugMem consistently outperforms task-agnostic baselines and exceeds task-specific memory designs, while also achieving the highest information density under a unified information-theoretic analysis. Code and data are available at https://github.com/TIMAN-group/PlugMem.
Read the full abstract

All authors Ke Yang, Zixi Chen, Xuan He, Jize Jiang, Michel Galley, Chenglong Wang, Jianfeng Gao, Jiawei Han, ChengXiang Zhai

Computation and Language

Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task

Gaia Molinaro, Dave August, Danielle Perszyk et al. (+1 more)

Whether in agentic workflows, social studies, or chat settings, large language models (LLMs) are increasingly being asked to replace humans in choosing which goals to pursue, rather than completing predefined tasks. However, the assumption that LLMs accurately reflect human preferences for goal setting remains largely untested. We assess the validity of LLMs as proxies for human goal selection in a controlled, self-directed learning task borrowed from cognitive science. Across five models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Qwen3 32B, and Centaur), we find substantial divergence from human behavior. While people gradually explore and learn to achieve goals with diversity across individuals, most models exploit a single identified solution or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model. Chain-of-thought reasoning and persona steering provide limited improvements, and our conclusions hold across experimental settings. While they await confirmation in applied settings, these findings highlight the uniqueness of human goal selection and caution against its replacement with current models.
Read the full abstract

All authors Gaia Molinaro, Dave August, Danielle Perszyk, Anne G. E. Collins

Computation and Language

Fine-Tuning and Evaluating Conversational AI for Agricultural Advisory

Sanyam Singh, Naga Ganesh, Vineet Singh et al. (+10 more)

Large Language Models show promise for agricultural advisory, yet vanilla models exhibit unsupported recommendations, generic advice lacking specific, actionable detail, and communication styles misaligned with smallholder farmer needs. In high stakes agricultural contexts, where recommendation accuracy has direct consequences for farmer outcomes, these limitations pose challenges for responsible deployment. We present a hybrid LLM architecture that decouples factual retrieval from conversational delivery: supervised fine-tuning with LoRA on expert-curated GOLDEN FACTS (atomic, verified units of agricultural knowledge) optimizes fact recall, while a separate stitching layer transforms retrieved facts into culturally appropriate, safety-aware responses. Our evaluation framework, DG-EVAL, performs atomic fact verification (measuring recall, precision, and contradiction detection) against expert-curated ground truth rather than Wikipedia or retrieved documents. Experiments across multiple model configurations on crops and queries from Bihar, India show that fine-tuning on curated data substantially improves fact recall and F1, while maintaining high relevance. Using a fine-tuned smaller model achieves comparable or better factual quality at a fraction of the cost of frontier models. A stitching layer further improves safety subscores while maintaining high conversational quality. We release the farmerchat-prompts library to enable reproducible development of domain-specific agricultural AI.
Read the full abstract

All authors Sanyam Singh, Naga Ganesh, Vineet Singh, Lakshmi Pedapudi, Ritesh Kumar, SSP Jyothi, Archana Karanam, Waseem Pasha, Ekta Kumari, C. Yashoda, Mettu Vijaya Rekha Reddy, Shesha Phani Debbesa, Chandan Dash

Computation and Language

SE-Search: Self-Evolving Search Agent via Memory and Dense Reward

Jian Li, Yizhang Jin, Dongqi Liu et al. (+9 more)

Retrieval augmented generation (RAG) reduces hallucinations and factual errors in large language models (LLMs) by conditioning generation on retrieved external knowledge. Recent search agents further cast RAG as an autonomous, multi-turn information-seeking process. However, existing methods often accumulate irrelevant or noisy documents and rely on sparse reinforcement learning signals. We propose Self-Evolving Search, a Self-Evolving Search agent that improves online search behavior through three components, memory purification, atomic query training, and dense rewards. SE-Search follows a Think-Search-Memorize strategy that retains salient evidence while filtering irrelevant content. Atomic query training promotes shorter and more diverse queries, improving evidence acquisition. Dense rewards provide fine-grained feedback that speeds training. Experiments on single-hop and multi-hop question answering benchmarks show that SE-Search-3B outperforms strong baselines, yielding a $10.8$ point absolute improvement and a $33.8\%$ relative gain over Search-R1.\footnote{We will make the code and model weights publicly available upon acceptance.}
Read the full abstract

All authors Jian Li, Yizhang Jin, Dongqi Liu, Hang Ding, Jiafu Wu, Dongsheng Chen, Yunhang Shen, Yulei Qin, Ying Tai, Chengjie Wang, Xiaotong Yuan, Yabiao Wang

Computation and Language

From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG

Wenhao Wu, Zhentao Tang, Yafu Li et al. (+5 more)

