论文检索

Fei Wu, Zhenrong Zhang, Qikai Chang, Jianshu Zhang, Quan Liu, Jun Du

Reinforcement Learning with Verifiable Rewards (RLVR) elicits long chain-of-thought reasoning in large language models (LLMs), but outcome-based rewards lead to coarse-grained advantage estimation. While existing approaches improve RLVR via token-level entropy or sequence-level length control, they lack a semantically grounded, step-level measure of reasoning progress. As a result, LLMs fail to distinguish necessary deduction from redundant verification: they may continue checking after reaching a correct solution and, in extreme cases, overturn a correct trajectory into an incorrect final answer. To remedy the lack of process supervision, we introduce a training-free probing mechanism that extracts intermediate confidence and correctness and combines them into a Step Potential signal that explicitly estimates the reasoning state at each step. Building on this signal, we propose Step Potential Advantage Estimation (SPAE), a fine-grained credit assignment method that amplifies potential gains, penalizes potential drops, and applies penalty after potential saturates to encourage timely termination. Experiments across multiple benchmarks show SPAE consistently improves accuracy while substantially reducing response length, outperforming strong RL baselines and recent efficient reasoning and token-level advantage estimation methods. The code is available at https://github.com/cii030/SPAE-RL.

Sander Land, Yuval Pinter

The Unigram tokenization algorithm offers a probabilistic alternative to the greedy heuristics of Byte-Pair Encoding. Despite its theoretical elegance, its implementation in practice is complex, limiting its adoption to the SentencePiece package and adapters thereof. We bridge this gap between theory and practice by providing a clear guide to implementation and parameter choices. We also identify a simpler algorithm that accepts slightly higher training loss in exchange for improved compression.

Xiyin Zeng, Yi Lu, Hao Wang

Visual Question Answering (VQA) requires models to identify the correct answer options based on both visual and textual evidence. Recent Mixture-of-Experts (MoE) methods improve option reasoning by grouping similar concepts or routing based on examples. However, unstable routing can lead to inconsistent expert selection in the same question type, while overly stable routing may reduce flexibility. To address this, we propose Concept-Guided Routing framework (CoGR-MoE), which incorporates semantics of the answer options to guide expert selection in the training phase.Next, option features are used to reweight the selected experts, producing discriminative representations for each candidate option. These option-level representations are further used for option comparison and optimized via contrastive learning. The experimental results indicate that CoGR-MoE delivers strong performance across multiple VQA tasks, demonstrating the effectiveness of our approach.

Bowen Zheng, Ming Ma, Zhongqiao Lin, Tianming Yang

Large Language Models (LLMs) have demonstrated remarkable abilities, one of the most important being in-context learning (ICL). With ICL, LLMs can derive the underlying rule from a few demonstrations and provide answers that comply with the rule. Previous work hypothesized that the network creates a task vector in specific positions during ICL. The task vector can be computed by averaging across the dataset. It conveys the overall task information and can thus be considered global. Patching the global task vector allows LLMs to achieve zero-shot performance with dummy inputs comparable to few-shot learning. However, we find that such a global task vector does not exist in all tasks, especially in tasks that rely on rules that can only be inferred from multiple demonstrations, such as categorization tasks. Instead, the information provided by each demonstration is first transmitted to its answer position and forms a local task vector associated with the demonstration. In some tasks but not in categorization tasks, all demonstrations’ local task vectors converge in later layers, forming the global task vector. We further show that local task vectors encode a high-level abstraction of rules extracted from the demonstrations. Our study provides novel insights into the mechanism underlying ICL in LLMs, demonstrating how ICL may be achieved through an information aggregation mechanism.

Xingwu Chen, Zhanqiu Zhang, Steven Y. Guo, Difan Zou

While LLMs demonstrate strong reasoning capabilities when provided with full information in a single turn, they exhibit substantial vulnerability in multi-turn interactions. Specifically, when information is revealed incrementally or requires updates, models frequently fail to integrate new constraints, leading to a collapse in performance compared to their single-turn baselines. We term the root cause as Contextual Inertia: a phenomenon where models rigidly adhere to previous reasoning traces. Even when users explicitly provide corrections or new data in later turns, the model ignores them, preferring to maintain consistency with its previous (incorrect) reasoning path. To address this, we introduce Reinforcement Learning with Single-Turn Anchors (RLSTA), a generalizable training approach designed to stabilize multi-turn interaction across diverse scenarios and domains. RLSTA leverages the model’s superior single-turn capabilities as stable internal anchors to provide reward signals. By aligning multi-turn responses with these anchors, RLSTA empowers models to break contextual inertia and self-calibrate their reasoning based on the latest information. Experiments show that RLSTA significantly outperforms standard fine-tuning and abstention-based methods. Notably, our method exhibits strong cross-domain generalization (e.g., math to code) and proves effective even without external verifiers, highlighting its potential for general-domain applications.

