Peer Revenue Play credits. Real papers.

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Thursday 15 January

Computation and Language

Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment

Cameron Tice, Puria Radmard, Samuel Ratnam et al. (+3 more)

Pretraining corpora contain extensive discourse about AI systems, yet the causal influence of this discourse on downstream alignment remains poorly understood. If prevailing descriptions of AI behaviour are predominantly negative, LLMs may internalise corresponding behavioural priors, giving rise to self-fulfilling misalignment. This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse. We find that discussion of AI contributes to misalignment. Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour. Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%. We consider this evidence of self-fulfilling alignment. These effects are dampened, but persist through post-training. Our findings establish the study of how pretraining data shapes alignment priors, or alignment pretraining, as a complement to post-training. We recommend practitioners consider pretraining for alignment alongside capabilities. We share our models, data, and evaluations at AlignmentPretraining.ai.
Read the full abstract

All authors Cameron Tice, Puria Radmard, Samuel Ratnam, Andy Kim, David Africa, Kyle O'Brien

Computation and Language

What Gets Activated: Uncovering Domain and Driver Experts in MoE Language Models

Guimin Hu, Meng Li, Qiwei Peng et al. (+3 more)

Most interpretability work focuses on layer- or neuron-level mechanisms in Transformers, leaving expert-level behavior in MoE LLMs underexplored. Motivated by functional specialization in the human brain, we analyze expert activation by distinguishing domain and driver experts. In this work, we study expert activation in MoE models across three public domains and address two key questions: (1) which experts are activated, and whether certain expert types exhibit consistent activation patterns; and (2) how tokens are associated with and trigger the activation of specific experts. To answer these questions, we introduce entropy-based and causal-effect metrics to assess whether an expert is strongly favored for a particular domain, and how strongly expert activation contributes causally to the model's output, thus identify domain and driver experts, respectively. Furthermore, we explore how individual tokens are associated with the activation of specific experts. Our analysis reveals that (1) Among the activated experts, some show clear domain preferences, while others exert strong causal influence on model performance, underscoring their decisive roles. (2) tokens occurring earlier in a sentence are more likely to trigger the driver experts, and (3) adjusting the weights of domain and driver experts leads to significant performance gains across all three models and domains. These findings shed light on the internal mechanisms of MoE models and enhance their interpretability.
Read the full abstract

All authors Guimin Hu, Meng Li, Qiwei Peng, Lijie Hu, Boyan Xu, Ruichu Cai

Computation and Language

ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback

Yutao Mou, Zhangchi Xue, Lijun Li et al. (+4 more)

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains under-explored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action-attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, we introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65 percent on average and improves benign task completion by approximately 10 percent under prompt injection attacks.
Read the full abstract

All authors Yutao Mou, Zhangchi Xue, Lijun Li, Peiyang Liu, Shikun Zhang, Wei Ye, Jing Shao

Computation and Language

Role-Playing Agents Driven by Large Language Models: Current Status, Challenges, and Future Trends

Ye Wang, Jiaxing Chen and Hongjiang Xiao

In recent years, with the rapid advancement of large language models (LLMs), role-playing language agents (RPLAs) have emerged as a prominent research focus at the intersection of natural language processing (NLP) and human-computer interaction. This paper systematically reviews the current development and key technologies of RPLAs, delineating the technological evolution from early rule-based template paradigms, through the language style imitation stage, to the cognitive simulation stage centered on personality modeling and memory mechanisms. It summarizes the critical technical pathways supporting high-quality role-playing, including psychological scale-driven character modeling, memory-augmented prompting mechanisms, and motivation-situation-based behavioral decision control. At the data level, the paper further analyzes the methods and challenges of constructing role-specific corpora, focusing on data sources, copyright constraints, and structured annotation processes. In terms of evaluation, it collates multi-dimensional assessment frameworks and benchmark datasets covering role knowledge, personality fidelity, value alignment, and interactive hallucination, while commenting on the advantages and disadvantages of methods such as human evaluation, reward models, and LLM-based scoring. Finally, the paper outlines future development directions of role-playing agents, including personality evolution modeling, multi-agent collaborative narrative, multimodal immersive interaction, and integration with cognitive neuroscience, aiming to provide a systematic perspective and methodological insights for subsequent research.
Read the full abstract

Computation and Language · Cross-listed from cs.MA

TopoDIM: One-shot Topology Generation of Diverse Interaction Modes for Multi-Agent Systems

Rui Sun, Jie Ding, Chenghua Gong et al. (+5 more)

Optimizing communication topology in LLM-based multi-agent system is critical for enabling collective intelligence. Existing methods mainly rely on spatio-temporal interaction paradigms, where the sequential execution of multi-round dialogues incurs high latency and computation. Motivated by the recent insights that evaluation and debate mechanisms can improve problem-solving in multi-agent systems, we propose TopoDIM, a framework for one-shot Topology generation with Diverse Interaction Modes. Designed for decentralized execution to enhance adaptability and privacy, TopoDIM enables agents to autonomously construct heterogeneous communication without iterative coordination, achieving token efficiency and improved task performance. Experiments demonstrate that TopoDIM reduces total token consumption by 46.41% while improving average performance by 1.50% over state-of-the-art methods. Moreover, the framework exhibits strong adaptability in organizing communication among heterogeneous agents. Code is available at: https://github.com/Sundiasy/TopoDIM.
Read the full abstract

