Computation and Language
Self-Evolving LLM Memory Extraction Across Heterogeneous Tasks
Yuqing Yang, Tengxiao Liu, Wang Bill Zhu et al. (+3 more)
As LLM-based assistants become persistent and personalized, they must extract and retain useful information from past conversations as
memory. However, the types of information worth remembering vary considerably across tasks. We formalize the heterogeneous
memory extraction task and introduce BEHEMOTH, a
benchmark that repurposes 18 existing
datasets spanning personalization, problem-solving, and agentic tasks, using a downstream utility-driven metric for systematic evaluation. Our empirical analysis confirms that no single static extraction
prompt dominates across all task categories, and that existing self-evolving
prompt optimization frameworks, originally designed for homogeneous distributions, degrade when training tasks are heterogeneous. To address this, we propose CluE, a cluster-based self-evolving strategy that groups training examples into clusters by extraction scenarios, analyzes each cluster independently, and synthesizes cross-cluster insights to update the extraction
prompt. Experiments on BEHEMOTH show that CluE generalizes effectively across heterogeneous tasks ($+$9.04% relative gain), consistently outperforming prior self-evolving frameworks.
Read the full abstract ▾
All authors Yuqing Yang, Tengxiao Liu, Wang Bill Zhu, Taiwei Shi, Linxin Song, Robin Jia
Computation and Language
A Triadic Suffix Tokenization Scheme for Numerical Reasoning
Olga Chetverina
Standard subword tokenization methods fragment numbers inconsistently, causing large
language models (LLMs) to lose positional and decimal structure - a primary driver of errors in arithmetic and scientific
reasoning. We introduce Triadic Suffix Tokenization (TST), a deterministic scheme that partitions digits into three-digit triads and annotates each triad with an explicit magnitude marker. Critically, the scheme defines a fixed, one-to-one mapping between suffixes and orders of magnitude for the integer part (thousands, millions, billions, etc.) and a parallel system of replicated markers for fractional depth (tenths, thousandths, millionths, etc.). Unlike approaches that rely on positional inference, this method provides a consistent gradient signal, which should ensure stable convergence. Two implementation variants are proposed: (1) a vocabulary-based approach that adds at most 10,000 fixed
tokens to an existing vocabulary, covering 33 orders of magnitude ($10^{-15}$ to $10^{18}$); and (2) a suffix-marker approach that uses a small set of special
tokens to denote magnitude dynamically. Both variants preserve exact digits while making order-of-magnitude relationships transparent at the
token level. While we focus on 3-digit groups (Triadic), the framework is inherently scalable to any group size for precise vocabulary optimization. Furthermore, it allows for linear vocabulary expansion to accommodate arbitrary precision and range. TST is architecture-agnostic and can be integrated as a drop-in preprocessing step. Experimental validation is deferred to future work.
Read the full abstract ▾
Computation and Language
Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking
Solomon Messing
LLM evaluations drive which models get deployed, what safety standards get adopted, which research conclusions get published, and how projections of AI's labor-
market impact get made. Yet standard confidence intervals ignore variability from judge model choice, model temperature, and
prompt phrasing, producing under-coverage that worsens with more data. The omitted variance can shift results enough to reverse conclusions \citep{baumann2025llmhacking, huang2026dropping}; pipelines that fail to average over it leave the surface that “
benchmark hacking” exploits \citep{singh2025leaderboard}. This paper decomposes LLM pipeline uncertainty into its sources, distinguishes variance that shrinks with more data from sensitivity to researcher design choices, and uses design-study projections to reduce total evaluation error (TEE). Across the demonstrations, naive standard errors are 40 - 60% smaller than the TEE-corrected SE. Using Chatbot Arena data, we show naive 95% CI coverage drops as $n$ grows while TEE-corrected coverage holds at 95%, and TEE-guided pipelines restrict the
benchmark gaming surface from 56 to 32 Elo ($K=27$), below the human-leaderboard baseline. We show further that a small pilot recovers honest CIs and projects which design changes most improve precision. Acting on those projections halves MMLU estimation error against the answer key at equivalent cost, and raises per-match agreement with human votes by 7.9 percentage points on Chatbot Arena.
Read the full abstract ▾
Computation and Language
MIXAR: Scaling Autoregressive Pixel-based Language Models to Multiple Languages and Scripts
Chen Hu, Yintao Tai, Antonio Vergari et al. (+2 more)
Pixel-based
language models are gaining momentum as alternatives to traditional
token-based approaches, promising to circumvent tokenization challenges; however, existing models remain English-centric and are hard to train as generators. This paper introduces MIXAR, the first generative
multilingual pixel-based
language model, pretrained on eight high-resource
languages spanning Latin and CJK scripts and scaled up to 0.5B parameters. Although pretrained on high-resource
languages only, MIXAR generalizes to unseen low-resource
languages better than a tokenizer-based model pretrained on the same corpus, with its largest gains on the scripts most distant from the pretraining mix, taking a step toward more equitable
language technology. To make free generation work in pixel space, MIXAR replaces the notoriously unstable
adversarial training stage of prior pixel-based models with a GRPO-based reinforcement-learning stage that rewards the model on its own free-running generations: it proves more stable and multiplies free-generation accuracy by up to 2.3x on
multilingual benchmarks. Finally, MIXAR performs on par with comparable tokenizer-based models on discriminative
benchmarks, improves consistently with scale, and withstands orthographic attacks more gracefully than tokenizer-based models. Code and models are available at https://github.com/april-tools/mixar.
