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.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
One LLM Does Not Simulate All Students: Ability-Aware Student Simulation via Cognitive Diagnosis Guided LLM Assignment
PDF ↗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.
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.
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.
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.
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.
Entity alignment (EA) aims to identify entities referring to the same real-world object across different knowledge graphs (KGs). Recent approaches based on large language models (LLMs) typically obtain entity embeddings through knowledge representation learning and use embedding similarity to identify an alignment-uncertain entity set. For each uncertain entity, a candidate entity set (CES) is then retrieved based on embedding similarity to support subsequent alignment reasoning and decision making. However, the reliability of the CES and the reasoning capability of LLMs critically affect the effectiveness of subsequent alignment decisions. To address this issue, we propose AgentEA, a reliable EA framework based on multi-agent debate. AgentEA first improves embedding quality through entity representation preference optimization, and then introduces a two-stage multi-role debate mechanism consisting of lightweight debate verification and deep debate alignment to progressively enhance the reliability of alignment decisions while enabling more efficient debate-based reasoning. Extensive experiments on public benchmarks under cross-lingual, sparse, large-scale, and heterogeneous settings demonstrate the effectiveness of AgentEA.
The autonomous synthesis of deep research reports represents a critical frontier for Large Language Models (LLMs), demanding sophisticated information orchestration and non-linear narrative logic. Current approaches rely on rigid predefined linear workflows, which cause error accumulation, preclude global restructuring from subsequent insights, and ultimately limit in-depth multimodal fusion and report quality. We propose CogGen, a Cognitively inspired recursive framework for deep research report Generation. Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning and global restructuring. To extend this recursivity to multimodal content, we introduce Abstract Visual Representation (AVR): a concise intent-driven language that iteratively refines visual-text layouts without pixel-level regeneration overhead. We further present CLEF, a Cognitive Load Evaluation Framework, and curate a new benchmark from Our World in Data (OWID). Extensive experiments show CogGen achieves state-of-the-art results among open-source systems, generating reports comparable to professional analysts’ outputs and surpassing Gemini Deep Research. Our code and dataset will be publicly available upon publication.
Real-time, continuous understanding of visual signals is essential for real-world interactive AI applications, and poses a fundamental system-level challenge. Existing research on streaming video understanding, however, typically focuses on isolated aspects such as question-answering accuracy under limited visual context or improvements in encoding efficiency, while largely overlooking practical deployability under realistic resource constraints. To bridge this gap, we introduce StreamingEval, a unified evaluation framework for assessing the streaming video understanding capabilities of Video-LLMs under realistic constraints. StreamingEval benchmarks both mainstream offline models and recent online video models under a standardized protocol, explicitly characterizing the trade-off between efficiency, storage and accuracy. Specifically, we adopt a fixed-capacity memory bank to normalize accessible historical visual context, and jointly evaluate visual encoding efficiency, text decoding latency, and task performance to quantify overall system deployability. Extensive experiments across multiple datasets reveal substantial gaps between current Video-LLMs and the requirements of realistic streaming applications, providing a systematic basis for future research in this direction. Codes will be released upon acceptance.
Role-playing prompts effectively steer Large Language Models (LLMs), yet the neural mechanism driving this behavioral shift remains unclear. In this work, we identify Role-Sensitive Neurons (RSNs)—a sparse sub-network (≈ 0.5% of all neurons) governing the transition from hesitation to action. Using a novel evaluation framework with explicit abstention (MMLU-E), we reveal a Confidence-Performance Decoupling: roles primarily modulate the model’s probabilistic "willingness to act" rather than its underlying knowledge representation. We demonstrate that RSNs function as a mechanistic gain control system: causal intervention on this subspace allows precise regulation of abstention behavior. Furthermore, cross-model transfer experiments confirm that these circuits are indigenous to pre-training, with Instruction Tuning (SFT) acting merely as a "signal sharpener" to refine latent gain dynamics. Finally, we identify a critical safety boundary: in knowledge-deficient models, amplifying RSNs induces "unwarranted certainty," highlighting decisiveness as a tunable gain parameter distinct from epistemic truth.
Multimodal Large Language Models (MLLMs) excel at structural reasoning yet suffer from a sharp logical brittleness in structural consistency. We term this phenomenon Structural Cognitive Overload (SCO), a byproduct of the contention between deep reasoning and safety alignment. However, prior work has predominantly targeted typographic and pixel-level perturbations, leaving the study of SCO largely unexplored. To this end, we propose StructBreak, an automated end-to-end framework designed to quantify SCO. By leveraging StructBreak, we uncover a novel higher-order cognitive overload attack paradigm; notably, this attack operates under a practical black-box setting, requiring no internal model access. Consequently, we utilize this framework to establish a comprehensive benchmark spanning ten diverse threat scenarios. Empirical evaluations on six leading MLLMs reveal that SCO readily triggers toxic generation, yielding a 92% average ASR (up to 97% on Gemini 2.5). To elucidate the mechanism of SCO, we further conduct model-level interpretations spanning attention dynamics, latent space topology, and geometric analysis. Our findings reveal that StructBreak acts as a novel structural channel to circumvent safety filters. Furthermore, the limited efficacy of inherent safety mechanisms underscores that current alignment paradigms are insufficient for the era of complex multimodal reasoning.
