Computation and Language
Multilingual Amnesia: On the Transferability of Unlearning in Multilingual LLMs
Alireza Dehghanpour Farashah, Aditi Khandelwal, Marylou Fauchard et al. (+3 more)
As
multilingual large
language models become more widely used, ensuring their safety and
fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has primarily focused on monolingual settings, typically English,
multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we study
multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend
benchmarks for factual knowledge and stereotypes to ten
languages through translation: English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian. These
languages span five
language families and a wide range of resource levels. Our experiments show that unlearning in high-resource
languages is generally more stable, with asymmetric transfer effects observed between typologically related
languages. Furthermore, our analysis of linguistic distances indicates that syntactic similarity is the strongest predictor of cross-lingual unlearning behavior.
Read the full abstract ▾
All authors Alireza Dehghanpour Farashah, Aditi Khandelwal, Marylou Fauchard, Zhuan Shi, Negar Rostamzadeh, Golnoosh Farnadi
Computation and Language · Cross-listed from cs.CR
Continual Pretraining on Encrypted Synthetic Data for Privacy-Preserving LLMs
Honghao Liu, Xuhui Jiang, Chengjin Xu et al. (+4 more)
Preserving
privacy in sensitive data while pretraining large
language models on small, domain-specific corpora presents a significant challenge. In this work, we take an exploratory step toward
privacy-preserving continual pretraining by proposing an entity-based framework that synthesizes encrypted training data to protect personally identifiable information (PII). Our approach constructs a weighted entity graph to guide data synthesis and applies deterministic encryption to PII entities, enabling LLMs to encode new knowledge through continual pretraining while granting authorized access to sensitive data through decryption keys. Our results on limited-scale
datasets demonstrate that our pretrained models outperform base models and ensure PII security, while exhibiting a modest performance gap compared to models trained on unencrypted synthetic data. We further show that increasing the number of entities and leveraging graph-based synthesis improves model performance, and that encrypted models retain instruction-following capabilities with long retrieved contexts. We discuss the security implications and limitations of deterministic encryption, positioning this work as an initial investigation into the design space of encrypted data pretraining for
privacy-preserving LLMs. Our code is available at https://github.com/DataArcTech/SoE.
Read the full abstract ▾
All authors Honghao Liu, Xuhui Jiang, Chengjin Xu, Cehao Yang, Yiran Cheng, Lionel Ni, Jian Guo
Computation and Language
GIFT: Games as Informal Training for Generalizable LLMs
Nuoyan Lyu, Bingbing Xu, Xueyun Tian et al. (+6 more)
Recent LLMs excel at formal tasks such as mathematical
reasoning and code generation, but still struggle with broader abilities such as planning, creativity, and social intelligence. Inspired by human learning, where formal instruction and informal experience jointly shape intelligence, we introduce informal learning into LLM training and use games as annotation-free, feedback-driven environments. To cover diverse abilities including abstract
reasoning, planning, creativity, and social interaction, we combine formal math tasks with three representative game tasks, including
Matrix Games, TicTacToe, and Who's the Spy. However, directly mixing these tasks under a unified RL objective can blur task-specific learning signals and provides no explicit guidance for coordinating task-gradient directions. To combat these, we propose Coordinated Subtask Training (CST), which replaces a single mixed update with sequential subtask-specific updates, separating heterogeneous RL signals while implicitly promoting coordination among subtasks. Experiments on ability-oriented
benchmarks show that game-based informal learning improves generalization beyond formal training alone, while CST further enhances multi-task RL by preserving in-domain subtask performance and improving broader general abilities. Code and data are publicly available.
Read the full abstract ▾
All authors Nuoyan Lyu, Bingbing Xu, Xueyun Tian, Weihao Meng, Yige Yuan, Yang Zhang, Zhiyong Huang, Tat-Seng Chua, Huawei Shen
Computation and Language
Text Detoxification in isiXhosa and Yorùbá: A Cross-Lingual Machine Learning Approach for Low-Resource African Languages
Abayomi O. Agbeyangi
Toxic
language is one of the major barrier to safe online participation, yet robust mitigation tools are scarce for African
languages. This study addresses this critical gap by investigating automatic text detoxification (toxic to neutral rewriting) for two low-resource African
languages, isiXhosa and Yorùbá. The work contributes a novel, pragmatic hybrid methodology: a lightweight, interpretable TF-IDF and Logistic Regression model for transparent toxicity detection, and a controlled lexicon- and
token-guided rewriting component. A parallel corpus of toxic to neutral rewrites, which captures idiomatic usage, diacritics, and code switching, was developed to train and evaluate the model. The detection component achieved stratified K-fold accuracies of 61-72% (isiXhosa) and 72-86% (Yorùbá), with per-
language ROC-AUCs up to 0.88. The rewriting component successfully detoxified all detected toxic sentences while preserving 100% of non-toxic sentences. These results demonstrate that scalable, interpretable machine learning detectors combined with rule-based edits offer a competitive and resource-efficient solution for culturally adaptive safety tooling, setting a new
benchmark for low-resource Text Style Transfer (TST) in African
languages.
