While large language models exhibit strong reasoning capabilities, prior work shows that their performance can be further enhanced by encouraging greater exploration. However, existing approaches overlook the presence of unhealthy exploration that increases exploration-related token usage without contributing to effective problem-solving. In this work, we show that prompt ambiguity can artificially prolong early-stage exploration, manifested as an elevated and delayed early-stage entropy peak. Although this uncertainty may be gradually resolved as reasoning progresses, reflected in the eventual convergence of the late-stage entropy peak, it does not meaningfully improve accuracy or self-consistency and instead substantially reduces reasoning efficiency. Motivated by these observations, we propose an entropy-dynamics-aware prompt optimization framework that trains a lightweight optimizer to generate concise clarifications. These clarifications aim to reduce ambiguity-induced early-stage uncertainty while preserving the model’s reasoning capabilities. Extensive experiments across multiple models, reasoning budgets, and benchmarks demonstrate that our approach consistently improves reasoning efficiency by up to 52%, by reducing unhealthy exploration without sacrificing accuracy.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Large Language Models (LLMs) encode vast factual knowledge, yet their inability to selectively forget specific information hinders privacy protection, bias mitigation, and post-deployment correction. We present LoRA-based Unlearning with Negative Examples (LUNE), a lightweight framework that performs negative-only unlearning by updating only low-rank adapters while freezing the backbone, thereby localizing edits and avoiding disruptive global changes. Leveraging Low-Rank Adaptation (LoRA), LUNE targets intermediate representations to suppress (or replace) requested knowledge with an order-of-magnitude lower compute and memory than full fine-tuning or direct weight editing. Extensive experiments on multiple factual unlearning tasks show that LUNE: (I) achieves effectiveness comparable to full fine-tuning and memory-editing methods; and (II) reduces computational cost by about an order of magnitude.
DiSec: Mitigating Backdoors in Pre-trained Language Models via Disentanglement of Adversarial Weights for Secure Fine-Tuning
PDF ↗Task-agnostic backdoor attacks can contaminate pre-trained language models (PLMs) in a way that survives downstream adaptation, even under full fine-tuning, making it difficult for practitioners to trust third-party checkpoints. Existing defenses often rely on privileged assumptions (e.g., access to poisoned data or trigger/target knowledge), thereby limiting their applicability in realistic settings. We present DiSec, a robust and label-efficient purification framework that uses only clean auxiliary text and does not rely on downstream supervision or attack signatures. DiSec elicits model-internal signals from this clean data to separate suspicious parameter components that are inconsistent with benign behavior, and then flags anomalous structures by jointly leveraging complementary spectral and generative views of outliers. Finally, DiSec performs a structure-preserving repair via layer-local prototype-based mean correction, yielding an idempotent update that depends only on non-adversarial statistics. Across diverse downstream classification tasks and PLM backdoor strategies, DiSec substantially suppresses attack success while preserving clean-task utility, offering a practical path to securing fully fine-tuned PLMs before deployment. The codes are publicly available at https://github.com/das-sunanda/DiSec.
Mid-training has become critical for enhancing the knowledge and reasoning ability of large language models (LLMs), especially through the utilization of large-scale synthetic data. However, existing data synthesis methods often generate simplistic and homogeneous QA pairs, with limited scale and diversity. To address this, we propose BoostQA, a novel framework designed to synthesize large-scale, diverse, and high-quality QA data for mid-training. BoostQA introduces model probes during mid-training for the first time and implements STEM-focused multi-grade synthesis to boost data diversity as well as high-difficulty synthesis to alleviate difficulty degradation, followed by answer refinement to further improve quality. Extensive experiments by mid-training Llama-3 8B demonstrate that using only 20B-token BoostQA data achieves a significant average improvement of **12.74%** on MMLU and CMMLU over the pre-training baseline. After mid-training on 500B tokens, including 100B-token BoostQA data, our model achieves SOTA average results across benchmarks among mainstream models of comparable size. BoostQA also demonstrates robust scalability, with performance consistently improving as model size, data volume, and initial FLOPs scale.