Read the full abstract ▾
All authors Chen Hu, Yintao Tai, Antonio Vergari, Frank Keller, Alessandro Suglia
Computation and Language
Phonological distances for linguistic typology and the origin of Indo-European languages
Marius Mavridis, Juan De Gregorio, Raul Toral et al. (+1 more)
We show that short-range phoneme dependencies encode large-scale patterns of linguistic relatedness, with direct implications for quantitative typology and evolutionary linguistics. Specifically, using an information-theoretic framework, we argue that phoneme sequences modeled as second-order Markov chains essentially capture the statistical correlations of a phonological system. This finding enables us to quantify distances among 67 modern
languages from a
multilingual parallel corpus employing a distance metric that incorporates articulatory features of phonemes. The resulting phonological distance
matrix recovers major
language families and reveals signatures of contact-induced convergence. Remarkably, we obtain a clear correlation with geographic distance, allowing us to constrain a plausible homeland region for the Indo-European family, consistent with the Steppe hypothesis.
Read the full abstract ▾
All authors Marius Mavridis, Juan De Gregorio, Raul Toral, David Sanchez
Computation and Language
Synthius-Mem: Brain-Inspired Hallucination-Resistant Persona Memory Achieving 94.4% Memory Accuracy and 99.6% Adversarial Robustness on LoCoMo
Artem Gadzhiev and Andrew Kislov
Providing AI
agents with reliable long-term
memory that does not hallucinate remains an open problem. Current approaches to
memory for LLM
agents – sliding windows, summarization,
embedding-based RAG, and flat fact extraction – each reduce
token cost but introduce catastrophic information loss, semantic drift, or uncontrolled
hallucination about the user. The structural reason is architectural: every published
memory system on the LoCoMo
benchmark treats conversation as a
retrieval problem over raw or lightly summarized dialogue segments, and none reports
adversarial robustness, the ability to refuse questions about facts the user never disclosed. We present Synthius-Mem, a
brain-inspired structured persona
memory system that takes a fundamentally different approach. Instead of retrieving what was said, Synthius-Mem extracts what is known about the person: a full persona extraction pipeline decomposes conversations into six cognitive domains (biography, experiences, preferences, social circle, work, psychometrics), consolidates and deduplicates per domain, and retrieves structured facts via CategoryRAG at 21.79 ms latency. On the LoCoMo
benchmark (ACL 2024, 10 conversations, 1,813 questions), Synthius-Mem achieves 94.37% accuracy, exceeding all published systems including MemMachine (91.69%,
adversarial score is not reported) and human performance (87.9 F1). Core
memory fact accuracy reaches 98.64%.
Adversarial robustness, the
hallucination resistance metric that no competing system reports, reaches 99.55%. Synthius-Mem reduces
token consumption by ~5x compared to full-context replay while achieving higher accuracy. Synthius-Mem achieves state-of-the-art results on LoCoMo and is, to our knowledge, the only persona
memory system that both exceeds human-level performance and reports
adversarial robustness.
Read the full abstract ▾
Computation and Language
Relax: An Asynchronous Reinforcement Learning Engine for Omni-Modal Post-Training at Scale
Liujie Zhang, Benzhe Ning, Rui Yang et al. (+8 more)
Reinforcement learning (RL) post-training has proven effective at unlocking
reasoning, self-reflection, and tool-use capabilities in large
language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness – throughput tradeoff. We present Relax (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an omni-native architecture builds
multimodal support into the full stack – from data preprocessing and modality-aware parallelism to inference generation – rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20$\times$ end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76$\times$ speedup over colocate on Qwen3-4B and a 2.00$\times$ speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~\cite{ma2025r3} for MoE models with only 1.9% overhead, compared to 32% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on
video without degradation. Relax is available at https://github.com/rednote-ai/Relax.
Read the full abstract ▾
All authors Liujie Zhang, Benzhe Ning, Rui Yang, Xiaoyan Yu, Jiaxing Li, Lumeng Wu, Jia Liu, Minghao Li, Weihang Chen, Weiqi Hu, Lei Zhang
Computation and Language · Cross-listed from cs.SD
MimicLM: Zero-Shot Voice Imitation through Autoregressive Modeling of Pseudo-Parallel Speech Corpora
Tao Feng, Yuxiang Wang, Yuancheng Wang et al. (+5 more)
Voice imitation aims to transform source
speech to match a reference speaker's timbre and speaking style while preserving linguistic content. A straightforward approach is to train on triplets of (source, reference, target), where source and target share the same content but target matches the reference's voice characteristics, yet such data is extremely scarce. Existing approaches either employ carefully designed disentanglement architectures to bypass this data scarcity or leverage external systems to synthesize pseudo-parallel training data. However, the former requires intricate model design, and the latter faces a quality ceiling when synthetic
speech is used as training targets. To address these limitations, we propose MimicLM, which takes a novel approach by using synthetic
speech as training sources while retaining real recordings as targets. This design enables the model to learn directly from real
speech distributions, breaking the synthetic quality ceiling. Building on this data construction approach, we incorporate interleaved text-audio modeling to guide the generation of content-accurate
speech and apply post-training with preference
alignment to mitigate the inherent distributional mismatch when training on synthetic data. Experiments demonstrate that MimicLM achieves superior voice imitation quality with a simple yet effective architecture, significantly outperforming existing methods in naturalness while maintaining competitive similarity scores across speaker identity, accent, and
emotion dimensions.