Guijin Luo, Zequn Xie, Sihang Cai, Chuxin Wang, Zhou Zhao, Yixuan Tang

Person Re-Identification (ReID) has long struggled with the semantic gap between low-level visual features and high-level identity concepts. While Vision-Language Models (VLMs) offer promising semantic understanding, existing methods typically adopt a static "one-pass" paradigm, converting images to text once for retrieval. This approach suffers from two critical flaws: Information Bottleneck, where converting rich visuals into text causes detail loss, and Open-Loop Failure, where initial hallucinations propagate without recourse. To address this, we propose Auto-ReID, a novel framework that reformulates ReID as an iterative "Think-and-Refine" process. We first introduce a Hierarchical Progressive Tuning strategy to transform a generic VLM into a specialized Re-ID expert. During inference, we deploy a closed-loop architecture comprising a Reasoner for structured attribute extraction, a Hybrid Retriever that anchors dynamic semantic queries with stable visual features to prevent drift, and a Corrector that deconstructs and verifies candidates to iteratively optimize the search. Extensive experiments on ReID datasets demonstrate that our method significantly outperforms state-of-the-art approaches, particularly in complex occlusion scenarios.

Zimu Jia, Mingjie Xu, Andrew Estornell, Jiaheng Wei

The efficacy of Large Vision-Language Models (LVLMs) is critically dependent on the quality of their training data, requiring a precise balance between visual fidelity and instruction-following capability. Existing datasets, however, are plagued by inconsistent quality, and current data filtering methods rely on coarse-grained scores that lack the granularity to identify nuanced semantic flaws like logical fallacies or factual errors. This creates a fundamental bottleneck in developing more reliable models. To address this, we make three core contributions. First, we construct a large-scale, 300K-sample benchmark by systematically injecting diverse, subtle defects to provide a challenging testbed for data auditing. Second, we introduce a novel "Decomposition-then-Evaluation" paradigm that breaks model responses into constituent cognitive components: visual description, subjective inference, and factual claim, enabling targeted analysis. Third, we instantiate this paradigm via EVIAN (Explainable Visual Instruction-tuning Data AuditiNg), an automated framework that evaluates these components along the orthogonal axes of Image-Text Consistency, Logical Coherence, and Factual Accuracy. Our empirical findings challenge the prevailing scale-centric paradigm: a model fine-tuned on a compact, high-quality subset curated by EVIAN consistently surpassed models trained on orders-of-magnitude larger datasets. We also reveal that dividing complex auditing into verifiable subtasks enables robust curation, and that Logical Coherence is the most critical factor in data quality evaluation.

Hao Liao, Jiwei Zhang, Jianxun Lian, Wensheng Lu, Mingqi Wu, Shuowangg, Yong Zhang, Yitian Huang, Mingyang Zhou, Rui Mao

Recommender systems based on Large Language Models (LLMs) are often plagued by hallucinations of out-of-domain (OOD) items. To address this, we propose RecLM, a unified framework that bridges the gap between retrieval and generation by instantiating three grounding paradigms under a single architecture: embedding-based retrieval, constrained generation over rewritten item titles, and discrete item-tokenizer generation. Using the same backbone LLM and prompts, we systematically compare these three views on public benchmarks. RecLM strictly eradicates OOD recommendations (OOD@10 = 0) across all variants, and the constrained generation variants RecLM-cgen and RecLM-token achieve overall state-of-the-art accuracy compared to both strong ID-based and LLM-based baselines. Our unified view provides a systematic basis for comparing three distinct paradigms to reduce item hallucinations, offering a practical framework to facilitate the application of LLMs to recommendation tasks. Source code is at https://github.com/microsoft/RecAI.