All authors Rui Sun, Jie Ding, Chenghua Gong, Tianjun Gu, Yihang Jiang, Juyuan Zhang, Liming Pan, Linyuan Lü

Computation and Language

Skill-Aware Data Selection and Fine-Tuning for Data-Efficient Reasoning Distillation

Lechen Zhang, Yunxiang Zhang, Wei Hu et al. (+1 more)

Large reasoning models such as DeepSeek-R1 and their distilled variants achieve strong performance on complex reasoning tasks. Yet, distilling these models often demands large-scale data for supervised fine-tuning (SFT), motivating the pursuit of data-efficient training methods. To address this, we propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model's weaker skills, and (2) Skill-aware fine-tuning, which encourages explicit skill decomposition during problem solving. With only 1,000 training examples selected from a 100K teacher-generated corpus, our method surpasses random SFT baselines by +1.6% on Qwen3-4B and +1.4% on Qwen3-8B across five mathematical reasoning benchmarks. Further analysis confirms that these gains concentrate on skills emphasized during training, highlighting the effectiveness of skill-centric training for efficient reasoning distillation.
Read the full abstract

All authors Lechen Zhang, Yunxiang Zhang, Wei Hu, Lu Wang

Computation and Language

SIN-Bench: Tracing Native Evidence Chains in Long-Context Multimodal Scientific Interleaved Literature

Yiming Ren, Junjie Wang, Yuxin Meng et al. (+11 more)

Evaluating whether multimodal large language models truly understand long-form scientific papers remains challenging: answer-only metrics and synthetic "Needle-In-A-Haystack" tests often reward answer matching without requiring a causal, evidence-linked reasoning trace in the document. We propose the "Fish-in-the-Ocean" (FITO) paradigm, which requires models to construct explicit cross-modal evidence chains within native scientific documents. To operationalize FITO, we build SIN-Data, a scientific interleaved corpus that preserves the native interleaving of text and figures. On top of it, we construct SIN-Bench with four progressive tasks covering evidence discovery (SIN-Find), hypothesis verification (SIN-Verify), grounded QA (SIN-QA), and evidence-anchored synthesis (SIN-Summary). We further introduce "No Evidence, No Score", scoring predictions when grounded to verifiable anchors and diagnosing evidence quality via matching, relevance, and logic. Experiments on eight MLLMs show that grounding is the primary bottleneck: Gemini-3-pro achieves the best average overall score (0.573), while GPT-5 attains the highest SIN-QA answer accuracy (0.767) but underperforms on evidence-aligned overall scores, exposing a gap between correctness and traceable support.
Read the full abstract

All authors Yiming Ren, Junjie Wang, Yuxin Meng, Yihang Shi, Zhiqiang Lin, Ruihang Chu, Yiran Xu, Ziming Li, Yunfei Zhao, Zihan Wang, Yu Qiao, Ruiming Tang, Minghao Liu, Yujiu Yang

Computation and Language · Cross-listed from cs.AI

Matrix as Plan: Structured Logical Reasoning with Feedback-Driven Replanning

Ke Chen, Jiandian Zeng, Zihao Peng et al. (+3 more)

As knowledge and semantics on the web grow increasingly complex, enhancing Large Language Models (LLMs)' comprehension and reasoning capabilities has become particularly important. Chain-of-Thought (CoT) prompting has been shown to enhance the reasoning capabilities of LLMs. However, it still falls short on logical reasoning tasks that rely on symbolic expressions and strict deductive rules. Neuro-symbolic methods address this gap by enforcing formal correctness through external solvers. Yet these solvers are highly format-sensitive, and small instabilities in model outputs can lead to frequent processing failures. The LLM-driven approaches avoid parsing brittleness, but they lack structured representations and process-level error-correction mechanisms. To further enhance the logical reasoning capabilities of LLMs, we propose MatrixCoT, a structured CoT framework with a matrix-based plan. Specifically, we normalize and type natural language expressions and attach explicit citation fields, and introduce a matrix-based planning method to preserve global relations among steps. The plan thus becomes a verifiable artifact and execution becomes more stable. For verification, we also add a feedback-driven replanning mechanism. Under semantic-equivalence constraints, it identifies omissions and defects, rewrites and compresses the dependency matrix, and produces a more trustworthy final answer. Experiments on five logical-reasoning benchmarks and five LLMs show that, without relying on external solvers, MatrixCoT enhances both the robustness and interpretability of LLMs when tackling complex symbolic reasoning tasks, while maintaining competitive performance.
Read the full abstract

All authors Ke Chen, Jiandian Zeng, Zihao Peng, Guo Li, Guangxue Zhang, Tian Wang

Computation and Language

CALM-IT: Enhancing Synthetic Motivational Interviewing Dialogue Quality with Conversational Dynamics

