Papers by Zhouhao Sun
Towards Generalizable and Faithful Logic Reasoning over Natural Language via Resolution Refutation (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have achieved significant performance in various natural language reasoning tasks, but struggle with performing first-order logic reasoning over formal logical theories expressed in natural language. |
| Approach: | They propose a framework which introduces the paradigm of resolution refutation to solve first-order logic reasoning problems by extending reasoning rules and employing the principle of proof by contradiction. |
| Outcome: | The proposed framework outperforms existing models while maintaining performance in simple scenarios. |
Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration (2026.acl-long)
Copied to clipboard
Yang Zhao, Yangou Ouyang, Xiao Ding, Hepeng Wang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu
| Challenge: | Existing data arbitration strategies for large language model training rely on surface-level heuristics that fail to diagnose intrinsic learning needs. |
| Approach: | They propose a framework that arbitrates data based on its degree of cognitive conflict with the model's existing knowledge. |
| Outcome: | Extensive experiments on WebShop and ALFWorld show that PRISM outperforms state-of-the-art hybrid methods while reducing computational costs by up to 3.22 . |
Beyond Similarity: A Gradient-based Graph Method for Instruction Tuning Data Selection (2025.acl-long)
Copied to clipboard
Yang Zhao, Li Du, Xiao Ding, Yangou Ouyang, Hepeng Wang, Kai Xiong, Jinglong Gao, Zhouhao Sun, Dongliang Xu, Qing Yang, Dongchen Li, Bing Qin, Ting Liu
| Challenge: | Existing methods for selecting training data from general datasets fail to account for the joint distribution of instructions, resulting in inefficient learning and suboptimal knowledge transfer. |
| Approach: | They propose a method that constructs a mixed gradient-based instruction graph to capture the joint distribution and interdependencies among instructions. |
| Outcome: | The proposed method outperforms existing methods on domain adaptation tasks and in complex, data-scarce scenarios. |
MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization (2026.acl-long)
Copied to clipboard
Yang Zhao, Hepeng Wang, Xiao Ding, Yangou Ouyang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu
| Challenge: | Group-Relative Policy Optimization (GRPO) has emerged as an efficient paradigm for aligning Large Language Models (LLMs), but its efficacy is confined to domains with verifiable ground truths. |
| Approach: | They propose a meta-cognitive orchestration layer that treats reward scalarization as a dynamic latent policy, leveraging the model’s terminal hidden states as 'a semantic bottleneck' . Across seven benchmarks, MAESTRO consistently outperforms single-reward and static multi-objective baselines while preserving the efficiency advantages of GRPO. |
| Outcome: | The proposed model outperforms single-reward and static multi-objective baselines while preserving efficiency advantages. |
Deciphering the Impact of Pretraining Data on Large Language Models through Machine Unlearning (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have suggested that the composition of the pretraining corpus exerts a significant impact upon the performance of LLMs. |
| Approach: | They analyze the impact of 48 datasets from 5 major categories of pretraining data of Large Language Models and measure their impacts on LLMs using benchmarks about nine major categories. |
| Outcome: | The proposed analysis provides insights into the organization of data to support more efficient pretraining of Large Language Models. |
Towards Stable Natural Language Understanding via Information Entropy Guided Debiasing (2023.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to debiase Natural Language Understanding models use dataset biases instead of learning the intended task. |
| Approach: | They propose a debiasing framework that detects and purifies dataset biases using information entropy. |
| Outcome: | The proposed framework improves the stability of performance on out-of-distribution datasets for a set of widely adopted NLU models. |
Large Language Models Are Still Misled by Simple Bias Ensembles (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for large language models are constrained to datasets where each sample is manually injected with only one type of bias. |
| Approach: | They propose a multi-bias benchmark where each sample contains multiple types of biases. |
| Outcome: | The proposed benchmark shows that existing LLMs and debiasing methods perform poorly on this benchmark, highlighting the challenge of eliminating compounded biases. |