Xiaochuan Liu, Yuanfeng Song, Xiaoming Yin, Xing Chen

In today’s data-driven era, fully automated end-to-end data analytics, particularly insight discovery, is critical for discovering actionable insights that assist organizations in making effective decisions. With the rapid advancement of large language models (LLMs), LLM-driven agents have emerged as a promising paradigm for automating insight discovery. However, existing data insight agents remain limited in several key aspects, often failing to deliver satisfactory results due to: (1) insufficient utilization of domain knowledge, (2) shallow analytical depth, and (3) error-prone code generation. To address these issues, we propose DataSage, a novel multi-agent framework that incorporates three innovative features including external knowledge retrieval to enrich the analytical context, a multi-role debating mechanism to simulate diverse analytical perspectives and deepen analytical depth, and multi-path reasoning to improve the accuracy of the generated code and insights. Extensive experiments on InsightBench demonstrate that DataSage consistently outperforms existing data insight agents across all difficulty levels, improving by 7.5% and 13.9% respectively in the insight-level and summary-level metrics. It offers an effective solution for automated data insight discovery.

Gaspard Michel, Elena V. Epure, Christophe Cerisara

Recent advances in text-to-speech (TTS) have been driven by large, multi-domain speech corpora, yet the expressive potential of audiobook data remains underexamined. We argue that human-narrated audiobooks, particularly fictional works, contain rich and diverse prosodic cues arising from the natural alternation between neutral narration and expressive character dialogue. Building from this observation, we introduce LibriQuote, a large-scale 5.3K hours of expressive speech drawn from character quotations.Each quote is supplemented with contextual pseudo-labels for speech verbs and adverbs that characterize the intended delivery of direct speech (e.g., “he whispered softly”).We found that fine-tuning a flow-matching model on LibriQuote yields substantial improvements in expressivity and intelligibility, while training from scratch enhances expressiveness of an autoregressive TTS model.Benchmarking on LibriQuote-test highlights significant variability across systems in generating expressive speech.We publicly release the dataset, code, and evaluation resources to facilitate reproducibility.Audio samples can be found at https://libriquote.github.io/.

Yao Chen, Yilong Chen, Yinqi Yang, Junyuan Shang, Zhenyu Zhang, Zefeng Zhang, Shuaiyi Nie, Shuohuan Wang, Yu Sun, Hua Wu 等

Existing approaches to increasing the effective depth of Transformers predominantly rely on parameter reuse, extending computation through recursive execution.Under this paradigm, the network structure remains static along the training timeline, and additional computational depth is uniformly assigned to entire blocks at the parameter level.This rigidity across training time and parameter space leads to substantial computational redundancy during training.In contrast, we argue that depth allocation during training should not be a static preset, but rather a progressively growing structural process. Our systematic analysis reveals a deep-to-shallow maturation trajectory across layers, where high-entropy attention heads play a crucial role in semantic integration. Motivated by this observation, we introduce the Sparse Growing Transformer (SGT).SGT is a training-time sparse depth allocation framework that progressively extends recurrence from deeper to shallower layers via targeted attention looping on informative heads. This mechanism induces structural sparsity by selectively increasing depth only for a small subset of parameters as training evolves.Extensive experiments across multiple parameter scales demonstrate that SGT consistently outperforms training-time static block-level looping baselines under comparable settings, while reducing the additional training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone.

Yinghao Chen, Wantong Xie, Shuli Zeng, Sijia Zhang, Xiaotian Pan, Feng Wu, Xiangyang Li

Large language models (LLMs) and vision-language models (VLMs) are increasingly used as optimization assistants to produce solutions, generate solver-executable programs, or both. However, current evaluations are misaligned with deployment in three ways: they (P1) fail to represent multimodal problem specifications, (P2) score outcomes only and cannot localize where failures occur along the modeling pipeline, and (P3) rarely report inference cost, obscuring reliability–cost trade-offs. We introduce Graph Optimization benchmark (GOBench), an aligned multimodal benchmark with solver-derived oracles and a four-layer diagnostic protocol that evaluates intermediate artifacts as well as end results, together with the Visual Inference Penalty (VIP) to measure multimodal overhead. Across frontier and open-weight models under paired text-only vs. T+V settings, we find that vision reliably increases inference cost, while its reliability impact is regime-dependent: frontier models often benefit from visual grounding, whereas several mid-tier/open models exhibit a Visual Paradox where vision reduces downstream executability and verification coverage. End-to-end success is frequently bottlenecked by intermediate-stage dropout; supervised fine-tuning on intermediate targets can mitigate this attrition in open models, enabling a reproducible harness for diagnosing failure modes and quantifying reliability–cost trade-offs.