Read the full abstract ▾
Computation and Language
Data Augmented Pipeline for Legal Information Extraction and Reasoning
Nguyen Minh Phuong, Ha-Thanh Nguyen, May Myo Zin et al. (+1 more)
In this paper, we propose a pipeline leveraging Large
Language Models (LLMs) for data augmentation in Information Extraction tasks within the legal domain. The proposed method is both simple and effective, significantly reducing the manual effort required for data annotation while enhancing the robustness of Information Extraction systems. Furthermore, the method is generalizable, making it applicable to various Natural
Language Processing (NLP) tasks beyond the legal domain.
Read the full abstract ▾
All authors Nguyen Minh Phuong, Ha-Thanh Nguyen, May Myo Zin, Ken Satoh
Computation and Language · Cross-listed from cs.CV
SceneAlign: Aligning Multimodal Reasoning to Scene Graphs in Complex Visual Scenes
Chuhan Wang, Xintong Li, Jennifer Yuntong Zhang et al. (+5 more)
Multimodal large
language models often struggle with faithful
reasoning in complex visual scenes, where intricate entities and relations require precise visual grounding at each step. This
reasoning unfaithfulness frequently manifests as hallucinated entities, mis-grounded relations, skipped steps, and over-specified
reasoning. Existing preference-based approaches, typically relying on textual perturbations or answer-conditioned rationales, fail to address this challenge as they allow models to exploit
language priors to bypass visual grounding. To address this, we propose SceneAlign, a framework that leverages scene graphs as structured visual information to perform controllable structural interventions. By identifying
reasoning-critical nodes and perturbing them through four targeted strategies that mimic typical grounding failures, SceneAlign constructs hard negative rationales that remain linguistically plausible but are grounded in inaccurate visual facts. These contrastive pairs are used in Direct Preference Optimization to steer models toward fine-grained, structure-faithful
reasoning. Across seven visual
reasoning benchmarks, SceneAlign consistently improves answer accuracy and
reasoning faithfulness, highlighting the effectiveness of grounding-aware
alignment for
multimodal reasoning.
Read the full abstract ▾
All authors Chuhan Wang, Xintong Li, Jennifer Yuntong Zhang, Junda Wu, Chengkai Huang, Lina Yao, Julian McAuley, Jingbo Shang
Computation and Language
ACR: Adaptive Context Refactoring via Context Refactoring Operators for Multi-Turn Dialogue
Jiawei Shen, Jia Zhu, Hanghui Guo et al. (+9 more)
Large
Language Models (LLMs) have shown remarkable performance in multi-turn dialogue. However, in multi-turn dialogue, models still struggle to stay aligned with what has been established earlier, follow dependencies across many turns, and avoid drifting into incorrect facts as the interaction grows longer. Existing approaches primarily focus on extending the context window, introducing external
memory, or applying context compression, yet these methods still face limitations such as contextual inertia and state drift. To address these challenges, we propose the Adaptive Context Refactoring (ACR) Framework, which dynamically monitors and reshapes the interaction history to mitigate contextual inertia and state drift actively. ACR is built on a library of context refactoring operators and a teacher-guided self-evolving training paradigm that learns when to intervene and how to refactor, thereby decoupling context management from the
reasoning process. Extensive experiments on multi-turn dialogue demonstrate that our method significantly outperforms existing baselines while reducing
token consumption.
Read the full abstract ▾
All authors Jiawei Shen, Jia Zhu, Hanghui Guo, Weijie Shi, Yue Cui, Qingyu Niu, Guoqing Ma, Yidan Liang, Jingjiang Liu, Yiling Wang, Shimin Di, Jiajie Xu
Computation and Language
Can large language models interpret unstructured chat data on dynamic group decision-making processes? Evidence on joint destination choice
Sung-Yoo Lim, Koki Sato, Kiyoshi Takami et al. (+2 more)
Social activities result from complex joint activity-travel decisions between group members. While observing the decision-making process of these activities is difficult via traditional travel surveys, the advent of new types of data, such as unstructured chat data, can help shed some light on these complex processes. However, interpreting these decision-making processes requires inferring both explicit and implicit factors. This typically involves the labor-intensive task of manually annotating dialogues to capture context-dependent meanings shaped by the social and cultural norms. This study evaluates the potential of Large
Language Models (LLMs) to automate and complement human annotation in interpreting decision-making processes from group chats, using data on joint eating-out activities in Japan as a case study. We designed a prompting framework inspired by the knowledge acquisition process, which sequentially extracts key decision-making factors, including the group-level restaurant choice set and outcome, individual preferences of each alternative, and the specific attributes driving those preferences. This structured process guides the LLM to interpret group chat data, converting unstructured dialogues into structured tabular data describing decision-making factors. To evaluate LLM-driven outputs, we conduct a quantitative analysis using a human-annotated ground truth
dataset and a qualitative error analysis to examine model limitations. Results show that while the LLM reliably captures explicit decision-making factors, it struggles to identify nuanced implicit factors that human annotators readily identified. We pinpoint specific contexts when LLM-based extraction can be trusted versus when human oversight remains essential. These findings highlight both the potential and limitations of LLM-based analysis for incorporating non-traditional data sources on social activities.
