📝 Publications

Soft Reasoning: Navigating Solution Spaces in Large Language Models through Controlled Embedding Exploration
Qinglin Zhu, Runcong Zhao, Hanqi Yan, Yulan He, Yudong Chen, Lin Gui.

• Proposes an embedding-based search framework that guides LLM generation by optimising the first token’s embedding.
  
• Combines embedding perturbation for controlled exploration with Bayesian optimisation via a verifier-guided objective, balancing exploration and exploitation.

Mirror: A Multiple-perspective Self-Reflection Method for Knowledge-rich Reasoning
Hanqi Yan*, Qinglin Zhu* , Xinyu Wang, Lin Gui, Yulan He.

• Proposes Mirror, enabling LLMs to reflect from multiple perspectives via Navigator–Reasoner cooperation.
  
• Encourages both diverse and consistent reasoning to overcome self-reflection traps.
  

Latent Refinement Decoding: Enhancing Diffusion-Based Language Models by Refining Belief States
Qinglin Zhu, Yizhen Yao, Runcong Zhao, Yanzheng Xiang, Amrutha Saseendran, Chen Jin, Philip Alexander Teare, Bin Liang, Yulan He, Lin Gui.

• Proposes Latent Refinement Decoding (LRD), a two-stage framework that tackles information loss and premature commitment in diffusion-based language models via latent refinement and predictive feedback.
  
• Enables faster, globally consistent parallel generation as a principled alternative to autoregressive decoding.
  

Pull Requests as a Training Signal for Repo-Level Code Editing
Qinglin Zhu, Tianyu Chen, Shuai Lu, Lei Ji, Runcong Zhao, Murong Ma, Xiangxiang Dai, Yulan He, Lin Gui, Yeyun Gong.

• Introduces Clean-PR, a pipeline that converts noisy PR diffs into structured edit blocks, yielding 2M training samples across 12 languages.
  
• Achieves +13.6% on SWE-bench Lite and +12.3% on SWE-bench Verified, demonstrating the value of real-world PRs for repo-level code editing.

Large Language Models Fall Short: Understanding Complex Relationships in Detective Narratives
Runcong Zhao*, Qinglin Zhu* , Hainiu Xu, Jiazheng Li, Yuxiang Zhou, Yulan He, Lin Gui.

• Existing datasets for narrative understanding often fail to represent the complexity and uncertainty of relationships in real-life social scenarios.
• To address this gap, we introduce a new benchmark, Conan, designed for extracting and analysing intricate character relation graphs from detective narratives.

SymbolicThought: Integrating Language Models and Symbolic Reasoning for Consistent and Interpretable Human Relationship Understanding
Runcong Zhao, Qinglin Zhu*, Hainiu Xu, Bin Liang, Yulan He, Lin Gui.

• Proposes SymbolicThought, a human-in-the-loop system combining LLM extraction and symbolic reasoning for character relationship understanding.
• Supports editable relationship graphs, logical constraints, and interactive conflict resolution.

PLAYER*: Enhancing LLM-based Multi-Agent Communication and Interaction in Murder Mystery Games
Qinglin Zhu, Runcong Zhao, Jinhua Du, Lin Gui, Yulan He.

• We propose PLAYER*, a novel framework for Murder Mystery Games (剧本杀) using an anytime sampling-based planner and a questioning-driven search framework.

Spectrum Projection Score: Aligning Retrieved Summaries with Reader Models in Retrieval-Augmented Generation
Zhanghao Hu, Qinglin Zhu, Siya Qi, Yulan He, Hanqi Yan, Lin Gui.

• Proposes SPS, a supervision-free metric to assess semantic alignment between retrieved summaries and LLM representations.
  
• Introduces xCompress, an inference-time controller that ranks and compresses retrievals to improve generation and clarify retrieval–generation interaction.
  

Beyond Prompting: An Efficient Embedding Framework for Open-Domain Question Answering
Zhanghao Hu, Hanqi Yan, Qinglin Zhu†, Zhenyi Shen, Yulan He, Lin Gui.

• Proposes EmbQA, an embedding-level framework for open-domain QA that optimizes retrieval with unsupervised contrastive learning and improves answer diversity via exploratory embeddings.
  

Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
Zhanghao Hu, Qinglin Zhu, Hanqi Yan, Yulan He, Lin Gui.

• Proposes xMemory, which decouples agent memories into semantic components and organises them hierarchically.
  
• Retrieves via top-down aggregation to capture diverse themes, outperforming standard RAG on long-horizon agent tasks.