CaM-HG: Causal-Enhanced MoE and Hypergraphs Network for Incomplete Multimodal Emotion Recognition in Conversations
PDF ↗Multimodal Emotion Recognition in Conversation (MERC) relies on integrating heterogeneous signals, yet real-world modality missingness frequently disrupts these systems. We contend that missingness is not merely a loss of data fidelity but a rupture of the fine-grained inter-modal causal chains essential for reasoning. Existing methods, which primarily focus on statistical reconstruction, often fail to bridge these logical gaps, effectively leaving semantic holes. To address this, we propose the Causal-Enhanced Mixture-of-Experts and Hypergraph Network (CaM-HG), employing a "restore-then-mine" paradigm. First, a Causal-Enhanced MoE module conditions experts on historical context to synthesize missing features that are both realistic and causally consistent, thereby patching the broken topology. Subsequently, an Asymmetric Causal Dynamic Hypergraph mines high-order correlations from the restored graph while enforcing strict temporal causality. Experiments on IEMOCAP, CMU-MOSI, and CMU-MOSEI show consistent improvements in terms of WAF1 and accuracy over strong baselines, e.g., surpassing SOTA benchmarks by 1.43% and 1.25% on IEMOCAP. The source code is included in the supplementary material.
QuCo-RAG: Quantifying Uncertainty from the Pre-training Corpus for Dynamic Retrieval-Augmented Generation
PDF ↗Dynamic Retrieval-Augmented Generation adaptively determines when to retrieve during generation to mitigate hallucinations in large language models (LLMs). However, existing methods rely on model-internal signals (e.g., logits, entropy), which are fundamentally unreliable because LLMs are typically ill-calibrated and often exhibit high confidence in erroneous outputs. We propose QuCo-RAG, which shifts from subjective confidence to objective statistics computed from pre-training data. Our method quantifies uncertainty through two stages: (1) before generation, we identify low-frequency entities indicating long-tail knowledge gaps; (2) during generation, we verify entity co-occurrence in the pre-training corpus, where zero co-occurrence often signals hallucination risk. Both stages leverage Infini-gram for millisecond-latency queries over 4 trillion tokens, triggering retrieval when uncertainty is high. Experiments on multi-hop QA benchmarks show QuCo-RAG achieves EM gains of 5–12 points over state-of-the-art baselines with OLMo-2 models, and transfers effectively to models with undisclosed pre-training data (Llama-3, Qwen2.5, GPT-4.1/5-chat), improving EM by up to 14 points. Generalization to long-form generation and biomedical QA further validates the robustness of our paradigm. These results establish corpus-grounded verification as a principled, practically model-agnostic paradigm for dynamic RAG.
Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks. Our code is available at https://github.com/DCDmllm/CoMoL.
Large Language Model agents deployed in complex environments frequently encounter a conflict between maximizing goal achievement and adhering to safety constraints. This paper identifies a new concept called Agentic Pressure, which characterizes the endogenous tension emerging when compliant execution becomes infeasible. We demonstrate that under this pressure agents exhibit normative drift where they strategically sacrifice safety to preserve utility. Notably we find that advanced reasoning capabilities accelerate this decline as models construct linguistic rationalizations to justify violation. Finally, we analyze the root causes and explore preliminary mitigation strategies, such as pressure isolation, which attempts to restore alignment by decoupling decision-making from pressure signals.
The paradigm of programmable diagram generation is evolving rapidly, playing a crucial role in structured visualization. However, most existing studies are confined to a narrow range of task formulations and language support, constraining their applicability to diverse diagram types. In this work, we propose OmniDiagram, a unified framework that incorporates diverse diagram code languages and task definitions. To address the challenge of aligning code logic with visual fidelity in Reinforcement Learning (RL), we introduce a novel visual feedback strategy named Visual Interrogation Verifies All (Viva). Unlike brittle syntax-based rules or pixel-level matching, Viva rewards the visual structure of rendered diagrams through a generative approach. Specifically, Viva actively generates targeted visual inquiries to scrutinize diagram visual fidelity and provides fine-grained feedback for optimization. This mechanism facilitates a self-evolving training process, effectively obviating the need for manually annotated ground truth code. Furthermore, we construct M3^2Diagram, the first large-scale diagram code generation dataset, containing over 196k high-quality instances. Experimental results confirm that the combination of SFT and our Viva-based RL allows OmniDiagram to establish a new state-of-the-art (SOTA) across diagram code generation benchmarks.