Viet Cuong Nguyen, Nhi Yen Nguyen, Kristin A. Candan et al. (+5 more)

Conversational dynamics describe how rapport, emotion, and goals change over a dialogue. These changes are important in therapeutic conversations, where they can influence when and how therapists intervene. However, large language model (LLM)-based synthetic therapeutic dialogue generators rarely track how these properties evolve together, which may limit the quality of longer conversations. We introduce CALM-IT, a dual-agent framework that tracks client states and therapist-side inferences to guide strategy selection and utterance generation. We evaluate synthetic Motivational Interviewing (MI) dialogues using automated Motivational Interviewing Treatment Integrity (MITI) ratings, conversational metrics, and comparisons across dialogue lengths, with ratings from four licensed psychologists providing support for key findings. Across 8,232 dialogues covering 686 grounded scenarios, four frameworks, and three target lengths, CALM-IT achieves the highest mean scores on all four MITI 4.2 global ratings and shows the smallest aggregate change in quality between short- and long-form conditions. CALM-IT also achieves the highest simulated client redirection acceptance rate (64.3%), with its strongest redirections followed by increased change talk and reduced sustain talk. Ablation experiments show that removing conversational dynamics components reduces dialogue quality. Together, these results suggest that explicitly modeling conversational dynamics can improve the quality and consistency of long-form synthetic MI dialogues.
Read the full abstract

All authors Viet Cuong Nguyen, Nhi Yen Nguyen, Kristin A. Candan, Mary Conlon, Vanessa Rumie, Kristen Risola, Michael L. Birnbaum, Munmun De Choudhury

Computation and Language

Is MT Ready for the Next Crisis or Pandemic?

Vipasha Bansal, William D. Lewis, Nicole Mihaylova et al. (+2 more)

Communication in times of crisis is essential. However, there is often a mismatch between the language of governments, aid providers, doctors, and those to whom they are providing aid. It is reasonable to turn to commercial MT tools in these scenarios. But how effective are these tools for translating to and from low resource languages, particularly in the crisis or medical domain? In this study, we evaluate 5 commercial systems used for MT (Google Translate, Microsoft Translator, GPT, Gemini and Claude) against the TICO-19 dataset, which contains pandemic-related sentences from a large set of high-priority languages spoken by communities most likely to be adversely affected in the next pandemic. We assess the current degree of "readiness" for this next pandemic based on the usability of the output translations.
Read the full abstract

All authors Vipasha Bansal, William D. Lewis, Nicole Mihaylova, Elizabeth Brown, Benjamin Pong

Computation and Language

Deriving Character Logic from Storyline as Codified Decision Trees

Letian Peng, Kun Zhou, Longfei Yun et al. (+2 more)

Role-playing (RP) agents rely on behavioral profiles to act consistently across diverse narrative contexts, yet existing profiles are largely unstructured, non-executable, and weakly validated, leading to brittle agent behavior. We propose Codified Decision Trees (CDT), a data-driven framework that induces an executable and interpretable decision structure from large-scale narrative data. CDT represents behavioral profiles as a tree of conditional rules, where internal nodes correspond to validated scene conditions and leaves encode grounded behavioral statements, enabling deterministic retrieval of context-appropriate rules at execution time. The tree is learned by iteratively inducing candidate scene-action rules, validating them against data, and refining them through hierarchical specialization, yielding profiles that support transparent inspection and principled updates. Across multiple benchmarks, CDT substantially outperforms human-written profiles and prior profile induction methods on $85$ characters across $16$ artifacts, indicating that codified and validated behavioral representations lead to more reliable agent grounding.
Read the full abstract

All authors Letian Peng, Kun Zhou, Longfei Yun, Yupeng Hou, Jingbo Shang

Computation and Language · Cross-listed from cs.LG

Sparse-RL: Breaking the Memory Wall in LLM Reinforcement Learning via Stable Sparse Rollouts

Sijia Luo, Xiaokang Zhang, Yuxuan Hu et al. (+6 more)

Reinforcement Learning (RL) has become essential for eliciting complex reasoning capabilities in Large Language Models (LLMs). However, the substantial memory overhead of storing Key-Value (KV) caches during long-horizon rollouts acts as a critical bottleneck, often prohibiting efficient training on limited hardware. While existing KV compression techniques offer a remedy for inference, directly applying them to RL training induces a severe policy mismatch, leading to catastrophic performance collapse. To address this, we introduce Sparse-RL empowers stable RL training under sparse rollouts. We show that instability arises from a fundamental policy mismatch among the dense old policy, the sparse sampler policy, and the learner policy. To mitigate this issue, Sparse-RL incorporates Sparsity-Aware Rejection Sampling and Importance-based Reweighting to correct the off-policy bias introduced by compression-induced information loss. Experimental results show that Sparse-RL reduces rollout overhead compared to dense baselines while preserving the performance. Furthermore, Sparse-RL inherently implements sparsity-aware training, significantly enhancing model robustness during sparse inference deployment. The corresponding training data and code are publicly available on the repository.
Read the full abstract

All authors Sijia Luo, Xiaokang Zhang, Yuxuan Hu, Bohan Zhang, Ke Wang, Jinbo Su, Mengshu Sun, Lei Liang, Jing Zhang

