Selected work

Representative Papers

Bootstrapping Niche Multilingual Code Translation via Reinforcement Learning with Execution-Based Verifiable Supervision

Aug 13, 2026

This study addresses the poor executability of code translation for low-resource languages caused by insufficient parallel supervision. We propose an execution feedback-based reinforcement learning framework that trains a reward model using execution-verified data and optimizes large language models via the GRPO algorithm to enhance cross-language code generation correctness. Additionally, we introduce Humaneval-X++, a multilingual evaluation benchmark. Experiments demonstrate that a 4B-parameter model achieves an average performance improvement of 13% on this benchmark, with gains reaching 21% for medium-resource languages. The approach successfully enables executable code translation across 600 language pairs, significantly mitigating alignment challenges in low-resource scenarios.

0 citationsRead paper

An Open and Reproducible Deep Research Agent for Long-Form Question Answering

Dec 15, 2025

This work addresses open-domain long-form question answering (LFQA) by proposing a reproducible deep-research agent framework. Methodologically, it integrates open-source large language models (LLMs) with web search APIs to construct an iterative retrieval–reasoning–synthesis pipeline and introduces the first LLM-as-a-judge-based multi-dimensional preference tuning mechanism—explicitly optimizing for clarity, insightfulness, and factual accuracy—to enable end-to-end optimization. Contributions include: (1) the first open-source system supporting multi-turn, real-world deep research; (2) an interpretable and scalable paradigm for automated multi-dimensional evaluation and alignment; and (3) state-of-the-art performance on the NeurIPS 2025 MMU-RAG text-to-text track, achieving systematic improvements across all three core evaluation metrics. All code and experimental configurations are publicly released.

0 citationsRead paper

Dynamic Injection of Entity Knowledge into Dense Retrievers

Jul 05, 2025

Dense retrievers suffer from limited performance on entity-centric queries due to insufficient prior knowledge of low-frequency entities. To address this, we propose a knowledge-augmented retrieval framework that requires no retraining: it dynamically integrates entity information from external knowledge bases via a context-entity cross-attention mechanism and introduces a plug-and-play dynamic entity embedding module enabling real-time injection of new knowledge. Built upon the BERT/bge-base architecture, our approach preserves model lightweightness and deployment efficiency. Extensive evaluation on EntityQuestions, WebQs, and TREC-QA demonstrates significant improvements—particularly a 12.6% gain in Recall@20 on EntityQuestions—and achieves state-of-the-art results among models of comparable size on two benchmarks. The method notably enhances recall and generalization for queries involving low-frequency entities.

0 citationsRead paper

Recent publications

Latest Papers

Bootstrapping Niche Multilingual Code Translation via Reinforcement Learning with Execution-Based Verifiable Supervision

Aug 13, 2026

This study addresses the poor executability of code translation for low-resource languages caused by insufficient parallel supervision. We propose an execution feedback-based reinforcement learning framework that trains a reward model using execution-verified data and optimizes large language models via the GRPO algorithm to enhance cross-language code generation correctness. Additionally, we introduce Humaneval-X++, a multilingual evaluation benchmark. Experiments demonstrate that a 4B-parameter model achieves an average performance improvement of 13% on this benchmark, with gains reaching 21% for medium-resource languages. The approach successfully enables executable code translation across 600 language pairs, significantly mitigating alignment challenges in low-resource scenarios.

0 citationsRead paper

An Open and Reproducible Deep Research Agent for Long-Form Question Answering

Dec 15, 2025

This work addresses open-domain long-form question answering (LFQA) by proposing a reproducible deep-research agent framework. Methodologically, it integrates open-source large language models (LLMs) with web search APIs to construct an iterative retrieval–reasoning–synthesis pipeline and introduces the first LLM-as-a-judge-based multi-dimensional preference tuning mechanism—explicitly optimizing for clarity, insightfulness, and factual accuracy—to enable end-to-end optimization. Contributions include: (1) the first open-source system supporting multi-turn, real-world deep research; (2) an interpretable and scalable paradigm for automated multi-dimensional evaluation and alignment; and (3) state-of-the-art performance on the NeurIPS 2025 MMU-RAG text-to-text track, achieving systematic improvements across all three core evaluation metrics. All code and experimental configurations are publicly released.

0 citationsRead paper

Dynamic Injection of Entity Knowledge into Dense Retrievers

Jul 05, 2025

Dense retrievers suffer from limited performance on entity-centric queries due to insufficient prior knowledge of low-frequency entities. To address this, we propose a knowledge-augmented retrieval framework that requires no retraining: it dynamically integrates entity information from external knowledge bases via a context-entity cross-attention mechanism and introduces a plug-and-play dynamic entity embedding module enabling real-time injection of new knowledge. Built upon the BERT/bge-base architecture, our approach preserves model lightweightness and deployment efficiency. Extensive evaluation on EntityQuestions, WebQs, and TREC-QA demonstrates significant improvements—particularly a 12.6% gain in Recall@20 on EntityQuestions—and achieves state-of-the-art results among models of comparable size on two benchmarks. The method notably enhances recall and generalization for queries involving low-frequency entities.

0 citationsRead paper