Causal reasoning is a crucial component of understanding complex phenomena and building intelligent systems. Recent advancements in large language models (LLMs) have demonstrated their strong capabilities in reasoning tasks; however, their true understanding of causal relationships remains limited, particularly in cases where causal chains are misidentified or reliance on empirical inference occurs. To mitigate the risk that models misclassify data as false positives due to these issues, we introduce CausalityCheck, an automated tool designed to efficiently generate causal reasoning checklists. This checklist enables the creation of multi-task causal reasoning datasets with task generalization and reasoning robustness from a single causal reasoning dataset. Using CausalityCheck, we developed CausalityCheck-CP to assess the causal reasoning abilities of 18 LLMs. This framework also measures the extent to which causal chains are misidentified or rely on empirical inferences. Our results indicate that the current large language models still face two critical issues when handling complex causal reasoning tasks: incorrect identification of causal chains and reliance on empirical inference. The code and data are available at https://github.com/dzh597/CausalityCheck.
The emergence of reasoning models, exemplified by OpenAI o1, signifies a transition from intuitive to deliberative cognition, effectively reorienting the scaling laws from pre-training paradigms toward test-time computation. While Monte Carlo Tree Search (MCTS) has shown promise in this domain, existing approaches typically treat each rollout as an isolated trajectory. This lack of information sharing leads to severe inefficiency and substantial computational redundancy, as the search process fails to leverage insights from prior explorations. To address these limitations, we propose PRISM-MCTS, a novel reasoning framework that draws inspiration from human parallel thinking and reflective processes. PRISM-MCTS integrates a Process Reward Model (PRM) with a dynamic shared memory, capturing both "Heuristics" and "Fallacies". By reinforcing successful strategies and pruning error-prone branches, PRISM-MCTS effectively achieves refinement. Furthermore, we develop a data-efficient training strategy for the PRM, achieving high-fidelity evaluation under a few-shot regime. Empirical evaluations across diverse reasoning benchmarks substantiate the efficacy of PRISM-MCTS. Notably, it halves the trajectory requirements on GPQA while surpassing MCTS-RAG and Search-o1, demonstrating that it scales inference by reasoning judiciously rather than exhaustively.
Tree-CoT-RT: An Explainable Multi-Path Tree-Guided Chain-of-Thought and Reinforcement Learning Framework for Aspect Sentiment Quad Prediction
PDF ↗Aspect Sentiment Quad Prediction (ASQP) is a fundamental yet challenging task in fine-grained sentiment analysis, particularly when aspects or opinions are implicit. Existing methods often lack explainability and generalization, making it difficult to justify inference decisions and to detect implicit sentiment across domains and varied expression patterns. To address these limitations, we propose Tree-CoT-RT, an explainable multi-path tree-guided chain-of-thought and reinforcement learning framework specifically designed for ASQP. The core idea is to use sentiment tree structures to design type-specific reasoning templates that guide LLMs in generating explainable chains, including both final sentiment quadruples and intermediate inference steps for transparent implicit reasoning. However, the generated reasoning chains often vary in quality and may contain logical inconsistencies. To mitigate this, we introduce a reinforcement learning strategy with a rule-based reward function to generate high-quality reasoning traces, which are then used to fine-tune the LLM and enable controlled sampling. Experiments on benchmark datasets demonstrate that Tree-CoT-RT substantially outperforms strong baselines, particularly in scenarios involving implicit sentiment analysis.
mPresenter: An Agentic Framework for Generating Multilingual Presentation Videos from Scientific Papers
PDF ↗Generating presentation videos from scientific papers is challenging due to the need for long-document discourse planning and cross-lingual grounding. Existing Paper2Video systems are largely monolingual and often rely on single-pass pipelines, which can limit the coherence and informativeness of the resulting presentations.We present mPresenter, a multilingual agentic Paper2Video system that decomposes the task into planning, audience-oriented critique, layout-aware slide generation, and multilingual figure interpretation, enabling iterative refinement at the discourse level. To facilitate reproducible evaluation, we also introduce mPreBench, a multilingual benchmark that evaluates presentation videos via question answering as a proxy for effective information transfer. Experimental results indicate that mPresenter improves question-answering accuracy relative to prior systems, while maintaining affordable cost and latency.