Read the full abstract ▾
All authors Tao Feng, Yuxiang Wang, Yuancheng Wang, Xueyao Zhang, Dekun Chen, Chaoren Wang, Xun Guan, Zhizheng Wu
Computation and Language · Cross-listed from cs.LG
Eliciting Medical Reasoning with Knowledge-enhanced Data Synthesis: A Semi-Supervised Reinforcement Learning Approach
Haolin Li, Shuyang Jiang, Ruipeng Zhang et al. (+3 more)
While large
language models hold promise for complex medical applications, their development is hindered by the scarcity of high-quality
reasoning data. To address this issue, existing approaches typically distill chain-of-thought
reasoning traces from large proprietary models via supervised fine-tuning, then conduct reinforcement learning (RL). These methods exhibit limited improvement on underrepresented domains like rare diseases while incurring substantial costs from generating complex
reasoning chains. To efficiently enhance medical
reasoning, we propose MedSSR, a Medical Knowledge-enhanced data Synthesis and Semi-supervised Reinforcement learning framework. Our framework first employs rare disease knowledge to synthesize distribution-controllable
reasoning questions. We then utilize the policy model itself to generate high-quality pseudo-labels. This enables a two-stage, intrinsic-to-extrinsic training paradigm: self-supervised RL on the pseudo-labeled synthetic data, followed by supervised RL on the human-annotated real data. MedSSR scales model training efficiently without relying on costly trace
distillation. Extensive experiments on Qwen and Llama demonstrate that our method outperforms existing methods across ten medical
benchmarks, achieving up to +5.93% gain on rare-disease tasks. Our code is available at https://github.com/tdlhl/MedSSR.
Read the full abstract ▾
All authors Haolin Li, Shuyang Jiang, Ruipeng Zhang, Jiangchao Yao, Ya Zhang, Yanfeng Wang
Computation and Language
Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory
Weixian Waylon Li, Jiaxin Zhang, Xianan Jim Yang et al. (+2 more)
Structured
memory representations such as knowledge graphs are central to autonomous
agents and other long-lived systems. However, most existing approaches model time as discrete metadata, either sorting by recency (burying old-yet-permanent knowledge), simply overwriting outdated facts, or requiring an expensive LLM call at every ingestion step, leaving them unable to distinguish persistent facts from evolving ones. To address this, we introduce RoMem, a drop-in temporal knowledge graph module for structured
memory systems, applicable to agentic
memory and beyond. A pretrained Semantic Speed Gate maps each relation's text
embedding to a
volatility score, learning from data that evolving relations (e.g., "president of") should rotate fast while persistent ones (e.g., "born in") should remain stable. Combined with continuous phase rotation, this enables
geometric shadowing: obsolete facts are rotated out of phase in complex vector space, so temporally correct facts naturally outrank contradictions without deletion. On temporal knowledge graph completion, RoMem achieves state-of-the-art results on ICEWS05-15 (72.6 MRR). Applied to agentic
memory, it delivers 2-3x MRR and answer accuracy on temporal
reasoning (MultiTQ), dominates hybrid
benchmark (LoCoMo), preserves static
memory with zero degradation (DMR-MSC), and generalises zero-shot to unseen financial domains (FinTMMBench).
Read the full abstract ▾
All authors Weixian Waylon Li, Jiaxin Zhang, Xianan Jim Yang, Tiejun Ma, Yiwen Guo
Computation and Language
NovBench: Evaluating Large Language Models on Academic Paper Novelty Assessment
Wenqing Wu, Yi Zhao, Yuzhuo Wang et al. (+4 more)
Novelty is a core requirement in academic publishing and a central focus of peer review, yet the growing volume of submissions has placed increasing pressure on human reviewers. While large
language models (LLMs), including those fine-tuned on peer review data, have shown promise in generating review comments, the absence of a dedicated
benchmark has limited systematic evaluation of their ability to assess research novelty. To address this gap, we introduce NovBench, the first large-scale
benchmark designed to evaluate LLMs' capability to generate novelty evaluations in support of human peer review. NovBench comprises 1,684 paper-review pairs from a leading NLP conference, including novelty descriptions extracted from paper introductions and corresponding expert-written novelty evaluations. We focus on both sources because the introduction provides a standardized and explicit articulation of novelty claims, while expert-written novelty evaluations constitute one of the current gold standards of human judgment. Furthermore, we propose a four-dimensional evaluation framework (including Relevance, Correctness, Coverage, and Clarity) to assess the quality of LLM-generated novelty evaluations. Extensive experiments on both general and specialized LLMs under different prompting strategies reveal that current models exhibit limited understanding of scientific novelty, and that fine–tuned models often suffer from instruction-following deficiencies. These findings underscore the need for targeted fine-tuning strategies that jointly improve novelty comprehension and instruction adherence.