Read the full abstract ▾
All authors Sung-Yoo Lim, Koki Sato, Kiyoshi Takami, Giancarlos Parady, Eui-Jin Kim
Computation and Language · Cross-listed from cs.DB
RISE: Rule-Driven SQL Dialect Translation via Query Reduction
Xudong Xie, Yuwei Zhang, Wensheng Dou et al. (+5 more)
Translating SQL dialects across different relational database management systems (RDBMSs) is crucial for migrating RDBMS-based applications to the cloud. Traditional SQL dialect translation tools rely on manually-crafted rules, necessitating significant manual effort to support new RDBMSs and dialects. Although large
language models (LLMs) can assist in translating SQL dialects, they often struggle with lengthy and complex SQL queries. In this paper, we propose RISE, a novel LLM-based SQL dialect translation approach that can accurately handle lengthy and complex SQL queries. Given a complex source query $Q_c$ that contains a SQL dialect $d$, we first employ a dialect-aware query reduction technique to derive a simplified query $Q_{s}$ by removing $d$-irrelevant SQL elements from $Q_c$. Subsequently, we utilize LLMs to translate $Q_{s}$ into $Q_{s^{'}}$, and automatically extract the translation rule $r_d$ for dialect $d$ based on the relationship between $Q_{s}$ and $Q_{s^{'}}$. By applying $r_d$ to $Q_c$, we can effectively translate the dialect $d$ within $Q_c$, thereby bypassing the complexity of the source query $Q_c$. We evaluate RISE on two real-world
benchmarks, i.e., TPC-DS and SQLProcBench, comparing its performance against both the traditional rule-based tools and the LLM-based approaches with respect to translation accuracy. RISE achieves accuracies of 97.98% on TPC-DS and 100% on SQLProcBench, outperforming the baselines by an average improvement of 24.62% and 238.41%, respectively.
Read the full abstract ▾
All authors Xudong Xie, Yuwei Zhang, Wensheng Dou, Yu Gao, Ziyu Cui, Jiansen Song, Rui Yang, Jun Wei
Computation and Language · Cross-listed from cs.AI
WildSci: Advancing Scientific Reasoning from In-the-Wild Literature
Tengxiao Liu, Deepak Nathani, Zekun Li et al. (+2 more)
Recent progress in large
language model (LLM)
reasoning has focused on domains like mathematics and coding, where abundant high-quality data and objective evaluation metrics are readily available. In contrast, progress in LLM
reasoning models remains limited in scientific domains such as medicine and materials science due to limited
dataset coverage and the inherent complexity of open-ended scientific questions. To address these challenges, we introduce WildSci, a new
dataset of domain-specific science questions automatically synthesized from peer-reviewed literature, covering 9 scientific disciplines and 26 subdomains. By framing complex scientific
reasoning tasks in a multiple-choice format, we enable scalable training with well-defined reward signals. We further apply reinforcement learning to finetune models on these data and analyze the resulting training dynamics, including domain-specific performance changes, response behaviors, and generalization trends. Experiments on a suite of scientific
benchmarks demonstrate the effectiveness of our
dataset and approach. We release WildSci to enable scalable and sustainable research in scientific
reasoning, available at https://huggingface.co/
datasets/JustinTX/WildSci.
Read the full abstract ▾
All authors Tengxiao Liu, Deepak Nathani, Zekun Li, Kevin Yang, William Yang Wang
Computation and Language · Cross-listed from cs.SD
The ICASSP 2026 HumDial Challenge: Benchmarking Human-like Spoken Dialogue Systems in the LLM Era
Zhixian Zhao, Shuiyuan Wang, Guojian Li et al. (+12 more)
Driven by the rapid advancement of Large
Language Models (LLMs), particularly Audio-LLMs and Omni-models, spoken dialogue systems have evolved significantly, progressively narrowing the gap between human-machine and human-human interactions. Achieving truly “human-like” communication necessitates a dual capability: emotional intelligence to perceive and resonate with users' emotional states, and robust interaction mechanisms to navigate the dynamic, natural flow of conversation, such as real-time turn-taking. Therefore, we launched the first Human-like Spoken Dialogue Systems Challenge (HumDial) at ICASSP 2026 to
benchmark these dual capabilities. Anchored by a sizable
dataset derived from authentic human conversations, this initiative establishes a fair evaluation platform across two tracks: (1) Emotional Intelligence, targeting long-term
emotion understanding and empathetic generation; and (2) Full-Duplex Interaction, systematically evaluating real-time decision-making under “ listening-while-speaking” conditions. This paper summarizes the
dataset, track configurations, and the final results.