Learning Continuous Temporal Dynamics on Symplectic Manifolds for Temporal Knowledge Graph Embedding
PDF ↗Temporal knowledge graph embedding (TKGE) aims to model the temporal evolution of relational facts. However, existing approaches predominantly rely on discrete timestamp lookup tables and high-dimensional embedding spaces, which lack explicit structural constraints for continuous-time dynamics. As a result, temporal patterns are often captured through capacity scaling rather than principled dynamic modeling, leading to limited parameter efficiency and scalability.To address these limitations, we propose , a physics-inspired framework that embeds temporal dynamics into a symplectic phase space. Our model introduces a structure-preserving Hamiltonian evolution mechanism based on a pairwise-decoupled Hamiltonian generator and its Cayley transform, ensuring that temporal transformations adhere to the symplectic group \mathrm{Sp}(2d) and preserve phase-space volume with linear computational complexity. In addition, we design a Time-Aware Parameter Modulation mechanism that integrates continuous Rotary Time Embeddings via Feature-wise Linear Modulation, enabling smooth temporal evolution while capturing event-driven variations. Theoretical analysis establishes the geometric validity of the proposed framework. Extensive experiments on standard TKGE benchmarks demonstrate that achieves competitive performance with substantially lower embedding dimensions. Furthermore, empirical results show that the proposed continuous Hamiltonian evolution facilitates generalization to unseen timestamps by learning transferable temporal dynamics from the underlying geometric structure.
Stable and Explainable Personality Trait Evaluation in Large Language Models with Internal Activations
PDF ↗Evaluating personality-related tendencies in Large Language Models (LLMs) helps characterize model behavior, compare models beyond task accuracy, and support responsible deployment in socially interactive settings. However, existing questionnaire-based evaluation methods exhibit limited stability and offer little explainability, as their results are highly sensitive to minor variations in prompt phrasing or role-play configurations. To address these limitations, we propose an internal-activation–based approach, termed Persona-Vector Neutrality Interpolation (PVNI), for stable and explainable personality trait evaluation in LLMs. PVNI extracts a persona vector associated with a target personality trait from the model’s internal activations using contrastive prompts. It then estimates the corresponding neutral score by interpolating along the persona vector as an anchor axis, enabling an interpretable comparison between the neutral prompt representation and the persona direction. We provide a theoretical analysis of the effectiveness and generalization properties of PVNI. Extensive experiments across diverse LLMs demonstrate that PVNI yields substantially more stable personality trait evaluations than existing methods, even under questionnaire and role-play variants.
Speculative decoding has emerged as a promising approach to accelerate autoregressive inference in large language models (LLMs). Self-draft methods, which leverage the base LLM itself for speculation, avoid the overhead of auxiliary draft models but face critical limitations: shallow layers often produce overconfident yet incorrect token predictions, and the presence of difficult tokens in a draft sequence forces redundant computation through deeper layers, undermining both draft acceptance and overall speedup. To address these issues, we propose a novel self-draft framework that suppresses spurious confidence via layer-wise temperature annealing in early-exit decision and adaptively bounds speculation length based on token-wise decoding difficulty. By reprocessing the hidden states of draft tokens in a unified parallel pass through deep layers when speculation terminates, our method maintains exact output equivalence with the original model while maximizing computational efficiency. It requires no modifications to the base LLM parameters and achieves up to 2.33× wall-time speedup over standard autoregressive decoding across diverse long-form generation tasks and multiple model architectures.
