Papers by Chonghua Liao
Exploring Forgetting in Large Language Model Pre-Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing research on task-level forgetting in LLMs has focused on pretraining . but, there is limited attention to finer-grained forgetting during training . |
| Approach: | They investigated the existence and measurement of forgetting in pre-training . they examined low-cost, straightforward methods to mitigate forgetting during the pre- training phase . |
| Outcome: | The proposed methods could be used to mitigate forgetting during the pre-training phase and offer insights into the dynamics of forgetting. |
FewNLU: Benchmarking State-of-the-Art Methods for Few-Shot Natural Language Understanding (2022.acl-long)
Copied to clipboard
Yanan Zheng, Jing Zhou, Yujie Qian, Ming Ding, Chonghua Liao, Li Jian, Ruslan Salakhutdinov, Jie Tang, Sebastian Ruder, Zhilin Yang
| Challenge: | Existing evaluation protocols for few-shot natural language understanding (NLU) tasks are inconsistent and hinder fair comparison and measuring progress. |
| Approach: | They propose an evaluation framework that improves previous evaluation procedures in three key aspects, i.e., test performance, dev-test correlation, and stability. |
| Outcome: | The proposed framework improves evaluation procedures in three key aspects, i.e., performance, dev-test correlation, and stability. |
TemplateRL: Structured Template-Guided Reinforcement Learning for LLM Reasoning (2026.findings-acl)
Copied to clipboard
Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao
| Challenge: | Existing RL methods rely on unstructured self-sampling to fit scalar rewards, resulting in inefficient rollouts. |
| Approach: | They propose a structured template-guided RL framework that augments policy optimization with explicit template guidance. |
| Outcome: | Experiments show that TemplateRL outperforms GRPO and GRPI by 99% on AIME and 41% on AMC with superior stability on weak models and remarkable cross-domain generalization. |
Beyond Examples: Towards Automated Thought-level In-Context Reasoning for Large Language Models (2026.acl-long)
Copied to clipboard
Jinyang Wu, Mingkuan Feng, Shuai Zhang, Feihu Che, Zhengqi Wen, Chonghua Liao, Ling Yang, Haoran Luo, Zheng Lian, Jianhua Tao
| Challenge: | In-context learning (ICL) struggles with complex reasoning due to superficial, example-level implicit imitation. |
| Approach: | They propose an automated method that shifts from surface-level examples to more guidance-oriented thought patterns. |
| Outcome: | The proposed method achieves 80.6% accuracy on MATH and 62.5% on AMC, surpassing GPT-4o’s 77.2% and 57.5% accuracy. |
MOA: Multi-Objective Alignment for Role-Playing Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work on role-playing agents relies on supervised fine-tuning or reinforcement learning with scalarized rewards, but these approaches do not address the coordination of multiple reward dimensions during optimization. |
| Approach: | They propose a reinforcement-learning framework that enables multi-dimensional, fine-grained rubric optimization for general RPAs. |
| Outcome: | Experiments on PersonaGym and RoleMRC show that MOA improves multi-dimensional role-playing performance over supervised and standard RL baselines. |
MemPO: Self-Memory Policy Optimization for Long-Horizon Agents (2026.findings-acl)
Copied to clipboard
Ruoran Li, Xinghua Zhang, Haiyang Yu, Shitong Duan, Xiang Li, Wenxin Xiang, Chonghua Liao, Xudong Guo, Yongbin Li, Jinli Suo
| Challenge: | Existing methods for long-horizon agents introduce the external memory module and look up the relevant information from the stored memory, which prevents the model from proactively managing its memory content and aligning with the agent’s overarching task objectives. |
| Approach: | They propose an algorithm which enables agents to autonomously manage their memory during interaction with environment and selectively retain crucial information. |
| Outcome: | Extensive experiments show that the proposed algorithm achieves absolute F1 score gains of 25.98 over the base model and 7.1 over the previous SOTA baseline while preserving task performance. |