Read the full abstract ▾
All authors Wenqing Wu, Yi Zhao, Yuzhuo Wang, Siyou Li, Juexi Shao, Yunfei Long, Chengzhi Zhang
Computation and Language
Triviality Corrected Endogenous Reward
Xinda Wang, Zhengxu Hou, Yangshijie Zhang et al. (+6 more)
Reinforcement learning for open-ended text generation is constrained by the lack of verifiable rewards, necessitating reliance on judge models that require either annotated data or powerful closed-source models. Inspired by recent work on unsupervised reinforcement learning for mathematical
reasoning using confidence-based endogenous rewards, we investigate whether this principle can be adapted to open-ended writing tasks. We find that directly applying confidence rewards leads to Triviality Bias: the policy collapses toward high-probability outputs, reducing diversity and meaningful content. We propose TCER (Triviality Corrected Endogenous Reward), which addresses this bias by rewarding the relative information gain between a specialist policy and a generalist reference policy, modulated by a probability-dependent correction mechanism. Across multiple writing
benchmarks and model architectures, TCER achieves consistent improvements without external supervision. Furthermore, TCER also transfers effectively to mathematical
reasoning, validating the generality of our approach across different generation tasks.
Read the full abstract ▾
All authors Xinda Wang, Zhengxu Hou, Yangshijie Zhang, Bingren Yan, Jialin Liu, Chenzhuo Zhao, Zhibo Yang, Bin-Bin Yang, Feng Xiao
Computation and Language · Cross-listed from cs.SE
DuET: Dual Execution for Test Output Prediction with Generated Code and Pseudocode
Hojae Han, Jaejin Kim, Seung-won Hwang et al. (+2 more)
This work addresses test output prediction, a key challenge in test case generation. To improve the reliability of predicted outputs by LLMs, prior approaches generate code first to ground predictions. One grounding strategy is direct execution of generated code, but even minor errors can cause failures. To address this, we introduce LLM-based pseudocode execution, which grounds prediction on more error-resilient pseudocode and simulates execution via LLM
reasoning. We further propose DuET, a dual-execution framework that combines both approaches by functional majority voting. Our analysis shows the two approaches are complementary in overcoming the limitations of direct execution suffering from code errors, and pseudocode
reasoning from
hallucination. On LiveCodeBench, DuET achieves the state-of-the-art performance, improving Pass@1 by 13.6 pp.
Read the full abstract ▾
All authors Hojae Han, Jaejin Kim, Seung-won Hwang, Yu Jin Kim, Moontae Lee
Computation and Language
Policy Split: Incentivizing Dual-Mode Exploration in LLM Reinforcement with Dual-Mode Entropy Regularization
Jiashu Yao, Heyan Huang, Daiqing Wu et al. (+2 more)
To encourage diverse exploration in reinforcement learning (RL) for large
language models (LLMs) without compromising accuracy, we propose Policy Split, a novel paradigm that bifurcates the policy into normal and high-
entropy modes with a high-
entropy prompt. While sharing model parameters, the two modes undergo collaborative dual-mode
entropy regularization tailored to distinct objectives. Specifically, the normal mode optimizes for task correctness, while the high-
entropy mode incorporates a preference for exploration, and the two modes learn collaboratively. Extensive experiments demonstrate that our approach consistently outperforms established
entropy-guided RL baselines across various model sizes in general and creative tasks. Further analysis reveals that Policy Split facilitates dual-mode exploration, where the high-
entropy mode generates distinct behavioral patterns to the normal mode, providing unique learning signals.
Read the full abstract ▾
All authors Jiashu Yao, Heyan Huang, Daiqing Wu, Zeming Liu, Yuhang Guo
Computation and Language
METER: Evaluating Multi-Level Contextual Causal Reasoning in Large Language Models
Pengfeng Li, Chen Huang, Chaoqun Hao et al. (+4 more)
Contextual causal
reasoning is a critical yet challenging capability for Large
Language Models (LLMs). Existing
benchmarks, however, often evaluate this skill in fragmented settings, failing to ensure context consistency or cover the full causal hierarchy. To address this, we pioneer METER to systematically
benchmark LLMs across all three levels of the causal ladder under a unified context setting. Our extensive evaluation of various LLMs reveals a significant decline in proficiency as tasks ascend the causal hierarchy. To diagnose this degradation, we conduct a deep mechanistic analysis via both error pattern identification and internal information flow tracing. Our analysis reveals two primary failure modes: (1) LLMs are susceptible to distraction by causally irrelevant but factually correct information at lower level of causality; and (2) as tasks ascend the causal hierarchy, faithfulness to the provided context degrades, leading to a reduced performance. We belive our work advances our understanding of the mechanisms behind LLM contextual causal
reasoning and establishes a critical foundation for future research. Our code and
dataset are available at https://github.com/SCUNLP/METER .