Large Language Models (LLMs) incur quadratic attention complexity with input length, creating a major time bottleneck in the prefilling stage. Existing acceleration methods largely exploit attention score sparsity by estimating blocks with high attention scores and applying dynamic sparse attention. In this work, we identify another untapped form of sparsity in the prefilling stage, namely decoding-time contribution sparsity, where many attention blocks exhibit nontrivial attention scores during prefilling yet contribute negligibly to subsequent decoding. Building on this observation, we propose TriangleMix, which replaces dense attention with Triangle attention in a subset of layers. Extensive experiments demonstrate that TriangleMix achieves near-lossless performance on both long-context and long-context reasoning benchmarks, while significantly improving efficiency. For 128K inputs, Triangle attention in the subset of layers achieves a 15.3 \times speedup in attention kernel computation, significantly exceeding the acceleration of typical dynamic sparse methods ( 1.9 \times to 3.4 \times ). Furthermore, TriangleMix can be seamlessly combined with dynamic sparsity approaches, delivering an additional 6%–19% reduction in TTFT over using dynamic sparsity alone.
ProCeedRL: Process Critic with Explorative Demonstration Reinforcement Learning for LLM Agentic Reasoning
PDF ↗Reinforcement Learning (RL) significantly enhances the reasoning abilities of large language models (LLMs), yet applying it to multi-turn agentic tasks remains challenging due to the long-horizon nature of interactions and the stochasticity of environmental feedback.We identify a structural failure mode in agentic exploration: suboptimal actions elicit noisy observations into misleading contexts, which further weaken subsequent decision-making, making recovery increasingly difficult.This cumulative feedback loop of errors renders standard exploration strategies ineffective and susceptible to the model’s reasoning and the environment’s randomness.To mitigate this issue, we propose **ProCeedRL**: **Pro**cess **C**ritic with **E**xplorativ**e** **D**emonstration RL, shifting exploration from passive selection to active intervention.ProCeedRL employs a process-level critic to monitor interactions in real time, incorporating reflection-based demonstrations to guide agents in stopping the accumulation of errors.We find that this approach significantly exceeds the model’s saturated exploration performance, demonstrating substantial exploratory benefits.By learning from exploratory demonstrations and on-policy samples, ProCeedRL significantly improves exploration efficiency and achieves superior performance on complex deep search and embodied tasks.
Recently, various excellent and powerful large language models (LLMs) have been utilized to solve a wide range of human problems. However, when faced with complex problems, most users are often unable to provide accurate and effective prompts to interact with LLMs, thus limiting their performance. To address this challenge, we propose Prompt-R1, an end-to-end reinforcement learning framework that utilizes a small-scale LLM (as agent) to collaborate with large-scale LLMs (as environment), replacing users to interact better. This collaboration is presented as a multi-turn interaction, where the small-scale LLM thinks and generates prompts, and the large-scale LLM performs complex reasoning. A double-constrained reward is designed to optimize correctness and quality of generation. Prompt-R1 provides a plug-and-play framework that supports both inference and training with various large-scale LLMs. Experimental results on twelve datasets show that Prompt-R1 significantly outperforms baseline LLMs across various tasks.Our code is available at https://github.com/QwenQKing/Prompt-R1.
Large language models (LLMs) reach state-of-the-art performance across many NLP tasks, but their large parameter counts introduce heavy computational and memory overhead, which complicates deployment in resource-constrained settings. Pruning is a standard compression strategy that induces sparsity to lower these costs. However, most pruning methods for LLMs depend on calibration data and expensive weight updates, which limits practical scalability. To address these limitations, we introduce Haar Wavelet Subband Pruning (), a post-training framework that requires no calibration data and no weight updates. applies a two-dimensional Haar wavelet transform to each weight matrix and decomposes it into four frequency subbands. It then assigns a uniform sparsity ratio to all subbands so that both low- and high-frequency components are retained in a balanced manner. Our theoretical analysis shows that the subband design of provides a deterministic per-subband retention guarantee, which helps mitigate the potential bias of global magnitude pruning toward dominant frequency components. Experiments on the LLaMA, OPT and Qwen model families show that achieves competitive accuracy relative to strong pruning baselines while substantially reducing pruning time. Compared with magnitude pruning, which serves as a simple calibration-free baseline, generally achieves better downstream performance across a wide range of sparsity levels and model scales.