Peiyang Liu, Yining Wang, Youru Li, Long Li, Zhi Cai, Wei Ye

While Chain-of-Thought (CoT) reasoning enhances code generation in Large Language Models (LLMs), it introduces a critical challenge in uncertainty estimation: Confidence Saturation. Existing calibration methods, such as Self-Consistency, rely on the assumption that consensus implies correctness. This assumption fails under systematic errors, where models confidently repeat flawed logic, leading to miscalibrated high-confidence predictions. To address this, we introduce NeuroSym-Cal, a hierarchical calibration framework. We posit that reliable confidence requires interrogating the model at two complementary levels: the extrinsic consensus of its symbolic outputs and the intrinsic sensitivity of its latent reasoning. Specifically, we propose Reasoning Sensitivity Analysis to measure the local curvature of the deductive process via latent perturbation, providing a fine-grained signal that persists even when output consensus saturates. These orthogonal features are fused by a Contextual Calibration Network to predict correctness. Experiments across state-of-the-art reasoning models (e.g., DeepSeek-R1) demonstrate that NeuroSym-Cal effectively de-saturates overconfident errors, achieving state-of-the-art Expected Calibration Error (ECE) and superior selective generation performance on Out-Of-Domain (OOD) benchmarks.

Yu-Xiang Lin, Cheng-Han Chiang, Hung-yi Lee

In this paper, we show that when spoken language models (SLMs) are instructed to speak in a specific speaking style at the beginning of a multi-turn conversation, they cannot maintain the required speaking styles after several turns of interaction; we refer to this as the style amnesia of SLMs. We focus on paralinguistic speaking styles, including emotion, accent, volume, and speaking speed. We evaluate three proprietary and two open-source SLMs, demonstrating that none of these models can maintain a consistent speaking style when instructed to do so. We further show that while SLMs can recall the style instruction when prompted in later turns, they still fail to express it, but through explicit recall can mitigate style amnesia. In addition, SLMs struggle more when the style instruction is placed in system messages rather than user messages, even though system messages are specifically designed to provide persistent, conversation-level instructions. Our findings highlight a systematic gap in current SLMs’ ability to maintain speaking styles, highlighting the need for improved style adherence in future models. Our code and evaluation data are publicly available at https://github.com/YuXiangLin1234/SLM-Style-Amnesia.

Hehai Lin, Shilei Cao, Sudong Wang, Haotian Wu, Minzhi Li, Linyi Yang, Juepeng Zheng, Chengwei Qin

Existing multi-agent learning approaches explicitly foster collaboration among Large Language Models (LLMs) to build stronger multi-agent systems (MAS), yet they still rely on re-executing the MAS during inference. This contrasts with human cognition, wherein individuals can internalize insights from interactions to improve later independent reasoning. To investigate whether multi-agent interaction can enhance LLMs’ independent problem-solving ability, we propose ILR (Interactive Learning for LLM Reasoning), a co-learning framework that integrates Dynamic Interaction and Perception Calibration. Dynamic Interaction adaptively selects cooperative or competitive strategies based on question difficulty and model capability, after which LLMs exchange information via Idea3 framework (Idea Sharing, Idea Analysis, and Idea Fusion), an interaction paradigm simulating human discussion, before producing final answers. Perception Calibration employs Group Relative Policy Optimization (GRPO) while integrating one LLM’s reward characteristics into another’s to strengthen interaction cohesion. We evaluate the effectiveness of ILR across three LLMs from two model families of varying scales on five mathematical and one coding benchmarks. We further investigate the advantage of Dynamic Interaction (i.e., boosting the robustness of stronger LLMs and surpassing pure strategy), and the scalability of ILR beyond two-model interactions.

Huixing Que, Qi Liu, Weibo Gao, Zhenya Huang

Large Language Models (LLMs) have become integral to personalized education systems, particularly in the realm of student behavior simulation. By predicting fine-grained learning behaviors, these simulations enable intelligent systems to provide tailored instructional support. However, most existing methods rely on a single high-capacity LLM to represent an entire population of diverse learners. In this work, we demonstrate that this “one-size-fits-all” approach induces a systematic ability-dependent bias, where high-capacity models tend to overestimate low-ability students while lower-capacity models underestimate high-ability ones. To mitigate this distortion, we propose an **ability-aware student simulation framework** that dynamically matches students with appropriate LLM backbones through cognitive alignment. We leverage Neural Cognitive Diagnosis (NeuralCD) to extract multidimensional cognitive profiles for both human students and LLM agents within a shared skill space, subsequently pairing each student with the most cognitively representative model. Extensive experiments demonstrate that our approach substantially reduces simulation bias and consistently outperforms single-model baselines across the entire proficiency spectrum. Our findings suggest that faithful behavior simulation necessitates the **alignment of model capacity with student ability**, establishing cognitive diagnosis as a principled mechanism for model assignment in educational AI.