Read the full abstract ▾
All authors Zhixian Zhao, Shuiyuan Wang, Guojian Li, Hongfei Xue, Chengyou Wang, Shuai Wang, Longshuai Xiao, Zihan Zhang, Hui Bu, Xin Xu, Xinsheng Wang, Hexin Liu, Eng Siong Chng, Hung-yi Lee, Lei Xie
Computation and Language
ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging
Junyao Yang, Chen Qian, Dongrui Liu et al. (+3 more)
Large
Reasoning Models (LRMs) with long chain-of-thought
reasoning have recently achieved remarkable success. Yet, equipping domain-specialized models with such
reasoning capabilities, referred to as "
Reasoning + X", remains a significant challenge. While model merging offers a promising training-free solution, existing methods often suffer from a destructive performance collapse: existing methods tend to both weaken
reasoning depth and compromise domain-specific utility. Interestingly, we identify a counter-intuitive phenomenon underlying this failure:
reasoning ability predominantly resides in parameter regions with low gradient sensitivity, contrary to the common assumption that domain capabilities correspond to high-magnitude parameters. Motivated by this insight, we propose ReasonAny, a novel merging framework that resolves the
reasoning-domain performance collapse through Contrastive Gradient Identification. Experiments across safety, biomedicine, and finance domains show that ReasonAny effectively synthesizes "
Reasoning + X" capabilities, significantly outperforming state-of-the-art baselines while retaining robust
reasoning performance.
Read the full abstract ▾
All authors Junyao Yang, Chen Qian, Dongrui Liu, Wen Shen, Yong Liu, Jing Shao
Computation and Language
Generation-Based and Emotion-Reflected Memory Update: Creating the KEEM Dataset for Better Long-Term Conversation
Jeonghyun Kang, Hongjin Kim and Harksoo Kim
In this work, we introduce the Keep Emotional and Essential
Memory (KEEM)
dataset, a novel generation-based
dataset designed to enhance
memory updates in long-term conversational systems. Unlike existing approaches that rely on simple accumulation or operation-based methods, which often result in information conflicts and difficulties in accurately tracking a user's current state, KEEM dynamically generates integrative
memories. This process not only preserves essential factual information but also incorporates emotional context and causal relationships, enabling a more nuanced understanding of user interactions. By seamlessly updating a system's
memory with both emotional and essential data, our approach promotes deeper empathy and enhances the system's ability to respond meaningfully in open-domain conversations.
Read the full abstract ▾
Computation and Language
Can Large Language Models Differentiate Harmful from Argumentative Essays? Steps Toward Ethical Essay Scoring
Hongjin Kim, Jeonghyun Kang and Harksoo Kim
This study addresses critical gaps in Automated Essay Scoring (AES) systems and Large
Language Models (LLMs) with regard to their ability to effectively identify and score harmful essays. Despite advancements in AES technology, current models often overlook ethically and morally problematic elements within essays, erroneously assigning high scores to essays that may propagate harmful opinions. In this study, we introduce the Harmful Essay Detection (HED)
benchmark, which includes essays integrating sensitive topics such as racism and gender bias, to test the efficacy of various LLMs in recognizing and scoring harmful content. Our findings reveal that: (1) LLMs require further enhancement to accurately distinguish between harmful and argumentative essays, and (2) both current AES models and LLMs fail to consider the ethical dimensions of content during scoring. The study underscores the need for developing more robust AES systems that are sensitive to the ethical implications of the content they are scoring.
Read the full abstract ▾
Computation and Language
Chaoren Wang, Heng Lu, Xueyao Zhang et al. (+4 more)
Although
Speech Large
Language Models have achieved notable progress, a substantial modality
reasoning gap remains: their
reasoning performance on
speech inputs is markedly weaker than on text. This gap could be associated with representational drift across
Transformer layers and behavior deviations in long-chain
reasoning. To address this issue, we introduce TARS, a reinforcement-learning framework that aligns text-conditioned and
speech-conditioned trajectories through an asymmetric reward design. The framework employs two dense and complementary signals: representation
alignment, which measures layer-wise hidden-state similarity between
speech- and text-conditioned trajectories, and behavior
alignment, which evaluates semantic consistency between generated outputs and reference text completions. Experiments on challenging
reasoning benchmarks, including MMSU and OBQA, show that our approach significantly narrows the modality
reasoning gap and achieves state-of-the-art performance among 7B-scale
Speech LLMs.