Computation and Language

Long-Chain Reasoning Distillation via Adaptive Prefix Alignment

Zhenghao Liu, Zhuoyang Wu, Xinze Li et al. (+6 more)

Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in solving complex mathematical problems. Recent studies show that distilling long reasoning trajectories can effectively enhance the reasoning performance of small-scale student models. However, teacher-generated reasoning trajectories are often excessively long and structurally complex, making them difficult for student models to learn. This mismatch leads to a gap between the provided supervision signal and the learning capacity of the student model. To address this challenge, we propose Prefix-ALIGNment distillation (P-ALIGN), a framework that fully exploits teacher CoTs for distillation through adaptive prefix alignment. Specifically, P-ALIGN adaptively truncates teacher-generated reasoning trajectories by determining whether the remaining suffix is concise and sufficient to guide the student model. Then, P-ALIGN leverages the teacher-generated prefix to supervise the student model, encouraging effective prefix alignment. Experiments on multiple mathematical reasoning benchmarks demonstrate that P-ALIGN outperforms all baselines by over 3%. Further analysis indicates that the prefixes constructed by P-ALIGN provide more effective supervision signals, while avoiding the negative impact of redundant and uncertain reasoning components. All code is available at https://github.com/NEUIR/P-ALIGN.
Read the full abstract

All authors Zhenghao Liu, Zhuoyang Wu, Xinze Li, Yukun Yan, Shuo Wang, Zulong Chen, Yu Gu, Ge Yu, Maosong Sun

Computation and Language

EmplifAI: a Fine-grained Dataset for Japanese Empathetic Medical Dialogues in 28 Emotion Labels

Wan Jou She, Lis Kanashiro Pereira, Fei Cheng et al. (+3 more)

This paper introduces EmplifAI, a Japanese empathetic dialogue dataset designed to support patients coping with chronic medical conditions. They often experience a wide range of positive and negative emotions (e.g., hope and despair) that shift across different stages of disease management. EmplifAI addresses this complexity by providing situation-based dialogues grounded in 28 fine-grained emotion categories, adapted and validated from the GoEmotions taxonomy. The dataset includes 280 medically contextualized situations and 4125 two-turn dialogues, collected through crowdsourcing and expert review. To evaluate emotional alignment in empathetic dialogues, we assessed model predictions on situation–dialogue pairs using BERTScore across multiple large language models (LLMs), achieving F1 scores of 0.83. Fine-tuning a baseline Japanese LLM (LLM-jp-3.1-13b-instruct4) with EmplifAI resulted in notable improvements in fluency, general empathy, and emotion-specific empathy. Furthermore, we compared the scores assigned by LLM-as-a-Judge and human raters on dialogues generated by multiple LLMs to validate our evaluation pipeline and discuss the insights and potential risks derived from the correlation analysis.
Read the full abstract

All authors Wan Jou She, Lis Kanashiro Pereira, Fei Cheng, Sakiko Yahata, Panote Siriaraya, Eiji Aramaki

Computation and Language

EHRNavigator: A Multi-Agent System for Patient-Level Clinical Question Answering over Heterogeneous Electronic Health Records

Lingfei Qian, Mauro Giuffre, Yan Wang et al. (+11 more)

Clinical decision-making increasingly relies on timely and context-aware access to patient information within Electronic Health Records (EHRs), yet most existing natural language question-answering (QA) systems are evaluated solely on benchmark datasets, limiting their practical relevance. To overcome this limitation, we introduce EHRNavigator, a multi-agent framework that harnesses AI agents to perform patient-level question answering across heterogeneous and multimodal EHR data. We assessed its performance using both public benchmark and institutional datasets under realistic hospital conditions characterized by diverse schemas, temporal reasoning demands, and multimodal evidence integration. Through quantitative evaluation and clinician-validated chart review, EHRNavigator demonstrated strong generalization, achieving 86% accuracy on real-world cases while maintaining clinically acceptable response times. Overall, these findings confirm that EHRNavigator effectively bridges the gap between benchmark evaluation and clinical deployment, offering a robust, adaptive, and efficient solution for real-world EHR question answering.
Read the full abstract

All authors Lingfei Qian, Mauro Giuffre, Yan Wang, Huan He, Qianqian Xie, Xuguang Ai, Xeuqing Peng, Fan Ma, Ruey-Ling Weng, Donald Wright, Adan Wang, Qingyu Chen, Vipina K. Keloth, Hua Xu

