Papers by Moxin Li
Doc2SoarGraph: Discrete Reasoning over Visually-Rich Table-Text Documents via Semantic-Oriented Hierarchical Graphs (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing work on document visual question answering fails to capture the differences and correlations between elements of a document and associated questions. |
| Approach: | They propose a document-visual question-answering challenge that exploits element-level semantics and employs hierarchical Graph structures to capture differences and correlations between elements. |
| Outcome: | The proposed model surpasses the state-of-the-art method and large language model in terms of Exact Match (EM) metric, demonstrating exceptional effectiveness. |
Dual-Phase Accelerated Prompt Optimization (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for prompt optimization make light of the importance of high-quality initialization and the identification of effective directions. |
| Approach: | They propose a method which uses a meta-instruction to generate high-quality initial prompts and iteratively optimize them at the sentence level. |
| Outcome: | The proposed method achieves consistent accuracy gain over baselines with less than five optimization steps. |
Assistant-Guided Mitigation of Teacher Preference Bias in LLM-as-a-Judge (2025.findings-emnlp)
Copied to clipboard
| Challenge: | LLM-as-a-Judge uses large language models to evaluate the quality of LLM generated responses, but training proxy judge models using evaluation data generated by powerful teacher models introduces a critical yet previously overlooked issue: teacher preference bias. |
| Approach: | They propose a new setting that incorporates an additional assistant model, which is not biased toward the teacher model’s responses, to complement the training data. |
| Outcome: | The proposed model reduces teacher preference bias while maintaining strong performance across six evaluation benchmarks. |
Hypothetical Training for Robust Machine Reading Comprehension of Tabular Context (2023.findings-acl)
Copied to clipboard
| Challenge: | Counterfactual training is expensive because of the complexity of tabular data. |
| Approach: | They propose a hypothetical training framework that uses paired examples with different hypothetical questions to supervise the direction of model gradient towards the counterfactual answer change. |
| Outcome: | The proposed framework improves on tabular MRC datasets. |
Counterfactual Debating with Preset Stances for Hallucination Elimination of LLMs (2025.coling-main)
Copied to clipboard
| Challenge: | Existing solutions to alleviate hallucination have considered utilizing LLMs’ inherent reasoning abilities to alleviating hallucinism, such as self-correction and diverse sampling methods. |
| Approach: | They propose a counterfactual multi-agent debate framework that predetermines LLMs' stances to override their inherent biases for answer inspection. |
| Outcome: | Extensive experiments on four datasets of three tasks demonstrate the superiority of the proposed framework over existing methods. |
Knowledge Boundary of Large Language Models: A Survey (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) store vast amount of knowledge in their parameters, but they still have limitations in the memorization and utilization of certain knowledge. |
| Approach: | They propose a comprehensive definition of the LLM knowledge boundary and introduce a formalized taxonomy categorizing knowledge into four distinct types. |
| Outcome: | The proposed definition of the LLM knowledge boundary and taxonomy categorizes knowledge into four distinct types . aims to offer a comprehensive overview, facilitate access to key issues, and inspire further advancements in LLM research. |
Learning to Imagine: Integrating Counterfactual Thinking in Neural Discrete Reasoning (2022.acl-long)
Copied to clipboard
| Challenge: | Existing NDR models suffer from large performance drop on hypothetical questions, e.g., “what the annualized rate of return would be if the revenue in 2020 was doubled”. |
| Approach: | They propose a learning to imagine module which can be seamlessly incorporated into NDR models to perform the imagination of unseen counterfactual. |
| Outcome: | The proposed model can perform the imagination of unseen counterfactuals on hypothetical questions. |
HellaSwag-Pro: A Large-Scale Bilingual Benchmark for Evaluating the Robustness of LLMs in Commonsense Reasoning (2025.findings-acl)
Copied to clipboard
Xiaoyuan Li, Moxin Li, Rui Men, Yichang Zhang, Keqin Bao, Wenjie Wang, Fuli Feng, Dayiheng Liu, Junyang Lin