Read the full abstract ▾
All authors Chaoren Wang, Heng Lu, Xueyao Zhang, Shujie Liu, Yan Lu, Jinyu Li, Zhizheng Wu
Computation and Language
Double: Breaking the Acceleration Limit via Double Retrieval Speculative Parallelism
Yuhao Shen, Tianyu Liu, Junyi Shen et al. (+4 more)
Parallel Speculative Decoding (PSD) accelerates traditional Speculative Decoding (SD) by overlapping draft generation with verification. However, it remains hampered by two fundamental challenges: (1) a theoretical speedup ceiling dictated by the speed ratio between the draft and target models, and (2) high computational waste and pipeline stall due to mid-sequence
token rejections of early errors. To address these limitations, we introduce Double (Double
Retrieval Speculative Parallelism). By bridging the gap between SD and PSD, our framework resolves the
Retrieval Precision-Efficiency Dilemma through a novel synchronous mechanism. Specifically, we enable the draft model to execute iterative
retrieval speculations to break the theoretical speedup limits; to alleviate rejections without rollback, the target model performs authoritative
retrieval to generate multi-
token guidance. Double is entirely training-free and lossless. Extensive experiments demonstrate state-of-the-art speedup of $\textbf{5.3}\times$ on LLaMA3.3-70B and $\textbf{2.8}\times$ on Qwen3-32B, significantly outperforming the advanced method EAGLE-3 that requires extensive model training.
Read the full abstract ▾
All authors Yuhao Shen, Tianyu Liu, Junyi Shen, Jinyang Wu, Quan Kong, Li Huan, Cong Wang
Computation and Language
CHisAgent: A Multi-Agent Framework for Event Taxonomy Construction in Ancient Chinese Cultural Systems
Xuemei Tang, Chengxi Yan, Jinghang Gu et al. (+1 more)
Despite strong performance on many tasks, large
language models (LLMs) show limited ability in historical and cultural
reasoning, particularly in non-English contexts such as Chinese history. Taxonomic structures offer an effective mechanism to organize historical knowledge and improve understanding. However, manual taxonomy construction is costly and difficult to scale. Therefore, we propose CHisAgent, a multi-
agent LLM framework for historical taxonomy construction in ancient Chinese contexts. CHisAgent decomposes taxonomy construction into three role-specialized stages: a bottom-up Inducer that derives an initial hierarchy from raw historical corpora, a top-down Expander that introduces missing intermediate concepts using LLM world knowledge, and an evidence-guided Enricher that integrates external structured historical resources to ensure faithfulness. Using the Twenty-Four Histories, we construct a large-scale, domain-aware event taxonomy covering politics, military, diplomacy, and social life in ancient China. Extensive reference-free and reference-based evaluations demonstrate improved structural coherence and coverage, while further analysis shows that the resulting taxonomy supports cross-cultural
alignment.
Read the full abstract ▾
All authors Xuemei Tang, Chengxi Yan, Jinghang Gu, Chu-Ren Huang
Computation and Language · Cross-listed from cs.CV
Enabling Stroke-Level Structural Analysis of Hieroglyphic Scripts without Language-Specific Priors
Fuwen Luo, Zihao Wan, Ziyue Wang et al. (+6 more)
Hieroglyphs, as logographic writing systems, encode rich semantic and cultural information within their internal structural composition. Yet, current advanced Large
Language Models (LLMs) and
Multimodal LLMs (MLLMs) usually remain structurally blind to this information. LLMs process characters as textual
tokens, while MLLMs additionally view them as raw pixel grids. Both fall short to model the underlying logic of character strokes. Furthermore, existing structural analysis methods are often script-specific and labor-intensive. In this paper, we propose Hieroglyphic Stroke Analyzer (HieroSA), a novel and generalizable framework that enables MLLMs to automatically derive stroke-level structures from character bitmaps without handcrafted data. It transforms modern logographic and ancient hieroglyphs character images into explicit, interpretable line-segment representations in a normalized coordinate space, allowing for cross-lingual generalization. Extensive experiments demonstrate that HieroSA effectively captures character-internal structures and semantics, bypassing the need for
language-specific priors. Experimental results highlight the potential of our work as a graphematics analysis tool for a deeper understanding of hieroglyphic scripts. View our code at https://github.com/THUNLP-MT/HieroSA.
Read the full abstract ▾
All authors Fuwen Luo, Zihao Wan, Ziyue Wang, Yaluo Liu, Pau Tong Lin Xu, Xuanjia Qiao, Xiaolong Wang, Peng Li, Yang Liu
Computation and Language
FlashMem: Distilling Intrinsic Latent Memory via Computation Reuse
Yubo Hou, Zhisheng Chen, Tao Wan et al. (+1 more)
The stateless architecture of Large
Language Models inherently lacks the mechanism to preserve dynamic context, compelling
agents to redundantly reprocess history to maintain long-horizon autonomy. While latent
memory offers a solution, current approaches are hindered by architectural segregation, relying on auxiliary encoders that decouple
memory from the
reasoning backbone. We propose FlashMem, a framework that distills intrinsic
memory directly from transient
reasoning states via computation reuse. Leveraging the property that internal representations uniquely encode input trajectories, FlashMem identifies the last hidden state as a sufficient statistic for the interaction history. This enables a Shared-KV Consolidator to synthesize
memory by attending directly to the backbone's frozen cache, eliminating redundant re-parameterization. Furthermore, a parameter-free Cognitive Monitor leverages
attention entropy to adaptively trigger consolidation only when high epistemic uncertainty is detected. Experiments demonstrate that FlashMem matches the performance of heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition.