Computation and Language · Cross-listed from cs.LG

Continuous-Depth Transformers with Learned Control Dynamics

Peter Jemley

We present a hybrid transformer architecture that replaces discrete middle layers with a continuous-depth Neural Ordinary Differential Equation (ODE) block, enabling inference-time control over generation attributes via a learned steering signal. Unlike standard transformers that process representations through fixed discrete layers, our approach treats depth as a continuous variable governed by a learned vector field $F_\theta(H, \tau, u)$, where $u$ is a low-dimensional control signal injected via explicit concatenation. We validate the architecture through four experiments: (1) gradient flow stability with zero exploding/vanishing gradient events, (2) semantic steering achieving 98%/88% accuracy for positive/negative sentiment control, (3) continuous interpolation validated by a negligible 0.068% trajectory divergence between fixed and adaptive solvers, and (4) efficiency benchmarking demonstrating latency parity with standard discrete baselines. Additionally, we show that adaptive ODE solvers reveal geometric structure in the learned dynamics: the control signal partitions the vector field into distinct dynamical regimes with different curvature characteristics. The adjoint method enables $O(1)$ memory training regardless of integration depth. Our results demonstrate that continuous-depth dynamics with learned control signals provide a viable, efficient mechanism for steerable language generation.
Read the full abstract

Computation and Language

SocraticKG: Knowledge Graph Construction via QA-Driven Fact Extraction

Sanghyeok Choi, Woosang Jeon, Kyuseok Yang et al. (+1 more)

Constructing Knowledge Graphs (KGs) from unstructured text provides a structured framework for knowledge representation and reasoning, yet current LLM-based approaches struggle with a fundamental trade-off: factual coverage often leads to relational fragmentation, while premature consolidation causes information loss. To address this, we propose SocraticKG, an automated KG construction method that introduces question-answer pairs as a structured intermediate representation to systematically unfold document-level semantics prior to triple extraction. By employing 5W1H-guided QA expansion, SocraticKG captures contextual dependencies and implicit relational links typically lost in direct KG extraction pipelines, providing explicit grounding in the source document that helps mitigate implicit reasoning errors. Evaluation on the MINE benchmark and HotpotQA downstream task demonstrates that our approach effectively addresses the coverage-connectivity trade-off, achieving superior factual retention and structural cohesion while supporting complex multi-hop reasoning.
Read the full abstract

All authors Sanghyeok Choi, Woosang Jeon, Kyuseok Yang, Taehyeong Kim

Computation and Language

Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG

David Samuel Setiawan, Raphaël Merx and Jey Han Lau

Neural Machine Translation (NMT) models for low-resource languages suffer significant performance degradation under domain shift. We quantify this challenge using Dhao, an indigenous language of Eastern Indonesia with no digital footprint beyond the New Testament (NT). When applied to the unseen Old Testament (OT), a standard NMT model fine-tuned on the NT drops from an in-domain score of 36.17 chrF++ to 27.11 chrF++. To recover this loss, we introduce a hybrid framework where a fine-tuned NMT model generates an initial draft, which is then refined by a Large Language Model (LLM) using Retrieval-Augmented Generation (RAG). The final system achieves 35.21 chrF++ (+8.10 recovery), effectively matching the original in-domain quality. Our analysis reveals that this performance is driven primarily by the number of retrieved examples rather than the choice of retrieval algorithm. Qualitative analysis confirms the LLM acts as a robust "safety net," repairing severe failures in zero-shot domains.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation

Seoyeon Kim and Jaehyung Kim

Personalizing Large Language Models typically relies on static retrieval or one-time adaptation, assuming user preferences remain invariant over time. However, real-world interactions are dynamic, where user interests continuously evolve, posing a challenge for models to adapt to preference drift without catastrophic forgetting. Standard continual learning approaches often struggle in this context, as they indiscriminately update on noisy interaction streams, failing to distinguish genuine preference shifts from transient contexts. To address this, we introduce SPRInG, a novel semi-parametric framework designed for effective continual personalization. During training, SPRInG employs drift-driven selective adaptation, which utilizes a likelihood-based scoring function to identify high-novelty interactions, selectively updating the user-specific adapter on drift signals while preserving hard-to-learn residuals in a replay buffer. During inference, we apply strict relevance gating and fuse parametric knowledge with retrieved history via probability interpolation. Experiments on the long-form personalized generation benchmark demonstrate that SPRInG significantly outperforms existing baselines, validating its robustness for real-world continual personalization.
Read the full abstract

Computation and Language

Take Out Your Calculators: Estimating the Real Difficulty of Question Items with LLM Student Simulations

Christabel Acquaye, Yi Ting Huang, Marine Carpuat et al. (+1 more)

Standardized math assessments require expensive human pilot studies to establish the difficulty of test items. We investigate the predictive value of open-source large language models (LLMs) for evaluating the difficulty of multiple-choice math questions for real-world students. We show that, while LLMs are poor direct judges of problem difficulty, simulation-based approaches with LLMs yield promising results under the right conditions. Under the proposed approach, we simulate a “classroom” of 4th, 8th, or 12th-grade students by prompting the LLM to role-play students of varying proficiency levels. We use the outcomes of these simulations to fit Item Response Theory (IRT) models, comparing learned difficulty parameters for items to their real-world difficulties, as determined by item-level statistics furnished by the National Assessment of Educational Progress (NAEP). We observe correlations as high as 0.75, 0.76, and 0.82 for grades 4, 8, and 12, respectively, on the item-level correctness rates. In our simulations, we experiment on math MCQs with different “classroom sizes,” showing tradeoffs between computation size and accuracy. We find that role-plays with diverse-named students improve predictions (compared to student IDs), and stratifying names across gender and race further improves predictions. Our results show that LLMs with relatively weaker mathematical abilities (Gemma) actually yield better real-world difficulty predictions than mathematically stronger models (Llama and Qwen), further underscoring the suitability of these models for the task.
Read the full abstract

