Selected work

Representative Papers

SOV-CAD: Stepwise Orthographic Views Guided CAD Modeling Sequence Reconstruction

Jul 05, 2026

Reconstructing CAD modeling sequences from images often fails to capture the iterative, feedback-driven nature of human design. This work formulates the task as a sequential decision-making problem and introduces a novel mechanism that combines stepwise orthographic view supervision with geometric alignment rewards. At each step, continuous visual feedback—comprising orthographic views, incremental model projections, and the current sketch—guides action selection. Built upon offline reinforcement learning and the Decision Transformer architecture, the proposed method significantly outperforms existing approaches in both reconstruction accuracy and data efficiency, achieving state-of-the-art performance while better reflecting authentic design workflows.

0 citationsRead paper

RoadMapper: A Multi-Agent System for Roadmap Generation of Solving Complex Research Problems

Apr 30, 2026

This study addresses the limitations of large language models in generating scientific research roadmaps—specifically, insufficient domain expertise, suboptimal task decomposition, and logical inconsistencies—by proposing RoadMapper, a multi-agent collaborative framework that structures the generation process into three phases: initial drafting, knowledge augmentation, and iterative critique-revision-evaluation. The work introduces RoadMap, the first benchmark dataset for research roadmap generation, and integrates knowledge enhancement with multi-agent coordination. Experimental results demonstrate that RoadMapper significantly outperforms baseline methods in domain specificity, logical coherence, and practical utility, achieving an average performance improvement of over 8% while reducing generation time to merely 16% of that required by human experts.

0 citationsRead paper

OGER: A Robust Offline-Guided Exploration Reward for Hybrid Reinforcement Learning

Apr 20, 2026

This work addresses the challenge that large language models struggle to effectively explore novel reasoning paths beyond their initial latent space in reinforcement learning settings. To overcome this limitation, the authors propose a hybrid framework that integrates offline multi-teacher collaborative guidance with online reinforcement learning. The approach leverages trajectory-level offline supervision, entropy-aware exploration reward modulation, and uncertainty-driven reward modeling to steer the model toward efficient autonomous exploration in both mathematical and general reasoning tasks. Empirical results demonstrate that the method significantly outperforms existing baselines on mathematical reasoning benchmarks and exhibits strong out-of-distribution generalization capabilities.

0 citationsRead paper

RETuning: Upgrading Inference-Time Scaling for Stock Movement Prediction with Large Language Models

Oct 24, 2025

Contemporary large language models (LLMs) exhibit two critical limitations in stock price forecasting: excessive reliance on analyst opinions and insufficient independent reasoning capability; and inability to effectively weigh contradictory evidence, resulting in unreliable predictions. To address these issues, we propose RETuning—a novel reflective evidence tuning framework that, for the first time, dynamically constructs analytical reasoning structures during inference. RETuning integrates heterogeneous long-context inputs (up to 32K tokens), including stock prices, news articles, research reports, fundamental data, and macroeconomic indicators, and performs evidence scoring and adaptive weighting. Crucially, it eliminates dependence on prior analyst views, enabling cold-start adaptation and out-of-distribution generalization. Evaluated on a comprehensive A-share dataset comprising over 200,000 samples, RETuning significantly improves accuracy across the three-class prediction task (upward/downward/hold), maintains robust inference stability for over six months, and demonstrates strong transferability to unseen stocks.

0 citationsRead paper

Exact Optimization for Minimum Dominating Sets

Aug 25, 2025

This paper addresses the NP-hard Minimum Dominating Set (MDS) problem by proposing ParDS, an exact algorithm based on the branch-and-bound framework. Methodologically, ParDS introduces two key innovations: (1) a high-accuracy linear programming relaxation to derive tighter theoretical lower bounds, and (2) dynamic graph reduction rules that simplify instances in real time during search, substantially shrinking the state space. Both components offer provable theoretical advantages over existing approaches in terms of bound quality and reduction efficiency. Experimental evaluation demonstrates that ParDS achieves the fastest runtime on 70% of benchmark graph classes, with up to 3,411× speedup on individual instances. Moreover, it successfully solves 16 previously unsolved instances—each exceeding the prior timeout threshold—within five hours. These results significantly extend the practical scalability of exact MDS computation on large-scale sparse graphs.