Hard-gated safety checkers often over-refuse and misalign with a vendor’s model spec; prevailing taxonomies also neglect robustness and honesty, yielding safer-on-paper yet less useful systems. This work introduces Guardian-as-an-Advisor (GaaA), a soft-gating pipeline where a guardian predicts a binary risk label plus a concise explanation and prepends this advice to the original query for re-inference, keeping the base model operating under its original spec. To support training and evaluation, GuardSet is constructed—a 208k+ multi-domain dataset unifying harmful and harmless cases with targeted robustness and honesty slices. GuardAdvisor is trained via SFT followed by RL to enforce label–explanation consistency. GuardAdvisor attains competitive detection accuracy while enabling the advisory workflow; when used to augment inputs, responses improve over unaugmented prompts. A latency study shows advisor inference uses below 5% of base-model compute and adds only 2–10% end-to-end overhead under realistic harmful-input rates. Overall, GaaA steers models to comply with the model spec, maintaining safety while reducing over-refusal.
Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated remarkable capabilities.However, their deployment is hindered by significant computational costs. Existing structured pruning methods, while hardware-efficient, often suffer from significant accuracy degradation. In this paper, we argue that this failure stems from a stage-agnostic pruning approach that overlooks the asymmetric roles between the prefill and decode stages. By introducing a virtual gate mechanism, our importance analysis reveals that deep layers are critical for next-token prediction (decode) but largely redundant for context encoding (prefill). Leveraging this insight, we propose Prefill-Only Pruning (POP), a stage-aware inference strategy that safely omits deep layers during the computationally intensive prefill stage while retaining the full model for the sensitive decode stage. To enable the transition between stages, we introduce independent Key-Value (KV) projections to maintain cache integrity, and a boundary handling strategy to ensure the accuracy of the first generated token. Extensive experiments on Llama-3.1, Qwen3-VL, and Gemma-3 across diverse modalities demonstrate that POP achieves up to 1.37\times speedup in prefill latency with minimal performance loss, effectively overcoming the accuracy-efficiency trade-off limitations of existing structured pruning methods.
Recent advances in Retrieval-Augmented Generation (RAG) have shifted from simple vector similarity to structure-aware approaches like HippoRAG, which leverage Knowledge Graphs (KGs) and Personalized PageRank (PPR) to capture multi-hop dependencies. However, these methods suffer from a "Static Graph Fallacy": fixed transition probabilities set during indexing ignore query-dependent edgerelevance, causing semantic drift where random walks are diverted into high-degree "hub" nodes before reaching critical evidence. Models often achieve high partial recall but fail to retrieve the complete evidence chain for multi-hop queries. To address this, we propose CatRAG, Context-Aware Traversal for robust RAG, which builds on the HippoRAG 2 and transforms the static KG into a query-adaptive navigation structure. CatRAG steers the random walk via three mechanisms: (1) Symbolic Anchoring, injecting weak entity constraints to regularize the random walk; (2) QueryAware Dynamic Edge Weighting, dynamically modulating graph structure to prune irrelevant paths and amplify query-aligned ones; and (3) Key-Fact Passage Weight Enhancement, a cost-efficient bias anchoring the walk to key evidence. Experiments across multi-hop benchmarks show that CatRAG outperforms state-of-the-art baselines. While standard Recall gains are modest, CatRAG achieves substantial improvements in reasoning completeness—the capacity to recover entire evidence chains without gaps. These results reveal that CatRAG effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning. Resources are available at https://github.com/kwunhang/CatRAG.
ARISE: An Adaptive Resolution-Aware Metric for Test-Time Scaling Evaluation in Large Reasoning Models
PDF ↗Test-time scaling has emerged as a transformative paradigm for enhancing the performance of large reasoning models, enabling dynamic allocation of computational resources during inference. However, as the landscape of reasoning models rapidly expands, a critical question remains: how can we systematically compare and evaluate the test-time scaling capabilities across different models? In this paper, we introduce ARISE (Adaptive Resolution-aware Scaling Evaluation), a novel metric specifically designed to assess the test-time scaling effectiveness of large reasoning models. Unlike existing evaluation approaches, ARISE incorporates two key innovations: (1) sample-level awareness that effectively penalizes negative scaling behaviors where increased computation leads to performance degradation, and (2) a dynamic sampling mechanism that mitigates the impact of accuracy fluctuations and token count instability on the final assessment. We conduct comprehensive experiments evaluating state-of-the-art reasoning models across diverse domains including mathematical reasoning, code generation, and agentic tasks. Our results demonstrate that ARISE provides a reliable and fine-grained measurement of test-time scaling capabilities, revealing significant variations in scaling efficiency across models. Notably, our evaluation identifies Claude Opus as exhibiting superior scaling characteristics compared to other contemporary reasoning models.