Read the full abstract ▾
All authors Yubo Hou, Zhisheng Chen, Tao Wan, Zengchang Qin
Computation and Language · Cross-listed from cs.LG
Hi-ZFO: Hierarchical Zeroth- and First-Order LLM Fine-Tuning via Importance-Guided Tensor Selection
Feihu Jin and Ying Tan
Fine-tuning large
language models (LLMs) using standard first-order (FO) optimization often drives training toward sharp, poorly generalizing minima. Conversely, zeroth-order (ZO) methods offer stronger exploratory behavior without relying on explicit gradients, yet suffer from slow convergence. More critically, our analysis reveals that in generative tasks, the vast output and search space significantly amplify estimation variance, rendering ZO methods both noisy and inefficient. To address these challenges, we propose Hi-ZFO (Hierarchical Zeroth- and First-Order optimization), a hybrid framework designed to synergize the precision of FO gradients with the exploratory capability of ZO estimation. Hi-ZFO adaptively partitions the model through layer-wise importance profiling, applying precise FO updates to critical layers while leveraging ZO optimization for less sensitive ones. Notably, ZO in Hi-ZFO is not merely a
memory-saving surrogate; it is intentionally introduced as a source of "beneficial stochasticity" to help the model escape the local minima where pure FO optimization tends to stagnate. Validated across diverse generative, mathematical, and code
reasoning tasks, Hi-ZFO consistently achieves superior performance while significantly reducing the training time. These results demonstrate the effectiveness of hierarchical hybrid optimization for LLM fine-tuning.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
MMViR: A Multi-Modal and Multi-Granularity Representation for Long-range Video Understanding
Zizhong Li, Haopeng Zhang and Jiawei Zhang
Long
videos, ranging from minutes to hours, present significant challenges for current Multi-modal Large
Language Models (MLLMs) due to their complex events, diverse scenes, and long-range dependencies. Direct encoding of such
videos is computationally too expensive, while simple
video-to-text conversion often results in redundant or fragmented content. To address these limitations, we introduce MMViR, a novel multi-modal, multi-grained structured representation for long
video understanding. MMViR identifies key turning points to segment the
video and constructs a three-level description that couples global narratives with fine-grained visual details. This design supports efficient query-based
retrieval and generalizes well across various scenarios. Extensive evaluations across three tasks, including QA, summarization, and
retrieval, show that MMViR outperforms the prior strongest method, achieving a 19.67% improvement in hour-long
video understanding while reducing processing latency to 45.4% of the original.
Read the full abstract ▾
Computation and Language
MemBuilder: Reinforcing LLMs for Long-Term Memory Construction via Attributed Dense Rewards
Zhiyu Shen, Ziming Wu, Fuming Lai et al. (+2 more)
Maintaining consistency in long-term dialogues remains a fundamental challenge for LLMs, as standard
retrieval mechanisms often fail to capture the temporal
evolution of historical states. While
memory-augmented frameworks offer a structured alternative, current systems rely on static prompting of closed-source models or suffer from ineffective training paradigms with sparse rewards. We introduce MemBuilder, a reinforcement learning framework that trains models to orchestrate multi-dimensional
memory construction with attributed dense rewards. MemBuilder addresses two key challenges: (1) Sparse Trajectory-Level Rewards: we employ synthetic session-level question generation to provide dense intermediate rewards across extended trajectories; and (2) Multi-Dimensional
Memory Attribution: we introduce contribution-aware gradient weighting that scales policy updates based on each component's downstream impact. Experimental results show that MemBuilder enables a 4B-parameter model to outperform state-of-the-art closed-source baselines, exhibiting strong generalization across long-term dialogue
benchmarks.