Read the full abstract ▾
All authors Pengfeng Li, Chen Huang, Chaoqun Hao, Hongyao Chen, Xiao-Yong Wei, Wenqiang Lei, See-Kiong Ng
Computation and Language · Cross-listed from cs.LG
Attention-Weighted Value Projection for KV-Cache Compression
Samuel Salfati
Rank reduction discards dimensions; quantization keeps them at lower precision. Comparing the two requires a choice of what compression should preserve. For
attention values, we study reconstruction of the
attention output rather than reconstruction of the values alone. With fixed
attention weights, the optimal orthogonal rank-$r$ projection uses the leading eigenvectors of $V^\top\alpha^\top\alpha V$, and its error is exactly the discarded
eigenvalue sum. We extend this objective to calibration
datasets and grouped query
attention, and describe rank allocation under an additive local error budget. We also examine the limits of using local error to predict downstream loss. Historical weight-perturbation experiments favor coefficient rounding over the value projections tested, but do not establish a comparison at equal cache storage. The resulting distinction is practical: the projection objective has an exact solution, while a comparison with cache quantization requires separate activation-level experiments.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
Revisiting Compositionality in Dual-Encoder Vision-Language Models: The Role of Inference
Imanol Miranda, Ander Salaberria, Eneko Agirre et al. (+1 more)
Dual-encoder
Vision-
Language Models (VLMs) such as CLIP are often characterized as bag-of-words systems due to their poor performance on compositional
benchmarks. We argue that this limitation may stem less from deficient representations than from the standard inference protocol based on global cosine similarity. First, through controlled diagnostic experiments, we show that explicitly enforcing fine-grained region-segment
alignment at inference dramatically improves compositional performance without updating pretrained encoders. We then introduce a lightweight
transformer that learns such
alignments directly from frozen patch and
token embeddings. Comparing against full fine-tuning and prior end-to-end compositional training methods, we find that although these approaches improve in-domain
retrieval, their gains do not consistently transfer under distribution shift. In contrast, learning localized
alignment over frozen representations matches full fine-tuning on in-domain
retrieval while yielding substantial improvements on controlled out-of-domain compositional
benchmarks. These results identify global
embedding matching as a key bottleneck in dual-encoder VLMs and highlight the importance of
alignment mechanisms for robust compositional generalization.
Read the full abstract ▾
All authors Imanol Miranda, Ander Salaberria, Eneko Agirre, Gorka Azkune
Computation and Language · Cross-listed from cs.AI
Samuel Cahyawijaya, Peerat Limkonchotiwat, Tack Hwa Wong et al. (+45 more)
While the field of
vision-
language (VL) has achieved remarkable success in integrating visual and textual information across multiple
languages and domains, there is still no dedicated framework for assessing human-centric
alignment in
vision-
language systems. We offer two contributions to address this gap. First, we introduce Anthropogenic Regional Adaptation: a novel paradigm that aims to optimize model relevance to specific regional contexts while ensuring the retention of global generalization capabilities. Second, we present a simple, but effective adaptation method named Geographical-generalization-made-easy (GG-EZ), which utilizes regional data filtering and model merging. Through comprehensive experiments on 3 VL architectures: large
vision-
language models, text-to-image
diffusion models, and
vision-
language embedding models, and a case study in Southeast Asia (SEA) regional adaptation, we demonstrate the importance of Anthropogenic Regional Adaptation and the effectiveness of GG-EZ, showing 5-15% gains in cultural relevance metrics across SEA while maintaining over 98% of global performance and even occasionally surpassing it. Our findings establish Anthropogenic Regional
Alignment as a foundational paradigm towards applicability of
multimodal vision-
language models in diverse regions and demonstrate a simple-yet-effective baseline method that optimizes regional value
alignment while preserving global generalization.
Read the full abstract ▾
All authors Samuel Cahyawijaya, Peerat Limkonchotiwat, Tack Hwa Wong, Hitesh Laxmichand Patel, Amit Agarwal, Manuel Antonio Rufino, Carlos Rafael Catalan, Muhammad Reza Qorib, Vicky Feliren, Holy Lovenia, Aye Hninn Khine, Frederikus Hudi, David Anugraha, Alham Fikri Aji, Romrawin Chumpu, Viet-Thanh Pham, Minghan Wang, Mohamed Fazli Imam, Ruochen Zhang, Joseph Marvin Imperial, Khumaisa Nur'aini, Do Xuan Long, Musa Izzanardi Wijanarko, Joel Ruben Antony Moniz, Patrick Amadeus Irawan, Hanif Muhammad Zhafran, Isaiah Flores, Salsabila Zahirah Pranida, Jun Kevin, Jostin Jerico Rosal, Patricia Nicole Monderin, Kun Kerdthaisong, Ahmad Mustafid, My Chiffon Nguyen, Natchapon Jongwiriyanurak, Siva Worajitwannakul, Haochen Li, Adrian Xuan Wei Lim, Bin Wang, Muhammad Ravi Shulthan Habibi, Lynnette Hui Xian Ng, Mithil Bangera, Yeshil Bangera, Priyaranjan Pattnayak, Dun Li Chan, Sherissa Caren Djuniwar, Cho Chan Myei Oo, Hee Ming Shan
Computation and Language · Cross-listed from cs.LG
Low-rank Optimization Trajectories Modeling for LLM RLVR Acceleration
Zhipeng Chen, Tao Qian, Wayne Xin Zhao et al. (+1 more)
Recently, scaling reinforcement learning with verifiable rewards (RLVR) for large
language models (LLMs) has emerged as an effective training paradigm for significantly improving model capabilities, which requires guiding the model to perform extensive exploration and learning, leading to substantial computational overhead and becoming a key challenge. To reduce the number of training steps, Prior work performs linear extrapolation of model parameters. However, the dynamics of model parameter updates during RLVR training remain insufficiently understood. To further investigate the
evolution of LLMs during RLVR training, we conduct empirical experiments and find that the rank-1 subspace of the model does not evolve linearly, and its dominance over the original parameters is further amplified during LoRA training. Based on the above insights, we propose the Nonlinear Extrapolation of low-rank trajectories (NExt), a novel framework that models and extrapolates low-rank parameter trajectories in a nonlinear manner. Concretely, we first train the model using LoRA and extract the rank-1 subspace of parameter differences at multiple training steps, which is then used for the subsequent nonlinear extrapolation. Afterward, we utilized the extracted rank-1 subspace to train a predictor, which can model the trajectory of parameter updates during RLVR, and then perform the predict-extend process to extrapolate model parameters, achieving the acceleration of RLVR. To further study and understand NExt, we conduct comprehensive experiments that demonstrate the effectiveness and robustness of the method. Our method reduces computational overhead by approximately 37.5% while remaining compatible with a wide range of RLVR
algorithms and tasks. We release our code in https://github.com/RUCAIBox/NExt.