Large Language Models (LLMs) exhibit high reasoning capacity in medical question-answering, but their tendency to produce hallucinations and outdated knowledge poses critical risks in healthcare fields. While Retrieval-Augmented Generation (RAG) mitigates these issues, existing methods rely on noisy token-level signals and lack the multi-round refinement required for complex reasoning. In this paper, we propose MA-RAG (Multi-Round Agentic RAG), a framework that facilitates test-time scaling for complex medical reasoning by iteratively evolving both external evidence and internal reasoning history within an agentic refinement loop. At each round, the agent transforms semantic conflict among candidate responses into actionable queries to retrieve external evidence, while optimizing history reasoning traces to mitigate long-context degradation. MA-RAG extends the self-consistency principle by leveraging the lack of consistency as a proactive signal for multi-round agentic reasoning and retrieval, and mirrors a boosting mechanism that iteratively minimizes the residual error toward a stable, high-fidelity medical consensus. Extensive evaluations across 7 medical Q&A benchmarks show that MA-RAG consistently surpasses competitive inference-time scaling and RAG baselines, delivering substantial +6.8 points on average accuracy over the backbone model. Our code is available at https://github.com/NJU-RL/MA-RAG.
Read the full abstract

All authors Wenhao Wu, Zhentao Tang, Yafu Li, Shixiong Kai, Mingxuan Yuan, Zhenhong Sun, Chunlin Chen, Zhi Wang

Computation and Language

One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models

Daniel Fein, Max Lamparth, Violet Xiang et al. (+2 more)

Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences. However, RM-based preference-tuning is vulnerable to reward hacking, whereby LM policies learn undesirable behaviors from flawed RMs. By systematically measuring biases in five high-quality RMs, including the state-of-the-art, we find that issues persist despite prior work with respect to length, sycophancy, and overconfidence. We also discover new issues related to bias toward model-specific “styles” and answer-order. We categorize RM failures as tractable or resistant to linear intervention and propose a simple post-hoc intervention to mitigate low-complexity biases that arise from spurious correlations. Our proposed mechanistic reward shaping reduces targeted biases without degrading reward quality and while using minimal labeled data. The method is extensible to new biases, model-internal, and generalizes out-of-distribution.
Read the full abstract

All authors Daniel Fein, Max Lamparth, Violet Xiang, Mykel J. Kochenderfer, Nick Haber

Computation and Language

A Zipf-preserving, long-range correlated surrogate for written language and other symbolic sequences

Marcelo A. Montemurro and Mirko Degli Esposti

Symbolic sequences such as written language and genomic DNA display characteristic frequency distributions and long-range correlations extending over many symbols. In language, this takes the form of Zipf's law for word frequencies together with persistent correlations spanning hundreds or thousands of tokens, while in DNA it is reflected in nucleotide composition and long-memory walks under purine-pyrimidine mappings. Existing surrogate models usually preserve either the frequency distribution or the correlation properties, but not both simultaneously. We introduce a surrogate model that retains both constraints: it preserves the empirical symbol frequencies of the original sequence and reproduces its long-range correlation structure, quantified by the detrended fluctuation analysis (DFA) exponent. Our method generates surrogates of symbolic sequences by mapping fractional Gaussian noise (FGN) onto the empirical histogram through a frequency-preserving assignment. The resulting surrogates match the original in first-order statistics and long-range scaling while randomising short-range dependencies. We validate the model on representative texts in English and Latin, and illustrate its broader applicability with genomic DNA, showing that base composition and DFA scaling are reproduced. This approach provides a principled tool for disentangling structural features of symbolic systems and for testing hypotheses on the origin of scaling laws and memory effects across language, DNA, and other symbolic domains.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

ACAR: Adaptive Complexity Routing for Multi-Model Ensembles with Auditable Decision Traces

Ramchand Kumaresan

We present ACAR (Adaptive Complexity and Attribution Routing), a measurement framework for studying multi-model orchestration under auditable conditions. ACAR uses self-consistency variance (sigma) computed from N=3 probe samples to route tasks across single-model, two-model, and three-model execution modes. The system is implemented on top of TEAMLLM, a deterministic execution substrate with immutable artifacts and complete decision traces. We evaluate ACAR on 1,510 tasks spanning four benchmarks: MathArena, Reasoning Gym, LiveCodeBench, and SuperGPQA, using Claude Sonnet 4, GPT-4o, and Gemini 2.0 Flash, producing more than 7,550 auditable runs. Results show that sigma-based routing achieves 55.6 percent accuracy, exceeding the two-model baseline of 54.4 percent while avoiding full ensembling on 54.2 percent of tasks. The routing mechanism is model-agnostic and requires no learned components. We also document negative results. First, retrieval augmentation reduced accuracy by 3.4 percentage points, as median retrieval similarity was only 0.167, demonstrating that experience injection without semantic alignment introduces noise rather than grounding. Second, when models agree on incorrect answers (sigma equals zero), no downstream ensemble can recover; this agreement-but-wrong failure mode is intrinsic to self-consistency and bounds achievable accuracy at approximately eight percentage points below full ensembling. Third, attribution estimates based on proxy signals such as response similarity and entropy showed weak correlation with ground-truth leave-one-out values, indicating that practical attribution requires explicit counterfactual computation. This work documents which assumptions fail in practice and provides falsifiable baselines for future research on routing, retrieval, and multi-model attribution.
Read the full abstract