All authors Christabel Acquaye, Yi Ting Huang, Marine Carpuat, Rachel Rudinger

Wednesday 14 January

Computation and Language · Cross-listed from cs.IR

Leveraging Large Language Models to Extract and Translate Medical Information in Doctors' Notes for Health Records and Diagnostic Billing Codes

Peter Hartnett, Chung-Chi Huang, Sarah Hartnett et al. (+1 more)

Physician burnout in the United States has reached critical levels, driven in part by the administrative burden of Electronic Health Record (EHR) documentation and complex diagnostic codes. To relieve this strain and maintain strict patient privacy, this thesis explores an on-device, offline automatic medical coding system. The work focuses on using open-weight Large Language Models (LLMs) to extract clinical information from physician notes and translate it into ICD-10-CM diagnostic codes without reliance on cloud-based services. A privacy-focused pipeline was developed using Ollama, LangChain, and containerized environments to evaluate multiple open-weight models, including Llama 3.2, Mistral, Phi, and DeepSeek, on consumer-grade hardware. Model performance was assessed for zero-shot, few-shot, and retrieval-augmented generation (RAG) prompting strategies using a novel benchmark of synthetic medical notes. Results show that strict JSON schema enforcement achieved near 100% formatting compliance, but accurate generation of specific diagnostic codes remains challenging for smaller local models (7B-20B parameters). Contrary to common prompt-engineering guidance, few-shot prompting degraded performance through overfitting and hallucinations. While RAG enabled limited discovery of unseen codes, it frequently saturated context windows, reducing overall accuracy. The findings suggest that fully automated unsupervised coding with local open-source models is not yet reliable; instead, a human-in-the-loop assisted coding approach is currently the most practical path forward. This work contributes a reproducible local LLM architecture and benchmark dataset for privacy-preserving medical information extraction and coding.
Read the full abstract

All authors Peter Hartnett, Chung-Chi Huang, Sarah Hartnett, David Hartnett

Computation and Language · Cross-listed from cs.AI

Knowledge Boundary Discovery for Large Language Models

Ziquan Wang and Zhongqi Lu

We propose Knowledge Boundary Discovery (KBD), a reinforcement learning based framework to explore the knowledge boundaries of the Large Language Models (LLMs). We define the knowledge boundary by automatically generating two types of questions: (i) those the LLM can confidently answer (within-knowledge boundary) and (ii) those it cannot (beyond-knowledge boundary). Iteratively exploring and exploiting the LLM's responses to find its knowledge boundaries is challenging because of the hallucination phenomenon. To find the knowledge boundaries of an LLM, the agent interacts with the LLM under the modeling of exploring a partially observable environment. The agent generates a progressive question as the action, adopts an entropy reduction as the reward, receives the LLM's response as the observation and updates its belief states. We demonstrate that the KBD detects knowledge boundaries of LLMs by automatically finding a set of non-trivial answerable and unanswerable questions. We validate the KBD by comparing its generated knowledge boundaries with manually crafted LLM benchmark datasets. Experiments show that our KBD-generated question set is comparable to the human-generated datasets. Our approach paves a new way to evaluate LLMs.
Read the full abstract

Computation and Language

Novelty Reviewer: A Large-Scale Dataset and Model for Reviewer-Aligned Novelty Assessment

Abeer Mostafa, Thi Huyen Nguyen and Zahra Ahmadi

Assessing novelty in AI research is arguably the most consequential yet least reliable step in peer review. Reviewer judgments of novelty remain inconsistent and dependent on comparisons to prior work that are often incomplete. We release a large-scale Novelty Reviewer Dataset built from 102,021 official reviews of 37,899 ICLR and NeurIPS submissions (2022-2025). For each submission, it provides the reviewers' novelty statements, a paper-level novelty assessment, and a novelty score on a four-point scale. On top of this dataset, we build Novelty Reviewer, a literature-aware framework that learns how reviewers judge novelty and grounds these judgments in semantic comparisons to existing research. We fine-tune a large language model on the dataset to capture reviewer-aligned novelty evaluation behavior. Given a manuscript, the framework extracts structured representations of its ideas, methods, and claims; retrieves semantically related papers; and constructs a knowledge graph that enables concept-level comparison to prior work. Based on the retrieved evidence, the model produces calibrated novelty scores alongside human-like explanatory assessments, reducing overestimation and improving consistency relative to existing approaches. Furthermore, we analyze different LLMs' abilities to assess originality and detect potential conceptual plagiarism. We publicly release the dataset, the model, and the code with documentation to support research on automated novelty assessment and peer review.
Read the full abstract

Computation and Language

PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models

Rajarshi Roy, Jonathan Raiman, Sang-gil Lee et al. (+5 more)