Read the full abstract ▾
All authors Zhiyu Shen, Ziming Wu, Fuming Lai, Shaobing Lian, Yanghui Rao
Computation and Language
The Facade of Truth: Uncovering and Mitigating LLM Susceptibility to Deceptive Evidence
Herun Wan, Jiaying Wu, Minnan Luo et al. (+3 more)
To reliably assist human decision-making, LLMs must maintain factual internal beliefs against misleading injections. While current models resist explicit misinformation, we uncover a fundamental vulnerability to sophisticated, hard-to-falsify evidence. To systematically probe this weakness, we introduce MisBelief, a framework that generates misleading evidence via collaborative, multi-round interactions among multi-role LLMs. This process mimics subtle, defeasible
reasoning and progressive refinement to create logically persuasive yet factually deceptive claims. Using MisBelief, we generate 4,800 instances across three difficulty levels to evaluate 7 representative LLMs. Results indicate that while models are robust to direct misinformation, they are highly sensitive to this refined evidence: belief scores in falsehoods increase by an average of 93.0%, fundamentally compromising downstream recommendations. To address this, we propose Deceptive Intent Shielding (DIS), a governance mechanism that provides an early warning signal by inferring the deceptive intent behind evidence. Empirical results demonstrate that DIS consistently mitigates belief shifts and promotes more cautious evidence evaluation.
Read the full abstract ▾
All authors Herun Wan, Jiaying Wu, Minnan Luo, Fanxiao Li, Zhi Zeng, Min-Yen Kan
Computation and Language · Cross-listed from cs.LG
MaxCode: A Max-Reward Reinforcement Learning Framework for Automated Code Optimization
Jiefu Ou, Sapana Chaudhary, Kaj Bostrom et al. (+4 more)
Large
Language Models (LLMs) demonstrate strong capabilities in general coding tasks but encounter two key challenges when optimizing code: (i) the complexity of writing optimized code (such as performant CUDA kernels and competition-level CPU code) requires expertise in systems,
algorithms and specific
languages and (ii) requires interpretation of performance metrics like timing and device utilization beyond binary correctness. In this work, we explore inference-time search
algorithms that guide the LLM to discover better solutions through iterative refinement based on execution feedback. Our approach, called MaxCode unifies existing search methods under a max-reward reinforcement learning framework, making the observation and action-value functions modular for modification. To enhance the observation space, we integrate a natural
language critique model that converts raw execution feedback into diagnostic insights about errors and performance bottlenecks, and the best-discounted reward seen so far. Together, these provide richer input to the code proposal function. To improve exploration during search, we train a generative reward-to-go model using action values from rollouts to rerank potential solutions. Testing on the KernelBench (CUDA) and PIE (C++) optimization
benchmarks shows that MaxCode improves optimized code performance compared to baselines, achieving 20.3% and 10.1% relative improvements in absolute speedup value and relative speedup ranking, respectively.
Read the full abstract ▾
All authors Jiefu Ou, Sapana Chaudhary, Kaj Bostrom, Nathaniel Weir, Shuai Zhang, Huzefa Rangwala, George Karypis
Computation and Language
Towards Valid Student Simulation with Large Language Models
Zhihao Yuan, Yunze Xiao, Ming Li et al. (+4 more)
This paper presents a conceptual and methodological framework for large
language model (LLM) based student simulation in educational settings. The authors identify a core failure mode, termed the "competence paradox" in which broadly capable LLMs are asked to emulate partially knowledgeable learners, leading to unrealistic error patterns and learning dynamics. To address this, the paper reframes student simulation as a constrained generation problem governed by an explicit Epistemic State Specification (ESS), which defines what a simulated learner can access, how errors are structured, and how learner state evolves over time. The work further introduces a Goal-by-Environment framework to situate simulated student systems according to behavioral objectives and deployment contexts. Rather than proposing a new system or
benchmark, the paper synthesizes prior literature, formalizes key design dimensions, and articulates open challenges related to validity, evaluation, and ethical risks. Overall, the paper argues for epistemic fidelity over surface realism as a prerequisite for using LLM-based simulated students as reliable scientific and pedagogical instruments.
Read the full abstract ▾
All authors Zhihao Yuan, Yunze Xiao, Ming Li, Weihao Xuan, Richard Tong, Mona Diab, Tom Mitchell
Computation and Language · Cross-listed from cs.CV
ROAP: A Reading-Order and Attention-Prior Pipeline for Optimizing Layout Transformers in Key Information Extraction
Tingwei Xie, Jinxin He and Yonghong Song
The efficacy of
Multimodal Transformers in visually-rich document understanding (VrDU) is critically constrained by two inherent limitations: the lack of explicit modeling for logical reading order and the interference of visual
tokens that dilutes
attention on textual semantics. To address these challenges, this paper presents ROAP, a lightweight and architecture-agnostic pipeline designed to optimize
attention distributions in Layout
Transformers without altering their pre-trained backbones. The proposed pipeline first employs an Adaptive-XY-Gap (AXG-Tree) to robustly extract hierarchical reading sequences from complex layouts. These sequences are then integrated into the
attention mechanism via a Reading-Order-Aware Relative Position Bias (RO-RPB). Furthermore, a Textual-
Token Sub-block
Attention Prior (TT-Prior) is introduced to adaptively suppress visual noise and enhance fine-grained text-text interactions. Extensive experiments on the FUNSD and CORD
benchmarks demonstrate that ROAP consistently improves the performance of representative backbones, including LayoutLMv3 and GeoLayoutLM. These findings confirm that explicitly modeling reading logic and regulating modality interference are critical for robust document understanding, offering a scalable solution for complex layout analysis. The implementation code will be released at https://github.com/KevinYuLei/ROAP.