Computation and Language · Cross-listed from cs.CY

How Well Can LLM Agents Simulate End-User Security and Privacy Attitudes and Behaviors?

Yuxuan Li, Leyang Li, Hao-Ping Lee et al. (+1 more)

A growing body of research assumes that large language model (LLM) agents can serve as proxies for how people form attitudes toward and behave in response to security and privacy (S&P) threats. If correct, these simulations could offer a scalable way to forecast S&P risks in products prior to deployment. We interrogate this assumption using SP-ABCBench, a new benchmark of 30 tests derived from validated S&P human-subject studies, which measures alignment between simulations and human-subjects studies on a 0-100 ascending scale, where higher scores indicate better alignment across three dimensions: Attitude, Behavior, and Coherence. Evaluating twelve LLMs, four persona construction strategies, and two prompting methods, we found that there remains substantial room for improvement: all models score between 50 and 64 on average. Newer, bigger, and smarter models do not reliably do better and sometimes do worse. Some simulation configurations, however, do yield high alignment: e.g., with scores above 95 for some behavior tests when agents are prompted to apply bounded rationality and weigh privacy costs against perceived benefits. We release SP-ABCBench to enable reproducible evaluation as methods improve.
Read the full abstract

All authors Yuxuan Li, Leyang Li, Hao-Ping Lee, Sauvik Das

Computation and Language · Cross-listed from cs.LG

A Case Study of Selected PTQ Baselines for Reasoning LLMs on Ascend NPU

Yuchen Luo, Fangyue Zhu, Ruining Zhou et al. (+4 more)

Post-Training Quantization (PTQ) is crucial for efficient model deployment, yet its effectiveness on Ascend NPU remains under-explored compared to GPU architectures. This paper presents a case study of representative PTQ baselines applied to reasoning-oriented models such as DeepSeek-R1-Distill-Qwen series (1.5B/7B/14B) and QwQ-32B. We evaluate four distinct algorithms, including AWQ, GPTQ, SmoothQuant, and FlatQuant, to cover the spectrum from weight-only compression to advanced rotation-based methods. Our empirical results reveal significant platform sensitivity. While 4-bit weight-only quantization proves viable for larger models, aggressive 4-bit weight-activation schemes suffer from layer-wise calibration instability on the NPU, leading to logic collapse in long-context reasoning tasks. Conversely, standard 8-bit quantization remains numerically stable. Furthermore, a real-world INT8 deployment demonstrates that although optimized kernels reduce latency, dynamic quantization overheads currently limit end-to-end acceleration. These findings offer a practical reference for the feasibility and limitations of deploying quantized reasoning models on Ascend NPU.
Read the full abstract

All authors Yuchen Luo, Fangyue Zhu, Ruining Zhou, Mingzhe Huang, Jian Zhu, Fanyu Fan, Wei Shao

Computation and Language · Cross-listed from cs.LG

Tethered Reasoning: Decoupling Entropy from Hallucination in Quantized LLMs via Manifold Steering

Craig Atkinson

Quantized language models face a fundamental dilemma: low sampling temperatures yield repetitive, mode-collapsed outputs, while high temperatures (T > 2.0) cause trajectory divergence and semantic incoherence. We present HELIX, a geometric framework that decouples output entropy from hallucination by tethering hidden-state trajectories to a pre-computed truthfulness manifold. HELIX computes a Unified Truth Score (UTS) combining token-level semantic entropy with Mahalanobis distance from the manifold. When UTS indicates trajectory divergence, graduated steering vectors redirect activations toward structurally coherent regions while affecting only 0.2-2.5% of tokens. On 4-bit quantized Granite 4.0 H Small (32B/9B active, hybrid Mamba-Transformer): GSM8K maintains 88.84% accuracy at T = 3.0 (2.81pp degradation from T = 0.5); MMLU maintains 72.49% across 14,042 questions (1.24pp degradation). This demonstrates that high-temperature hallucination is primarily trajectory divergence rather than semantic collapse. Notably, steering the sparse Transformer attention layers (~10% of layers) is sufficient to correct drift in the Mamba-2 state-space formulation. Geometric tethering reveals a previously-masked High-Entropy Creative Reservoir. At T > 2.0, steered outputs exhibit 5-20% idea duplication versus 70-80% at conservative settings. Cross-architecture validation (Qwen3-30B-A3B MOE) confirms this phenomenon is architecture-independent, with 46.7% higher unique concept generation. HELIX acts as a syntax tether, enabling exploration of semantic diversity without violating the logical backbone required for valid output. This enables Multi-Temperature Synthesis, generating 200% more unique concepts than single-temperature inference.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Robust Pre-Training of Medical Vision-and-Language Models with Domain-Invariant Multi-Modal Masked Reconstruction