Recent advances in duplex speech models have enabled natural, low-latency speech-to-speech interactions. However, existing models are restricted to a fixed role and voice, limiting their ability to support structured, role-driven real-world applications and personalized interactions. In this work, we introduce PersonaPlex, a duplex conversational speech model that incorporates hybrid system prompts, combining role conditioning with text prompts and voice cloning with speech samples. PersonaPlex is trained on a large-scale synthetic dataset of paired prompts and user-agent conversations, generated with open-source large language models (LLM) and text-to-speech (TTS) models. To evaluate role conditioning in real-world settings, we extend the Full-Duplex-Bench benchmark beyond a single assistant role to multi-role customer service scenarios. Experiments show that PersonaPlex achieves strong role-conditioned behavior, voice-conditioned speech, and natural conversational responsiveness, surpassing state-of-the-art duplex speech models and hybrid large language model-based speech systems in role adherence, speaker similarity, latency, and naturalness.
Read the full abstract

All authors Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim, Bryan Catanzaro

Computation and Language

A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents

Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang et al. (+57 more)

Research in artificial intelligence is shifting from model innovations and benchmark scores towards problem definition and rigorous real-world evaluation. As the field enters the "second half," the central challenge becomes real utility in long-horizon, dynamic, and user-dependent settings such as agentic coding, deep research, and computer use, where LLM-based agents face context explosion beyond fixed context windows and must continuously accumulate, manage, and selectively reuse information across extended interactions. Memory, with hundreds of papers released in 2025, therefore emerges as the critical solution to fill this utility gap. Beyond passive storage, memory is increasingly the substrate through which agents self-evolve: short-term memory gates which experiences are perceived and abstracted during execution, while long-term memory consolidates them into reusable knowledge and skills, forming the loop through which agents improve from their own experience. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal parametric state and external retrieval-augmented stores), cognitive mechanism (sensory, working, episodic, semantic, and procedural), and memory subject (user-centric personalization and agent-centric experience). We then analyze how memory is operated under single- and multi-agent topologies and highlight learning policies over memory operations, showing how memory management itself is becoming a trainable capability spanning reinforcement-learned context curation, experience consolidation at decision time, and the emerging ecosystem of portable, shareable agent skills. Finally, we review evaluation benchmarks and metrics for memory utility, and outline open challenges and future directions.
Read the full abstract

All authors Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang, Yuanchen Bei, Yankai Chen, Tao Feng, Xinyu Pan, Zhen Tan, Yu Wang, Tianxin Wei, Shanglin Wu, Ruiyao Xu, Liangwei Yang, Rui Yang, Wooseong Yang, Chin-Yuan Yeh, Hanrong Zhang, Haozhen Zhang, Siqi Zhu, Henry Peng Zou, Wanjia Zhao, Song Wang, Wujiang Xu, Zixuan Ke, Zheng Hui, Dawei Li, Yaozu Wu, Langzhou He, Chen Wang, Xiongxiao Xu, Baixiang Huang, Juntao Tan, Shelby Heinecke, Huan Wang, Caiming Xiong, Ahmed A. Metwally, Jun Yan, Chen-Yu Lee, Hanqing Zeng, Yinglong Xia, Xiaokai Wei, Ali Payani, Yu Wang, Haitong Ma, Wenya Wang, Chenguang Wang, Yu Zhang, Xin Eric Wang, Yongfeng Zhang, Jiaxuan You, Hanghang Tong, Xiao Luo, Xue Liu, Yizhou Sun, Wei Wang, Julian McAuley, James Zou, Jiawei Han, Philip S. Yu, Kai Shu

Computation and Language

CAST: Character-and-Scene Episodic Memory for Agents

Kexin Ma, Bojun Li, Yuhua Tang et al. (+2 more)

Episodic memory is a central component of human memory, which refers to the ability to recall coherent events grounded in who, when, and where. However, most agent memory systems only emphasize semantic recall and treat experience as structures such as key-value, vector, or graph, which makes them struggle to represent and retrieve coherent events. To address this challenge, we propose a Character-and-Scene based memory architecture(CAST) inspired by dramatic theory. Specifically, CAST constructs 3D scenes (time/place/topic) and organizes them into character profiles that summarize the events of a character to represent episodic memory. Moreover, CAST complements this episodic memory with a graph-based semantic memory, which yields a robust dual memory design. Experiments demonstrate that CAST has averagely improved 8.11% F1 and 10.21% J(LLM-as-a-Judge) than baselines on various datasets, especially on open and time-sensitive conversational questions.
Read the full abstract

All authors Kexin Ma, Bojun Li, Yuhua Tang, Liting Sun, Ruochun Jin

Computation and Language

Relevance-aware Multi-context Contrastive Decoding for Retrieval-augmented Visual Question Answering

Jongha Kim, Byungoh Ko, Jeehye Na et al. (+2 more)