| Challenge: | Existing studies show that large language models are robust in commonsense reasoning . however, some variations in questions can lead to incorrect responses . |
| Approach: | They propose a large-scale bilingual benchmark consisting of 11,200 cases . they conduct extensive experiments on 41 representative LLMs . |
| Outcome: | The proposed benchmark systematically evaluates the robustness of large language models in commonsense reasoning. |
MTR-Bench: A Comprehensive Benchmark for Multi-Turn Reasoning Evaluation (2026.acl-long)
Copied to clipboard
Xiaoyuan Li, Keqin Bao, Yubo Ma, Moxin Li, Wenjie Wang, Rui Men, Yichang Zhang, Fuli Feng, Dayiheng Liu
| Challenge: | Recent advances in Large Language Models (LLMs) have shown promising results in complex reasoning tasks. |
| Approach: | They propose to use a multi-turn reasoning evaluation framework to cover multi-turn interactions with the environments of large language models. |
| Outcome: | The proposed framework covers diverse reasoning capabilities, fine-grained difficulty granularity, and necessitates multi-turn interactions with the environments. |
Evaluating Mathematical Reasoning of Large Language Models: A Focus on Error Identification and Correction (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluations focus on problem-solving from examiner perspective, overlooking a dual perspective of examiner regarding error identification and correction. |
| Approach: | They propose to use an annotated dataset to evaluate large language models from the examiner perspective and to use diverse prompts to evaluate eleven representative LLMs. |
| Outcome: | The proposed model outperforms all models while LLaMA-2-7B has comparable abilities to closed-source models GPT-3.5 and Gemini Pro. |
Don’t Just Say “I don’t know”! Self-aligning Large Language Models for Responding to Unknown Questions with Explanations (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies investigate ways to refuse to answer unknown questions . Large Language Models (LLMs) display a significant level of overconfidence when answering questions that they are aware of. |
| Approach: | They propose a self-alignment method to utilize Large Language Models to enhance its response-ability to unknown questions. |
| Outcome: | The proposed method is superior to baseline methods on four types of unknown questions. |
Self-Improvement Towards Pareto Optimality: Mitigating Preference Conflicts in Multi-Objective Alignment (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to optimize large language models with human preferences suffer from preference conflicts in the data. |
| Approach: | They propose to construct Pareto-optimal responses to resolve preference conflicts by using a self-improving DPO framework that enables LLMs to self-generate and select Paret-optimized responses. |
| Outcome: | The proposed framework achieves superior Pareto Front performance over baselines on two datasets. |
Robust Prompt Optimization for Large Language Models Against Distribution Shifts (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing research has explored automatic prompt optimization methods to eliminate manual effort in identifying effective prompts for a given task. |
| Approach: | They propose a framework for prompt optimization that can be generalized to an unlabeled target group. |
| Outcome: | The proposed framework improves on target group and source group while generalizing to unlabeled target group. |
Think Twice Before Trusting: Self-Detection for Large Language Models through Comprehensive Answer Reflection (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to self-detection only retrospectively evaluate LLM-generated answers, leading to over-trust in incorrectly generated answers. |
| Approach: | They propose a self-detection paradigm that considers the comprehensive answer space beyond LLM-generated answers to mitigate the over-trust in LLM generated incorrect answers. |
| Outcome: | The proposed framework can be integrated with existing approaches for superior self-detection. |
TROJail: Trajectory-Level Optimization for Multi-Turn Large Language Model Jailbreaks with Process Rewards (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to training multi-turn attackers to probe model safety vulnerabilities rely on turn-level optimization, which is insufficient for learning long-term attack strategies. |
| Approach: | They propose a multi-turn reinforcement learning problem that optimizes the harmfulness of the final-turn response as the outcome reward. |
| Outcome: | The proposed approach improves attack success rates across multiple models and benchmarks, highlighting the effectiveness of the proposed approach. |