0 citationsRead paper

Recent publications

Latest Papers

SOV-CAD: Stepwise Orthographic Views Guided CAD Modeling Sequence Reconstruction

Jul 05, 2026

Reconstructing CAD modeling sequences from images often fails to capture the iterative, feedback-driven nature of human design. This work formulates the task as a sequential decision-making problem and introduces a novel mechanism that combines stepwise orthographic view supervision with geometric alignment rewards. At each step, continuous visual feedback—comprising orthographic views, incremental model projections, and the current sketch—guides action selection. Built upon offline reinforcement learning and the Decision Transformer architecture, the proposed method significantly outperforms existing approaches in both reconstruction accuracy and data efficiency, achieving state-of-the-art performance while better reflecting authentic design workflows.

0 citationsRead paper

RoadMapper: A Multi-Agent System for Roadmap Generation of Solving Complex Research Problems

Apr 30, 2026

This study addresses the limitations of large language models in generating scientific research roadmaps—specifically, insufficient domain expertise, suboptimal task decomposition, and logical inconsistencies—by proposing RoadMapper, a multi-agent collaborative framework that structures the generation process into three phases: initial drafting, knowledge augmentation, and iterative critique-revision-evaluation. The work introduces RoadMap, the first benchmark dataset for research roadmap generation, and integrates knowledge enhancement with multi-agent coordination. Experimental results demonstrate that RoadMapper significantly outperforms baseline methods in domain specificity, logical coherence, and practical utility, achieving an average performance improvement of over 8% while reducing generation time to merely 16% of that required by human experts.

0 citationsRead paper

OGER: A Robust Offline-Guided Exploration Reward for Hybrid Reinforcement Learning

Apr 20, 2026

This work addresses the challenge that large language models struggle to effectively explore novel reasoning paths beyond their initial latent space in reinforcement learning settings. To overcome this limitation, the authors propose a hybrid framework that integrates offline multi-teacher collaborative guidance with online reinforcement learning. The approach leverages trajectory-level offline supervision, entropy-aware exploration reward modulation, and uncertainty-driven reward modeling to steer the model toward efficient autonomous exploration in both mathematical and general reasoning tasks. Empirical results demonstrate that the method significantly outperforms existing baselines on mathematical reasoning benchmarks and exhibits strong out-of-distribution generalization capabilities.

0 citationsRead paper

RETuning: Upgrading Inference-Time Scaling for Stock Movement Prediction with Large Language Models

Oct 24, 2025

Contemporary large language models (LLMs) exhibit two critical limitations in stock price forecasting: excessive reliance on analyst opinions and insufficient independent reasoning capability; and inability to effectively weigh contradictory evidence, resulting in unreliable predictions. To address these issues, we propose RETuning—a novel reflective evidence tuning framework that, for the first time, dynamically constructs analytical reasoning structures during inference. RETuning integrates heterogeneous long-context inputs (up to 32K tokens), including stock prices, news articles, research reports, fundamental data, and macroeconomic indicators, and performs evidence scoring and adaptive weighting. Crucially, it eliminates dependence on prior analyst views, enabling cold-start adaptation and out-of-distribution generalization. Evaluated on a comprehensive A-share dataset comprising over 200,000 samples, RETuning significantly improves accuracy across the three-class prediction task (upward/downward/hold), maintains robust inference stability for over six months, and demonstrates strong transferability to unseen stocks.

0 citationsRead paper

Exact Optimization for Minimum Dominating Sets

Aug 25, 2025

This paper addresses the NP-hard Minimum Dominating Set (MDS) problem by proposing ParDS, an exact algorithm based on the branch-and-bound framework. Methodologically, ParDS introduces two key innovations: (1) a high-accuracy linear programming relaxation to derive tighter theoretical lower bounds, and (2) dynamic graph reduction rules that simplify instances in real time during search, substantially shrinking the state space. Both components offer provable theoretical advantages over existing approaches in terms of bound quality and reduction efficiency. Experimental evaluation demonstrates that ParDS achieves the fastest runtime on 70% of benchmark graph classes, with up to 3,411× speedup on individual instances. Moreover, it successfully solves 16 previously unsolved instances—each exceeding the prior timeout threshold—within five hours. These results significantly extend the practical scalability of exact MDS computation on large-scale sparse graphs.

0 citationsRead paper