Selected work

Representative Papers

MaskTab: Scalable Masked Tabular Pretraining with Scaling Laws and Distillation for Industrial Classification

May 11, 2026

Industrial-scale tabular data often confronts challenges such as high dimensionality, extensive missing values, and scarce annotations, with a notable absence of a general-purpose self-supervised pretraining framework. This work proposes MaskTab, a unified pretraining approach that employs learnable missing tokens to distinguish between structural and random missingness, integrates a dual-path hybrid supervision architecture to jointly optimize masked reconstruction and downstream task objectives, and incorporates a Mixture-of-Experts (MoE)-enhanced loss with knowledge distillation. Evaluated on industrial benchmarks, MaskTab achieves substantial performance gains (AUC +5.04%, KS +8.28%). Moreover, the distilled lightweight model retains strong performance under stringent latency and interpretability constraints (AUC +2.55%, KS +4.85%) while demonstrating enhanced robustness to distribution shifts.

0 citationsRead paper

All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG

Apr 22, 2026

This work addresses a significant language bias in multilingual retrieval-augmented generation (mRAG) systems during the reranking stage, where existing rerankers disproportionately favor evidence in English or the query language, thereby suppressing critical cross-lingual information. The study presents the first quantitative analysis of this bias and reveals a substantial performance gap between current rerankers and the theoretical upper bound through oracle evidence estimation. To mitigate this issue, the authors propose LAURA, a language-agnostic, utility-driven reranking alignment method that explicitly aligns multilingual evidence ranking with downstream generation objectives, eliminating reliance on monolingual or query-language cues. Experimental results demonstrate that LAURA consistently improves question-answering accuracy and generation quality across diverse languages and generative models, effectively alleviating language bias.

0 citationsRead paper

PolicySim: An LLM-Based Agent Social Simulation Sandbox for Proactive Policy Optimization

Mar 20, 2026

This work proposes PolicySim, a large language model (LLM)-based social simulation sandbox designed to proactively evaluate the societal risks of platform intervention policies—such as recommendation algorithms and content filtering—that may exacerbate filter bubbles and polarization. Unlike existing approaches reliant on post-hoc A/B testing, PolicySim introduces, for the first time in LLM-driven simulations, a dynamic feedback mechanism that models platform interventions through contextual bandits and message-passing architectures to capture evolving user networks. User agents are trained via supervised fine-tuning (SFT) and direct preference optimization (DPO) to adaptively respond to interventions. The framework significantly enhances simulation fidelity and operational utility at both micro-behavioral and macro-ecological levels, outperforming current simulation paradigms and enabling prospective assessment and refinement of intervention strategies.

0 citationsRead paper

Learning from the Irrecoverable: Error-Localized Policy Optimization for Tool-Integrated LLM Reasoning

Feb 10, 2026

This work addresses the inefficiency in tool-integrated reasoning caused by sparse delayed rewards and weak step-level credit assignment, particularly when early irreversible errors dominate task outcomes in long-horizon trajectories. We propose the first localized policy optimization framework specifically designed to handle irreversible errors: by identifying the first such error, we construct a binary-search-based rollback tree to generate fine-grained learning signals, and integrate hierarchical advantage attribution with error-local adaptive clipping to enable precise credit assignment and targeted correction of critical errors and their downstream actions. Evaluated on benchmarks spanning mathematical reasoning, scientific question answering, and code execution, our method significantly outperforms strong baselines, improving Pass@K and Major@K scalability, rollback ranking quality, and tool invocation efficiency.

0 citationsRead paper

VULSOVER: Vulnerability Detection via LLM-Driven Constraint Solving

Aug 31, 2025

Traditional vulnerability detection methods suffer from either over-reliance on handcrafted rules or hallucination and instability inherent in large language models (LLMs), compromising both accuracy and robustness. This paper proposes an LLM-driven constraint-solving framework that formulates vulnerability detection as a semantics-aware constraint satisfaction problem, synergistically integrating static application security testing (SAST) with LLM-based code understanding: SAST extracts structured, syntactic constraints, while the LLM performs semantic-level reasoning and validation—enabling expert-like collaborative inference. By decoupling structural analysis from semantic interpretation, the approach avoids the semantic blind spots of pure rule-based matching and mitigates the unreliability of end-to-end LLM invocation. Evaluated on the OWASP Benchmark, it achieves 96.29% accuracy, 96.55% F1-score, and 100% recall. Furthermore, it identifies 15 previously unreported high-severity vulnerabilities across real-world GitHub projects.