Melika Filvantorkaman and Mohsen Piri

Medical vision-language models show strong potential for joint reasoning over medical images and clinical text, but their performance often degrades under domain shift caused by variations in imaging devices, acquisition protocols, and reporting styles. Existing multi-modal pre-training methods largely overlook robustness, treating it as a downstream adaptation problem. In this work, we propose Robust Multi-Modal Masked Reconstruction (Robust-MMR), a self-supervised pre-training framework that explicitly incorporates robustness objectives into masked vision-language learning. Robust-MMR integrates asymmetric perturbation-aware masking, domain-consistency regularization, and modality-resilience constraints to encourage domain-invariant representations. We evaluate Robust-MMR on multiple medical vision-language benchmarks, including medical visual question answering (VQA-RAD, SLAKE, VQA-2019), cross-domain image-text classification (MELINDA), and robust image-caption retrieval (ROCO). Robust-MMR achieves 78.9% cross-domain accuracy on VQA-RAD, outperforming the strongest baseline by 3.8 percentage points, and reaches 74.6% and 77.0% accuracy on SLAKE and VQA-2019, respectively. Under perturbed evaluation, Robust-MMR improves VQA-RAD accuracy from 69.1% to 75.6%. For image-text classification, cross-domain MELINDA accuracy increases from 70.3% to 75.2%, while retrieval experiments show a reduction in mean rank degradation from over 16 to 4.1 under perturbation. Qualitative results further demonstrate improved clinical reasoning for disease detection and structural abnormality assessment. These findings show that explicitly modeling robustness during pre-training leads to more reliable and transferable medical vision-language representations for real-world deployment.
Read the full abstract

Computation and Language · Cross-listed from stat.AP

Daily and Weekly Periodicity in Large Language Model Performance and Its Implications for Research

Paul Tschisgale and Peter Wulff

Large language models (LLMs) are increasingly used in research as both tools and objects of study. Much of this work assumes that LLM performance under fixed conditions (identical model snapshot, hyperparameters, and prompt) is time-invariant, meaning that average output quality remains stable over time; otherwise, reliability and reproducibility would be compromised. To test the assumption of time invariance, we conducted a longitudinal study of GPT-4o's average performance under fixed conditions. The LLM was queried to solve the same physics task ten times every three hours over approximately three months. Spectral (Fourier) analysis of the resulting time series revealed substantial periodic variability, accounting for about 20% of total variance. The observed periodic patterns are consistent with interacting daily and weekly rhythms. These findings challenge the assumption of time invariance and carry important implications for research involving LLMs.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

BRIDGE: Predicting Human Task Completion Time From Model Performance

Fengyuan Liu, Jay Gala, Nilaksh et al. (+3 more)

Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns a latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR's exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.
Read the full abstract

All authors Fengyuan Liu, Jay Gala, Nilaksh, Dzmitry Bahdanau, Siva Reddy, Hugo Larochelle

Computation and Language · Cross-listed from cs.AI

From Out-of-Distribution Detection to Hallucination Detection: A Geometric View

Litian Liu, Reza Pourreza, Yubing Jian et al. (+2 more)

Detecting hallucinations in large language models is a critical open problem with significant implications for safety and reliability. While existing hallucination detection methods achieve strong performance in question-answering tasks, they remain less effective on tasks requiring reasoning. In this work, we revisit hallucination detection through the lens of out-of-distribution (OOD) detection, a well-studied problem in areas like computer vision. Treating next-token prediction in language models as a classification task allows us to apply OOD techniques, provided appropriate modifications are made to account for the structural differences in large language models. We show that OOD-based approaches yield training-free, single-sample-based detectors, achieving strong accuracy in hallucination detection for reasoning tasks. Overall, our work suggests that reframing hallucination detection as OOD detection provides a promising and scalable pathway toward language model safety.
Read the full abstract

All authors Litian Liu, Reza Pourreza, Yubing Jian, Yao Qin, Roland Memisevic