Read the full abstract ▾
All authors Zhipeng Chen, Tao Qian, Wayne Xin Zhao, Ji-Rong Wen
Computation and Language
Think Before you Write: QA-Guided Reasoning for Character Descriptions in Books
Argyrios Papoudakis, Mirella Lapata and Frank Keller
Character description generation is an important capability for narrative-focused applications such as summarization, story analysis, and character-driven simulations. However, generating accurate character descriptions from long-form narratives (e.g., novels) is challenging: models must track evolving attributes (e.g., relationships and events), integrate evidence scattered across the text, and infer implicit details. Despite the success of
reasoning-enabled LLMs on many
benchmarks, we find that for character description generation their performance improves when built-in
reasoning is disabled (i.e., an empty
reasoning trace). Motivated by this, we propose a training framework that decouples
reasoning from generation. Our approach, which can be applied on top of long-context LLMs or chunk-based methods, consists of a
reasoning model that produces a structured QA
reasoning trace and a generation model that conditions on this trace to produce the final character description. Experiments on two
datasets (BookWorm and CroSS) show that QA-guided
reasoning improves faithfulness, informativeness, and grounding over strong long-context baselines.
Read the full abstract ▾
Computation and Language
METRO: Towards Strategy Induction from Expert Dialogue Transcripts for Non-collaborative Dialogues
Haofu Yang, Jiaji Liu, Chen Huang et al. (+3 more)
Developing non-collaborative dialogue
agents traditionally requires the manual, unscalable codification of expert strategies. We propose \ours, a method that leverages large
language models to autonomously induce both strategy actions and planning logic directly from raw transcripts. METRO formalizes expert knowledge into a Strategy Forest, a hierarchical structure that captures both short-term responses (nodes) and long-term strategic foresight (branches). Experimental results across two
benchmarks show that METRO demonstrates promising performance, outperforming existing methods by an average of 9%-10%. Our further analysis not only reveals the success behind METRO (strategic behavioral diversity and foresight), but also demonstrates its robust cross-task transferability. This offers new insights into building non-collaborative
agents in a cost-effective and scalable way. Our code is available at https://github.com/Humphrey-0125/METRO.
Read the full abstract ▾
All authors Haofu Yang, Jiaji Liu, Chen Huang, Faguo Wu, Wenqiang Lei, See-Kiong Ng
Computation and Language
Bridging What the Model Thinks and How It Speaks: Expressive Speech Generation via Self-Aware Intent-Realization Alignment
Kuang Wang, Lai Wei, Ping Lin et al. (+8 more)
Speech Language Models (SLMs) exhibit strong semantic understanding, yet often fail to translate this capacity into expressive acoustic realization, producing
speech with flattened prosody and misaligned
emotion. We identify this mismatch as the semantic understanding-acoustic realization gap. Existing approaches typically rely on externally specified proxies, such as
emotion labels or style
prompts, which require annotations and struggle to capture dynamically evolving expressive intent throughout dialogue. To overcome these limitations, we propose SASLM (Self-Aware
Speech Language Model), a proxy-free framework that bridges what the model thinks and how it speaks through self-aware intent-realization
alignment: (1) Intent-Aware Bridging self-distills expressive intent from the model's own evolving semantic generation states via a Variational Information Bottleneck (VIB), thereby guiding expressive
speech realization without external expressive supervision; while (2) Realization-Aware
Alignment reflectively aligns generated acoustics with intended expression through self-reward optimization, progressively improving intent-realization consistency during
speech generation. Despite using only 3B parameters and 800 hours of expressive
speech data, SASLM achieves state-of-the-art performance on EchoMind among open-source systems, surpassing models over 10 times larger and approaching commercial systems.
Read the full abstract ▾
All authors Kuang Wang, Lai Wei, Ping Lin, Qibing Bai, Wenkai Fang, Li Zhou, Feng Jiang, Zhongjie Jiang, Jun Huang, Yannan Wang, Haizhou Li
Computation and Language
Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning
Bo Li, Mingda Wang, Gexiang Fang et al. (+2 more)
We revisit
retrieval-augmented generation (RAG) by
embedding retrieval control directly into generation. Instead of treating
retrieval as an external intervention, we express
retrieval decisions within
token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of
Retrieval as Generation, we propose GRIP (Generation-guided
Retrieval with Information Planning), a unified framework in which the model regulates
retrieval behavior through control-
token emission. Central to GRIP is Self-Triggered Information Planning, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples
retrieval and
reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific
token patterns. Experiments on five QA
benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters.