Jiahuan Cao, Yongxin Shi, Zeyu Shan, Zhengyang Lu, Lianwen Jin

Chinese historical documents encode millennia of cultural heritage, yet remain largely inaccessible to computational analysis. While multimodal large language models (MLLMs) have achieved strong performance on modern document OCR, their application to historical Chinese texts suffers from severe hallucinations, character fabrication, uncontrolled repetition, and semantic drift. We identify the root cause as visual-textual misalignment: models prioritize linguistic priors over visual evidence, particularly problematic when archaic orthography and degraded image quality destabilize cross-modal correspondences. To address this, we propose HisDoc-OCR, which restores visual grounding through three synergistic strategies: (1) Layout Injection, which encodes two-dimensional layout structures into textual outputs using layout-aware delimiters; (2) First-Occurrence Boost, which emphasizes vision-dependent characters during training by reweighting first-occurrence characters; (3) Self-Distilled Attention Focusing, which guides the model’s attention by distilling patterns from the most focused layer to the remaining layers. Extensive experiments demonstrate that HisDoc-OCR consistently outperforms general-purpose and OCR-specific MLLMs. The code will be publicly available.

Wenjie Yang, Mao Zheng, Mingyang Song, Zheng Li, Sitong Wang

Large language models (LLMs) have recently demonstrated remarkable capabilities in machine translation (MT). However, most advanced MT-specific LLMs rely heavily on external supervision during training, such as human-annotated reference data or trained reward models (RMs), which are expensive to obtain and difficult to scale. To address this limitation, we propose **Simple Self-Rewarding (SSR)**, a reinforcement learning (RL) framework for MT that is reference-free and relies solely on self-judging rewards. Using only 13K monolingual examples and Qwen-2.5-7B as the backbone, SSR-Zero-7B outperforms existing MT-specific LLMs as well as larger general LLMs such as Qwen2.5-32B-Instruct on English \leftrightarrow Chinese translation benchmarks including WMT23, WMT24, and FLORES200. It further demonstrates strong generalization to low-resource language pairs. In addition, when augmented with external supervision from COMET, our strongest model, SSR-X-Zero-7B, surpasses all existing open-source models under 72B parameters and performs competitively with leading closed-source systems in English \leftrightarrow Chinese translation. Our analysis highlights the effectiveness and generalizability of the self-rewarding mechanism relative to external LLM-as-a-judge approaches and demonstrates its complementary benefits when combined with trained RMs. We will publicly release our code, data, and models.

Van Yang, Shouren Wang, Debargha Ganguly, Xinpeng Li, Chaoda Song, Vikash Singh, Vipin Chaudhary, Xiaotian Han

Reasoning language models are controlled through explicit modes such as Think and No-think, yet we find that these behaviors are largely governed by a few token-level triggers rather than high-level instructions. Through attention analysis and controlled prompting experiments, we show that a leading “Okay” token induces reasoning behavior, while the newline pattern following ‘</think>‘ suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy–length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Ye Wang, Ruijun Jiang, Zhongqing Wang, Guodong Zhou

Aspect-based sentiment analysis has garnered increasing attention in the research community; however, most studies have predominantly focused on English datasets, with other languages such as Chinese, Japanese, and German being neglected due to the limited availability of adequately labeled data. Even within English, labeled data is scarce. To address these challenges, this study investigates the utilization of a multilingual pre-trained setting to leverage resources from diverse languages for aspect-based sentiment analysis. Specifically, we propose a Cross-lingual Knowledge Fusion framework that explores various single-round and two-round bilingual pre-training configurations. This framework utilizes both the original and translated texts, along with their corresponding labels, to pre-train the multilingual model. Evaluation results reveal that our model significantly outperforms state-of-the-art performance across multiple languages, highlighting the effectiveness of the proposed multilingual pre-trained language model for aspect-based sentiment analysis.