Papers by Keming Lu
LLM Critics Help Catch Bugs in Mathematics: Towards a Better Mathematical Verifier with Natural Language Feedback (2025.findings-acl)
Copied to clipboard
Bofei Gao, Zefan Cai, Runxin Xu, Peiyi Wang, Ce Zheng, Runji Lin, Keming Lu, Dayiheng Liu, Chang Zhou, Wen Xiao, Tianyu Liu, Baobao Chang
| Challenge: | Existing mathematical verifiers are trained with binary classification labels, which are not informative enough for the model to accurately assess the solutions. |
| Approach: | They propose a natural language feedback-enhanced verifier that can validate the correctness of response generated by policy models by constructing automatically generated training data and a two-stage training paradigm. |
| Outcome: | The proposed verifier significantly improves in verification and reinforcement learning and alleviates data-demanding problems of the reward model. |
ProcessBench: Identifying Process Errors in Mathematical Reasoning (2025.acl-long)
Copied to clipboard
Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, Junyang Lin
| Challenge: | Existing models fail to generalize to more challenging math problems, authors say . existing benchmarks related to assessing language models' reasoning process are limited . |
| Approach: | They propose a tool to measure language models' ability to identify erroneous steps in reasoning . they use two types of models: process reward models and critic models . |
| Outcome: | The proposed model outperforms existing models in evaluating language models' reasoning process . the best open-source model has demonstrated the critique capability competitive with the proprietary model . |
Multi-hop Evidence Retrieval for Cross-document Relation Extraction (2023.findings-acl)
Copied to clipboard
| Challenge: | Relation Extraction (RE) is a task that seeks to identify the relation of entities described according to some context. |
| Approach: | They propose a multi-hop evidence retrieval method based on evidence path mining and ranking to support cross-document relation extraction. |
| Outcome: | The proposed method acquires cross-document evidence and boosts performance in both closed and open environments. |
Speculative Contrastive Decoding (2024.acl-short)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit exceptional performance in language tasks, yet their auto-regressive inference is limited due to high computational requirements and is sub-optimal due to the exposure bias. |
| Approach: | They propose a decoding approach that leverages predictions from smaller language models to achieve both decoding acceleration and quality improvement. |
| Outcome: | The proposed method achieves both decoding acceleration and quality improvement on four diverse language tasks. |
MuggleMath: Assessing the Impact of Query and Response Augmentation on Math Reasoning (2024.acl-long)
Copied to clipboard
Chengpeng Li, Zheng Yuan, Hongyi Yuan, Guanting Dong, Keming Lu, Jiancan Wu, Chuanqi Tan, Xiang Wang, Chang Zhou
| Challenge: | In math reasoning with large language models, fine-tuning data augmentation by query evolution and diverse reasoning paths is empirically verified effective. |
| Approach: | They propose to fine-tune data augmentation by query evolution and diverse reasoning paths. |
| Outcome: | The proposed model achieves new state-of-the-art on GSM8K and MATH. |
Summarization as Indirect Supervision for Relation Extraction (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Relation extraction (RE) models rely on training data with expensive annotations . et al., 2018; Zhao e.t al, 2018) . |
| Approach: | They propose a method that converts RE into a summarization formulation by using constraint decoding techniques. |
| Outcome: | The proposed method improves relation extraction models with high-resource and high-contrast inferences. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |
Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models (2025.findings-acl)
Copied to clipboard
Hao Xiang, Bowen Yu, Hongyu Lin, Keming Lu, Yaojie Lu, Xianpei Han, Ben He, Le Sun, Jingren Zhou, Junyang Lin
| Challenge: | Prior research focused on developing data generation methods, while insufficient attention has been paid to quality control mechanisms and often produces inaccurate and unhelpful data. |
| Approach: | They propose an algorithm that automatically generates high-quality preference data, eliminating manual annotation requirements. |
| Outcome: | The proposed algorithm outperforms baselines in human preference alignment and reward optimization. |
How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Composition (2024.acl-long)
Copied to clipboard
Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, Jingren Zhou
| Challenge: | supervised fine-tuning (SFT) is a technique used to enhance multiple abilities in large language models. |
| Approach: | They propose to study the interplay of data composition between mathematical reasoning, code generation, and general human-aligning abilities during supervised fine-tuning. |
| Outcome: | The proposed model improves math reasoning and code generation with increasing data amount . the proposed model size and SFT strategies can be used to learn multiple skills with different scaling patterns. |
On the Editability of Delta Parameters in Post-Trained Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Several studies have explored delta parameter properties via pruning, quantization, low-rank approximation, and extrapolation, but what properties of delta parameters are essential for maintaining performance? |
| Approach: | They propose to examine delta parameter properties along magnitude and sign . they propose to use a loss-based local surrogate analysis to examine editing effects . |
| Outcome: | The proposed analysis shows that delta parameters can be edited while maintaining performance. |
Predicting Rewards Alongside Tokens: Non-disruptive Parameter Insertion for Efficient Inference Intervention in Large Language Model (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to fine tune LLMs produce unsafe responses and unreliable reasoning, but this solution introduces substantial time and space overhead due to the separate models required. |
| Approach: | They propose to insert extra parameters into transformer architecture to predict calibration signals along with original LLM output. |
| Outcome: | The proposed model reduces time and space costs while enabling seamless online deployment. |
Large Language Models are Superpositions of All Characters: Attaining Arbitrary Role-play via Self-Alignment (2024.acl-long)
Copied to clipboard
| Challenge: | Existing work cheaply emulates LLMs, allowing users to create profiles for their preferred characters. |
| Approach: | They propose a self-alignment method that encourages an instruction-following LLM to simulate role-play dialogues as a variant of reading comprehension. |
| Outcome: | The proposed model outperforms open-source role-play benchmarks and the roleplay subset of MT-Bench in multiple parameters. |
Routing to the Expert: Efficient Reward-guided Ensemble of Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing ensemble methods for Large Language Models focus on reward model ranking of outputs, leading to significant computation overhead. |
| Approach: | They propose a reward-guided routing method distilling rewards on training queries to train a routing function. |
| Outcome: | The proposed method outperforms the best single model and ranks first on 44% of tasks. |
PIVOINE: Instruction Tuning for Open-world Entity Profiling (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for information extraction focus on a closed-world setting, but PIVOINE is a promising solution to tackle the open-world problem of entity profiling. |
| Approach: | They propose to develop an LLM that performs Open-world Entity Profiling with instruction tuning to extract desirable entity profiles . they construct INSTRUCTOPENWIKI, a substantial instruction-tuning dataset for Open-World Entity Profiles . |
| Outcome: | The proposed model outperforms existing methods and ChatGPT-based baselines on unseen and out-of-ontology cases. |