Read the full abstract ▾
All authors Bo Li, Mingda Wang, Gexiang Fang, Shikun Zhang, Wei Ye
Computation and Language · Cross-listed from cs.CV
Lost in Adaptation: Layer-Selective Recovery of Temporal Reasoning in Video-Language Models
Zihang Fu, Haonan Wang, Jian Kang et al. (+2 more)
Multimodal adaptation can weaken temporal
reasoning (TR) in
video-
language models (VLMs) even when the visual evidence needed for an answer is correctly perceived. This weakened competence, however, remains partly recoverable from the model's paired pre-adaptation
language checkpoint. We introduce MERIT, a gradient-free framework that selectively restores this competence across the
language backbone while protecting temporal perception (TP). It assigns each self-
attention layer a VLM-dominant or LLM-dominant interpolation and uses the Covariance
Matrix Adaptation
Evolution Strategy (CMA-ES) to search the recipe space under a perception-constrained objective. Across three VLM families and five
video benchmarks, MERIT improves
reasoning-aligned performance in all 15 model–
benchmark evaluations while preserving TP across all three
Video-MME settings. Recipes selected using 232 diagnostic examples transfer to four unseen
benchmarks, with relative gains up to 27.8%, and outperform full-backbone, uniform-
attention, and alpha-tuned random-layer merging. Interventional masking shows that MERIT-selected layers are disproportionately important for TR, while frame-level attribution reveals increased reliance on temporally distributed, causally relevant evidence. These results show that adaptation-induced
reasoning degradation can be repaired post hoc through selective reuse of pre-adaptation competence, without learning new parameters or adding inference overhead.
Read the full abstract ▾
All authors Zihang Fu, Haonan Wang, Jian Kang, Kenji Kawaguchi, Jiaying Wu
Computation and Language · Cross-listed from cs.CV
What Do Vision-Language Models Encode for Personalized Image Aesthetics Assessment?
Koki Ryu and Hitomi Yanaka
Personalized image aesthetics assessment (PIAA) is an important research problem with practical real-world applications. While methods based on
vision-
language models (VLMs) are promising candidates for PIAA, it remains unclear whether they internally encode rich, multi-level aesthetic attributes required for effective personalization. In this paper, we first analyze the internal representations of VLMs to examine the presence and distribution of such aesthetic attributes, and then leverage them for lightweight, individual-level personalization without model fine-tuning. Our analysis reveals that VLMs encode diverse aesthetic attributes that propagate into the
language decoder layers. Building on these representations, we demonstrate that simple linear models can perform PIAA effectively. We further analyze how aesthetic information is transferred across layers in different VLM architectures and across image domains. Our findings provide insights into how VLMs can be utilized for modeling subjective, individual aesthetic preferences. Our code is available at https://github.com/ynklab/vlm-latent-piaa.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Learning from Contrasts: Synthesizing Reasoning Paths from Diverse Search Trajectories
Peiyang Liu, Zhirui Chen, Xi Wang et al. (+4 more)
Monte Carlo Tree Search (MCTS) has been widely used for automated
reasoning data exploration, but current supervision extraction methods remain inefficient. Standard approaches retain only the single highest-reward trajectory, discarding the comparative signals present in the many explored paths. Here we introduce Contrastive
Reasoning Path Synthesis (CRPS), a framework that transforms supervision extraction from a filtering process into a synthesis procedure. CRPS uses a structured reflective process to analyze the differences between high- and low-quality search trajectories, extracting explicit information about strategic pivots and local failure modes. These insights guide the synthesis of
reasoning chains that incorporate success patterns while avoiding identified pitfalls. We show empirically that models fine-tuned on just 60K CRPS-synthesized examples match or exceed the performance of baselines trained on 590K examples derived from standard rejection sampling, a 20$\times$ reduction in
dataset size. Furthermore, CRPS improves generalization on out-of-domain
benchmarks, demonstrating that learning from the contrast between success and failure produces more transferable
reasoning capabilities than learning from success alone.
Read the full abstract ▾
All authors Peiyang Liu, Zhirui Chen, Xi Wang, Di Liang, Youru Li, Zhi Cai, Wei Ye
Computation and Language · Cross-listed from cs.CR
Geometry-Aware Localized Watermarking for Copyright Protection in Embedding-as-a-Service
Zhimin Chen, Xiaojie Liang, Wenbo Xu et al. (+2 more)
Embedding-as-a-Service (EaaS) has become an important semantic infrastructure for natural
language and multimedia applications, but it is highly vulnerable to model stealing and copyright infringement. Existing EaaS watermarking methods face a fundamental robustness–utility–verifiability tension: trigger-based methods are fragile to paraphrasing, transformation-based methods are sensitive to dimensional perturbation, and region-based methods may incur false positives due to coincidental
geometric affinity. To address this problem, we propose GeoMark, a geometry-aware localized watermarking framework for EaaS copyright protection. GeoMark uses a natural in-
manifold embedding as a shared watermark target, constructs geometry-separated anchors with explicit target–anchor margins, and activates watermark injection only within adaptive local neighborhoods. This design decouples where watermarking is triggered from what ownership is attributed to, achieving localized triggering and centralized attribution. Experiments on four
benchmark datasets show that GeoMark preserves downstream utility and
geometric fidelity while maintaining robust copyright verification under paraphrasing, dimensional perturbation, and CSE (Clustering, Selection, Elimination) attacks, with improved verification stability and low false-positive risk.