Despite the remarkable capabilities of Large Vision Language Models (LVLMs), they still lack detailed knowledge about specific entities. Retrieval-augmented Generation (RAG) is a widely adopted solution that enhances LVLMs by providing additional contexts from an external Knowledge Base. However, we observe that previous decoding methods for RAG are sub-optimal as they fail to sufficiently leverage multiple relevant contexts and suppress the negative effects of irrelevant contexts. To this end, we propose Relevance-aware Multi-context Contrastive Decoding (RMCD), a novel decoding method for RAG. RMCD outputs a final prediction by combining outputs predicted with each context, where each output is weighted based on its relevance to the question. By doing so, RMCD effectively aggregates useful information from multiple relevant contexts while also counteracting the negative effects of irrelevant ones. Experiments show that RMCD consistently outperforms other decoding methods across multiple LVLMs, achieving the best performance on three knowledge-intensive visual question-answering benchmarks. Also, RMCD can be simply applied by replacing the decoding method of LVLMs without additional training. Analyses also show that RMCD is robust to the retrieval results, consistently performing the best across the weakest to the strongest retrieval results. Code is available at https://github.com/mlvlab/RMCD.
Read the full abstract

All authors Jongha Kim, Byungoh Ko, Jeehye Na, Jinsung Yoon, Hyunwoo J. Kim

Computation and Language

ROSA-Tuning: Enhancing Long-Context Modeling via Suffix Matching

Yunao Zheng, Xiaojie Wang, Lei Ren et al. (+1 more)

Long-context capability and computational efficiency are among the central challenges facing today's large language models. Existing efficient attention methods reduce computational complexity, but they typically suffer from a limited coverage of the model state. This paper proposes ROSA-Tuning, a retrieval-and-recall mechanism for enhancing the long-context modeling ability of pretrained models. Beyond the standard attention mechanism, ROSA-Tuning leverages in parallel a CPU-based ROSA (RWKV Online Suffix Automaton) retrieval module, which efficiently locates historical positions in long contexts that are relevant to the current query, and injects the retrieved information into the model state in a trainable manner; subsequent weighted fusion can then be handled by range-restricted attention. To enable end-to-end training, we employ the binary discretization strategy and the counterfactual gradient algorithm, and further optimize overall execution efficiency via an asynchronous CPU-GPU pipeline. Systematic evaluations on Qwen3-Base-1.7B show that ROSA-Tuning substantially restores the long-context modeling ability of windowed-attention models, achieving performance close to and in some cases matching global attention on benchmarks such as LongBench, while maintaining computational efficiency and GPU memory usage that are nearly comparable to windowed-attention methods, offering a new technical path for efficient long-context processing. The example code can be found at https://github.com/zyaaa-ux/ROSA-Tuning.
Read the full abstract

All authors Yunao Zheng, Xiaojie Wang, Lei Ren, Wei Chen

Computation and Language

Test-Time Detoxification without Training or Learning Anything

Baturay Saglam and Dionysis Kalogerias

Large language models can produce toxic or inappropriate text even for benign inputs, creating risks when deployed at scale. Detoxification is therefore important for safety and user trust, particularly when we want to reduce harmful content without sacrificing the model's generation quality. Many existing approaches rely on model retraining, gradients, or learned auxiliary components, which can be costly and may not transfer across model families or to truly black-box settings. We introduce a test-time procedure that approximates the gradient of completion toxicity with respect to the input embeddings and uses a small number of descent steps to steer generation toward less toxic continuations. This is achieved with zeroth-order optimization that requires only access to input embeddings, a toxicity scoring function, and forward evaluations of the model. Empirically, the approach delivers robust toxicity reductions across models and prompts and, in most settings, achieves the best overall toxicity-quality trade-off. More broadly, our work positions word embeddings as effective control variables and encourages wider use of black-box optimization to guide autoregressive language models toward scalable, safer text generation, without requiring any training or access to intermediate computations.
Read the full abstract

Computation and Language

STEMVerse: A Dual-Axis Diagnostic Framework for STEM Reasoning in Large Language Models

Xuzhao Li, Xuchen Li, Jian Zhao et al. (+1 more)

As Large Language Models (LLMs) achieve significant breakthroughs in complex reasoning tasks, evaluating their proficiency in science, technology, engineering, and mathematics (STEM) has become a primary method for measuring machine intelligence. However, current evaluation paradigms often treat benchmarks as isolated "silos," offering only monolithic aggregate scores that neglect the intricacies of both academic specialization and cognitive depth. This result-oriented approach fails to distinguish whether model errors stem from insufficient domain knowledge or deficiencies in cognitive capacity, thereby limiting the diagnostic value. To address this, we propose STEMVerse, a diagnostic framework designed to systematically analyze the STEM reasoning capabilities of LLMs. This framework characterizes model performance across academic specialization and cognitive complexity to map the capability required for reasoning. We re-aggregate over 20,000 STEM problems from mainstream benchmarks into a unified "Discipline $\times$ Cognition" capability space, assigning dual-axis labels to every instance. Utilizing this unified diagnostic framework, we systematically evaluate representative LLM families across varying parameter scales and training paradigms. Our empirical results reveal structural failure patterns in STEM reasoning. By integrating multi-disciplinary coverage and fine-grained cognitive stratification into a unified framework, STEMVerse provides a clear and actionable perspective for understanding the scientific reasoning characteristics of LLMs.
Read the full abstract

All authors Xuzhao Li, Xuchen Li, Jian Zhao, Shiyu Hu