CodeWiki: Evaluating AI’s Ability to Generate Holistic Documentation for Large-Scale Codebases
PDF ↗Comprehensive software documentation is crucial yet costly to produce. Despite recent advances in large language models (LLMs), generating holistic, architecture-aware documentation at the repository level remains challenging due to complex and evolving codebases that exceed LLM context limits. Existing automated methods struggle to capture rich semantic dependencies and architectural structure. We present \textbf{CodeWiki}, a unified framework for automated repository-level documentation across seven mainstream programming languages. CodeWiki combines top-down hierarchical decomposition with a divide-and-conquer agent system to preserve architectural context and scale documentation generation, and a bottom-up synthesis that integrates textual descriptions with visual artifacts such as architecture and data-flow diagrams. We also introduce \textbf{CodeWikiBench}, a benchmark with hierarchical rubrics and LLM-based evaluation protocols. Experiments show that CodeWiki achieves a 68.79% quality score with proprietary models, outperforming the closed-source DeepWiki baseline by 4.73%, with especially strong gains on scripting languages. CodeWiki is released as open source to support future research.
Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at https://github.com/yeasen-z/LeakDojo.
Progress Ratio Embeddings: An Impatience Signal for Robust Length Control in Neural Text Generation
PDF ↗Modern neural language models achieve high accuracy in text generation, yet precise control over generation length remains underdeveloped. In this paper, we first investigate a recent length control method based on Reverse Positional Embeddings (RPE) and show its limits when control is requested beyond the training distribution. In particular, using a discrete countdown signal tied to the absolute remaining token count leads to instability. To provide robust length control, we introduce Progress Ratio Embeddings (PRE), as continuous embeddings tied to a trigonometric impatience signal. PRE integrates seamlessly into standard Transformer architectures, providing stable length fidelity without degrading text accuracy under standard evaluation metrics. We further show that PRE generalizes well to unseen target lengths. Experiments on two widely used news-summarization benchmarks and a popular question generation dataset validate these findings.
With the rapid advancement of speech generation technologies, the threat posed by speech deepfakes in real-time communication (RTC) scenarios has intensified. However, existing detection studies mainly focus on offline simulations and struggle to cope with the complex distortions introduced during RTC transmission, including unknown speech enhancement processes (e.g., noise suppression) and codec compression. To address this challenge, we present the first large-scale speech deepfake dataset tailored for RTC scenarios, termed RTCFake, totaling approximately 600 hours. The dataset is constructed by transmitting speech through multiple mainstream social media and conferencing platforms (e.g., Zoom), enabling precise pairing between offline and online speech. In addition, we propose a phoneme-guided consistency learning (PCL) strategy that enforces models to learn platform-invariant semantic structural representations. In this paper, the RTCFake dataset is divided into training, development, and evaluation sets. The evaluation set further includes both unseen RTC platforms and unseen complex noise conditions, thereby providing a more realistic and challenging evaluation benchmark for speech deepfake detection. Furthermore, the proposed PCL strategy achieves significant improvements in both cross-platform generalization and noise robustness, offering an effective and generalizable modeling paradigm.
Federated LoRA Fine-Tuning with Pipelined Error-Mitigated Aggregation and Matrix-Wise Freezing
PDF ↗Federated low-rank adaptation (LoRA) enables multiple clients to collaboratively fine-tune large language models (LLMs) without disclosing their raw data. However, existing works often experience performance degradation due to biased model aggregation and are hindered by significant communication and computation burden, both limiting training efficiency. In this paper, we propose iFLoRA, an improved Federated LoRA fine-tuning system for LLMs featuring pipelined error-mitigated model aggregation and adaptive matrix-wise parameter freezing. Specifically, iFLoRA mitigates aggregation error by first reconstructing local update matrices from clients’ low-rank matrices. These are then aggregated into a global update, which is decomposed via singular value decomposition (SVD) to form low-rank matrices for the next round. To mitigate the overhead from SVD, iFLoRA employs a pipeline to overlap global aggregation, local computation, and communication. Additionally, iFLoRA implements an adaptive matrix-wise freezing scheme that assesses their stability and selectively freezes them for adaptively adjusted periods, alleviating client training overheads without compromising model performance. Extensive experiments on real-world datasets show that iFLoRA can improve time-to-target by 2.17-8.48× than state-of-the-art methods. Our code is available at: https://github.com/whr819987540/iflora.