Read the full abstract ▾
All authors Zhimin Chen, Xiaojie Liang, Wenbo Xu, Yuxuan Liu, Wei Lu
Computation and Language
Do LLMs Know Tool Irrelevance? Demystifying Structural Alignment Bias in Tool Invocations
Yilong Liu, Xixun Lin, Pengfei Cao et al. (+3 more)
Large
language models (LLMs) have demonstrated impressive capabilities in utilizing external tools. In practice, however, LLMs are often exposed to tools that are irrelevant to the user's query, in which case the desired behavior is to refrain from invocations. In this work, we identify a widespread yet overlooked mechanistic flaw in tool refusal, which we term structural
alignment bias: Even when a tool fails to serve the user's goal, LLMs still tend to invoke it whenever query attributes can be validly assigned to tool parameters. To systematically study this bias, we introduce SABEval, a new
dataset that decouples structural
alignment from semantic relevance. Our analysis shows that structural
alignment bias induces severe tool-invocation errors in LLMs, yet remains largely unaccounted for in existing evaluations. To investigate the internal mechanisms underlying this bias, we propose Contrastive
Attention Attribution, which reveals two competing pathways for semantic checking and structural matching. The relative strength of these pathways drives LLMs' tool invocation decisions. Based on these findings, we further introduce a rebalancing strategy that effectively mitigates structural
alignment bias, as demonstrated by extensive experiments, without degrading general tool-use capabilities.
Read the full abstract ▾
All authors Yilong Liu, Xixun Lin, Pengfei Cao, Ge Zhang, Fang Fang, Yanan Cao
Computation and Language · Cross-listed from cs.CR
The Salami Slicing Threat: Exploiting Cumulative Risks in LLM Systems
Yihao Zhang, Kai Wang, Jiangrong Wu et al. (+7 more)
Large
Language Models (LLMs) face prominent security risks from jailbreaking, a practice that manipulates models to bypass built-in security constraints and generate unethical or unsafe content. Among various jailbreak techniques, multi-turn jailbreak attacks are more covert and persistent than single-turn counterparts, exposing critical vulnerabilities of LLMs. However, existing multi-turn jailbreak methods suffer from two fundamental limitations that affect the actual impact in real-world scenarios: (a) As models become more context-aware, any explicit harmful trigger is increasingly likely to be flagged and blocked; (b) Successful final-step triggers often require finely tuned, model-specific contexts, making such attacks highly context-dependent. To fill this gap, we propose Salami Slicing Risk, which operates by chaining numerous low-risk inputs that individually evade
alignment thresholds but cumulatively accumulate harmful intent to ultimately trigger high-risk behaviors, without heavy reliance on pre-designed contextual structures. Building on this risk, we develop Salami Attack, an automatic framework universally applicable to multiple model types and modalities. Rigorous experiments demonstrate its state-of-the-art performance across diverse models and modalities, achieving over 90% Attack Success Rate on GPT-4o and Gemini, as well as robustness against real-world
alignment defenses. We also proposed a defense strategy to constrain the Salami Attack by at least 44.8% while achieving a maximum blocking rate of 64.8% against other multi-turn jailbreak attacks. Our findings provide critical insights into the pervasive risks of multi-turn jailbreaking and offer actionable mitigation strategies to enhance LLM security.
Read the full abstract ▾
All authors Yihao Zhang, Kai Wang, Jiangrong Wu, Haolin Wu, Yuxuan Zhou, Zeming Wei, Dongxian Wu, Xun Chen, Jun Sun, Meng Sun
Computation and Language
Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning
Rui Song, Lida Shi, Ruihua Qi et al. (+2 more)
In recent years, rapid advances in
Multimodal Large
Language Models (MLLMs) have increasingly stimulated research on ancient Chinese scripts. As the
evolution of written characters constitutes a fundamental pathway for understanding cultural transformation and historical continuity, how MLLMs can be systematically leveraged to support and advance text
evolution analysis remains an open and largely underexplored problem. To bridge this gap, we construct a comprehensive
benchmark comprising 11 tasks and over 130,000 instances, specifically designed to evaluate the capability of MLLMs in analyzing the
evolution of ancient Chinese scripts. We conduct extensive evaluations across multiple widely used MLLMs and observe that, while existing models demonstrate a limited ability in glyph-level comparison, their performance on core tasks-such as character recognition and evolutionary
reasoning-remains substantially constrained. Motivated by these findings, we propose a glyph-driven fine-tuning framework (GEVO) that explicitly encourages models to capture evolutionary consistency in glyph transformations and enhances their understanding of text
evolution. Experimental results show that even models at the 2B scale achieve consistent and comprehensive performance improvements across all evaluated tasks. To facilitate future research, we publicly release both the
benchmark and the trained models\footnote{https://github.com/songruiecho/GEVO}.
Read the full abstract ▾
All authors Rui Song, Lida Shi, Ruihua Qi, Yingji Li, Hao Xu