Read the full abstract ▾
Computation and Language
Do LLMs Need Inherent Reasoning Before Reinforcement Learning? A Study in Korean Self-Correction
Hongjin Kim, Jaewook Lee, Kiyoung Lee et al. (+3 more)
Large
Language Models (LLMs) demonstrate strong
reasoning and self-correction abilities in high-resource
languages like English, but their performance remains limited in low-resource
languages such as Korean. In this study, we investigate whether reinforcement learning (RL) can enhance Korean
reasoning abilities to a degree comparable to English. Our findings reveal that RL alone yields limited improvements when applied to models lacking inherent Korean
reasoning capabilities. To address this, we explore several fine-tuning strategies and show that aligning the model's internal
reasoning processes with Korean inputs-particularly by tuning Korean-specific
neurons in early layers-is key to unlocking RL's effectiveness. We introduce a self-correction code-switching
dataset to facilitate this
alignment and observe significant performance gains in both mathematical
reasoning and self-correction tasks. Ultimately, we conclude that the crucial factor in
multilingual reasoning enhancement is not injecting new linguistic knowledge, but effectively eliciting and aligning existing
reasoning capabilities. Our study provides a new perspective on how internal translation and
neuron-level tuning contribute to
multilingual reasoning alignment in LLMs.
Read the full abstract ▾
All authors Hongjin Kim, Jaewook Lee, Kiyoung Lee, Jong-hun Shin, Soojong Lim, Oh-Woog Kwon
Computation and Language · Cross-listed from cs.LG
RingSQL: Schema-Independent Synthetic Data Generation for Text-to-SQL Reinforcement Learning
Marko Sterbentz, Kevin Cushing, Cameron Barrie et al. (+1 more)
Recent advances in text-to-SQL have been driven by larger models, better
datasets, and new training methods like RLVR. However, progress remains limited by scarce high-quality training data, a problem RLVR is especially sensitive to since noisy data can produce spurious rewards. Manual data creation is expensive, and existing synthetic methods trade off reliability for scalability: template-based approaches guarantee correct SQL but need schema-specific templates and lack diversity, while LLM-based generation scales easily but lacks quality guarantees. We introduce RingSQL, a hybrid framework for generating question-SQL pairs that combines schema-independent query templates with LLM-based paraphrasing of natural
language questions. By grounding question generation in complete template questions, RingSQL preserves question-query correctness across all levels of query complexity, a property purely LLM-based methods fail to maintain. RingSQL also produces the only synthetic
dataset that improves RLVR training performance across all tested model architectures and
benchmarks, achieving 69.8% average accuracy and surpassing both the next-best synthetic
dataset by 2.1% and human-annotated data from Spider and BIRD. Code and data are available at https://github.com/nu-c3lab/RingSQL.
Read the full abstract ▾
All authors Marko Sterbentz, Kevin Cushing, Cameron Barrie, Kristian J. Hammond
Computation and Language
Tracing Moral Foundations in Large Language Models
Chenxiao Yu, Bowen Yi, Farzan Karimi-Malekabadi et al. (+5 more)
Large
language models often produce human-like moral judgments, but it is unclear whether this reflects an internal conceptual structure or superficial “moral mimicry.” Using Moral Foundations Theory (MFT) as an analytic framework, we study how moral foundations are encoded, organized, and expressed across 14 base and instruction-tuned LLMs spanning four model families (Llama, Qwen2.5, Qwen3-MoE, Mistral) and scales from 7B to 70B. We employ a multi-level approach combining (i) layer-wise analysis of MFT concept representations and their
alignment with human moral perceptions, (ii) pretrained sparse autoencoders (SAEs) over the residual stream to identify sparse features that support moral concepts, and (iii) causal steering interventions using dense MFT vectors and sparse SAE features. We find that models represent and distinguish moral foundations in a manner that aligns with human judgments, and that this moral geometry naturally emerges from pretraining and is selectively rewired by post-training. At a finer scale, SAE features show clear semantic links to specific foundations, suggesting partially disentangled mechanisms within shared representations. Finally, steering along either dense vectors or sparse features produces predictable shifts in foundation-relevant behavior, demonstrating a causal connection between internal representations and moral outputs. Together, our results provide mechanistic evidence that moral concepts in LLMs are distributed, layered, and partly disentangled, suggesting that pluralistic moral structure can emerge as a latent pattern from the statistical regularities of
language alone.
Read the full abstract ▾
All authors Chenxiao Yu, Bowen Yi, Farzan Karimi-Malekabadi, Suhaib Abdurahman, Jinyi Ye, Shrikanth Narayanan, Yue Zhao, Morteza Dehghani