Papers by Debing Zhang
MemSearcher: Iterative Memory Integration for Search Agent via End-to-End Reinforcement Learning (2026.findings-acl)
Copied to clipboard
Qianhao Yuan, Jie Lou, Zichao Li, Jiawei Chen, Yaojie Lu, Hongyu Lin, Le Sun, Debing Zhang, Xianpei Han
| Challenge: | Recent LLM-based search agents often concatenate the full interaction history into the context, producing long and noisy inputs and increasing compute cost and memory overhead. |
| Approach: | They propose an agent framework that maintains a compact memory during multi-turn interactions. |
| Outcome: | The proposed framework outperforms strong history-concatenation (ReAct-style) baselines on a range of public datasets while maintaining nearly constant token counts across multi-turn interactions. |
Critic-CoT: Boosting the Reasoning Abilities of Large Language Model via Chain-of-Thought Critic (2025.findings-acl)
Copied to clipboard
Xin Zheng, Jie Lou, Boxi Cao, Xueru Wen, Yuqiu Ji, Hongyu Lin, Yaojie Lu, Xianpei Han, Debing Zhang, Le Sun
| Challenge: | Existing approaches to improve the reasoning performance of large language models rely on intuitive instance-level feedback, which limits the reasoning capabilities. |
| Approach: | They propose a framework that pushes LLMs toward System-2-like critic capability by using a step-wise CoT reasoning paradigm and automatic construction of weak-supervision data without human annotation. |
| Outcome: | The proposed model significantly improves task-solving performance by filtering out invalid solutions or iterative refinement. |
On-Policy Self-Alignment with Fine-grained Knowledge Feedback for Hallucination Mitigation (2025.findings-acl)
Copied to clipboard
Xueru Wen, Jie Lou, Xinyu Lu, Yuqiu Ji, Xinyan Guan, Yaojie Lu, Hongyu Lin, Ben He, Xianpei Han, Debing Zhang, Le Sun
| Challenge: | Large language models exhibit behavior that deviates from the boundaries of their knowledge during response generation. |
| Approach: | They propose a framework that allows large language models to explore their knowledge boundaries and self-correct generation behavior through fine-grained feedback signals. |
| Outcome: | The proposed framework enables LLMs to explore their knowledge boundaries and self-correct generation behavior through fine-grained feedback signals. |
SedarEval: Automated Evaluation using Self-Adaptive Rubrics (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing evaluation paradigms rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process. |
| Approach: | They propose a new evaluation paradigm based on self-adaptive rubrics that mimic a human evaluator's analytical process. |
| Outcome: | The proposed evaluation paradigm achieves higher concordance rate with human graders than existing paradigms, including GPT-4. |
Untie the Knots: An Efficient Data Augmentation Strategy for Long-Context Pre-Training in Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have prioritized expanding the context window from which they can incorporate more information. |
| Approach: | They propose a data augmentation strategy to enable large language models to gain long-context capabilities without the need to modify existing data mixture. |
| Outcome: | The proposed model outperforms existing models on 20 billion tokens and achieves 75% and 84.5% accuracy on RULER at 128K context length. |
Cheems: A Practical Guidance for Building and Evaluating Chinese Reward Models from Scratch (2025.acl-long)
Copied to clipboard
Xueru Wen, Jie Lou, Zichao Li, Yaojie Lu, XingYu XingYu, Yuqiu Ji, Guohai Xu, Hongyu Lin, Ben He, Xianpei Han, Le Sun, Debing Zhang
| Challenge: | Existing Chinese resources are small in scale and limited to specific domains, making them insufficient for LLM post-training. |
| Approach: | They propose a Chinese-annotated reward model and a preference dataset to address this gap . they evaluate Chinese RMs on CheemsBench and construct an RM that captures human preferences . |
| Outcome: | The proposed RM achieves state-of-the-art performance on CheemsBench and CheeMePreference. |