0 citationsRead paper

Recent publications

Latest Papers

MaskTab: Scalable Masked Tabular Pretraining with Scaling Laws and Distillation for Industrial Classification

May 11, 2026

Industrial-scale tabular data often confronts challenges such as high dimensionality, extensive missing values, and scarce annotations, with a notable absence of a general-purpose self-supervised pretraining framework. This work proposes MaskTab, a unified pretraining approach that employs learnable missing tokens to distinguish between structural and random missingness, integrates a dual-path hybrid supervision architecture to jointly optimize masked reconstruction and downstream task objectives, and incorporates a Mixture-of-Experts (MoE)-enhanced loss with knowledge distillation. Evaluated on industrial benchmarks, MaskTab achieves substantial performance gains (AUC +5.04%, KS +8.28%). Moreover, the distilled lightweight model retains strong performance under stringent latency and interpretability constraints (AUC +2.55%, KS +4.85%) while demonstrating enhanced robustness to distribution shifts.

0 citationsRead paper

All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG

Apr 22, 2026

This work addresses a significant language bias in multilingual retrieval-augmented generation (mRAG) systems during the reranking stage, where existing rerankers disproportionately favor evidence in English or the query language, thereby suppressing critical cross-lingual information. The study presents the first quantitative analysis of this bias and reveals a substantial performance gap between current rerankers and the theoretical upper bound through oracle evidence estimation. To mitigate this issue, the authors propose LAURA, a language-agnostic, utility-driven reranking alignment method that explicitly aligns multilingual evidence ranking with downstream generation objectives, eliminating reliance on monolingual or query-language cues. Experimental results demonstrate that LAURA consistently improves question-answering accuracy and generation quality across diverse languages and generative models, effectively alleviating language bias.

0 citationsRead paper

PolicySim: An LLM-Based Agent Social Simulation Sandbox for Proactive Policy Optimization

Mar 20, 2026

This work proposes PolicySim, a large language model (LLM)-based social simulation sandbox designed to proactively evaluate the societal risks of platform intervention policies—such as recommendation algorithms and content filtering—that may exacerbate filter bubbles and polarization. Unlike existing approaches reliant on post-hoc A/B testing, PolicySim introduces, for the first time in LLM-driven simulations, a dynamic feedback mechanism that models platform interventions through contextual bandits and message-passing architectures to capture evolving user networks. User agents are trained via supervised fine-tuning (SFT) and direct preference optimization (DPO) to adaptively respond to interventions. The framework significantly enhances simulation fidelity and operational utility at both micro-behavioral and macro-ecological levels, outperforming current simulation paradigms and enabling prospective assessment and refinement of intervention strategies.

0 citationsRead paper

Learning from the Irrecoverable: Error-Localized Policy Optimization for Tool-Integrated LLM Reasoning

Feb 10, 2026

This work addresses the inefficiency in tool-integrated reasoning caused by sparse delayed rewards and weak step-level credit assignment, particularly when early irreversible errors dominate task outcomes in long-horizon trajectories. We propose the first localized policy optimization framework specifically designed to handle irreversible errors: by identifying the first such error, we construct a binary-search-based rollback tree to generate fine-grained learning signals, and integrate hierarchical advantage attribution with error-local adaptive clipping to enable precise credit assignment and targeted correction of critical errors and their downstream actions. Evaluated on benchmarks spanning mathematical reasoning, scientific question answering, and code execution, our method significantly outperforms strong baselines, improving Pass@K and Major@K scalability, rollback ranking quality, and tool invocation efficiency.

0 citationsRead paper

VULSOVER: Vulnerability Detection via LLM-Driven Constraint Solving

Aug 31, 2025

Traditional vulnerability detection methods suffer from either over-reliance on handcrafted rules or hallucination and instability inherent in large language models (LLMs), compromising both accuracy and robustness. This paper proposes an LLM-driven constraint-solving framework that formulates vulnerability detection as a semantics-aware constraint satisfaction problem, synergistically integrating static application security testing (SAST) with LLM-based code understanding: SAST extracts structured, syntactic constraints, while the LLM performs semantic-level reasoning and validation—enabling expert-like collaborative inference. By decoupling structural analysis from semantic interpretation, the approach avoids the semantic blind spots of pure rule-based matching and mitigates the unreliability of end-to-end LLM invocation. Evaluated on the OWASP Benchmark, it achieves 96.29% accuracy, 96.55% F1-score, and 100% recall. Furthermore, it identifies 15 previously unreported high-severity vulnerabilities across real-world GitHub projects.

0 citationsRead paper