P1: Mastering Physics Olympiads with Reinforcement Learning

Jiacheng Chen*, Qianjia Cheng*, Fangchen Yu*, Haiyuan Wan, Yuchen Zhang, Shenghe Zheng, Junchi Yao, Qingyang Zhang, Haonan He, Yun Luo, Yufeng Zhao, Futing Wang, Li Sheng, Chengxing Xie, Yuxin Zuo, Yizhuo Li, Wenxauan Zeng, Yulun Wu, Rui Huang, Dongzhan Zhou, Kai Chen, Yu Qiao, Lei Bai, Yu Cheng, Ning Ding, Bowen Zhou, Peng Ye, Ganqu Cui.

• More details can be found in our blog.

The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models

Ganqu Cui*, Yuchen Zhang*, Jiacheng Chen*, Lifan Yuan, Zhi Wang, Yuxin Zuo, Haozhan Li, Yuchen Fan, Huayu Chen, Weize Chen, Zhiyuan Liu, Hao Peng, Lei Bai, Wanli Ouyang, Yu Cheng, Bowen Zhou, Ning Ding.

• We conducted empirical and theoretical analysis at “entropy collapse” phenomena, and proposed new way of entropy control.

Reasoning in Reasoning: A Hierarchical Framework for Neural Theorem Proving (NeurIPS 2024 Workshop MATH-AI)

Ziyu Ye, Jiacheng Chen, Jonathan Light, Yifei Wang, Jiankai Sun, Mac Schwager, Philip Torr, Guohao Li, Yuxin Chen, Kaiyu Yang, Yisong Yue, Ziniu Hu.

SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024)

Jiacheng Chen*, Zeyuan Ma*, Hongshu Guo, Yining Ma, Jie Zhang, Yue-Jiao Gong.

Project

• Unlike previous methods incrementally auto-configuring some existing black-box algortithms, SYMBOL directly generate stepwise update rule in the form of symbolic eqution to achieve more flexible and interpretable optimization behaviour.

LLaMoCo: Instruction Tuning of Large Language Models for Optimization Code Generation

Zeyuan Ma, Hongshu Guo, Jiacheng Chen, Guojun Peng, Zhiguang Cao, Yining Ma, Yue-Jiao Gong.

• We propose to fine-tune language model to generate executable code that can be used for optimization tasks. We proposed a dataset that containing diversed optimization problems and corresponding algorithm in this paper, also leverage some tricks during training process and finally provided a fine-tuned LM for optimization tasks.

MetaBox: A Benchmark Platform for Meta-Black-Box Optimization with Reinforcement Learning (NeurIPS 2023 Oral)

Zeyuan Ma, Hongshu Guo, Jiacheng Chen, Zhenrui Li, Guojun Peng, Yue-Jiao Gong, Yining Ma, Zhiguang Cao.

Project

• We released a benchmark platform for Meta-Black-Box Optimization named MetaBox. We integrate three different testsuits, about 20 baselines including traditional black-box methods and Meta-Black-Box methods, and new evaluation metrics tailored for Meta-Black-Box optimization. The codebase can be found here.

Neural Exploratory Landscape Analysis (ICLR 2025)

Zeyuan Ma, Jiacheng Chen, Hongshu Guo, Yue-Jiao Gong.

• We developed an Neural-Network based lanscape analyser to replace the feature-extracting parts in Meta-Black-Box works which is usually manually designed. To ensure the generalization ability of the NeurELA, we let it operate in Multi-task setting and use neuroevolution to train it.

Auto-configuring Exploration-Exploitation Tradeoff in Evolutionary Computation via Deep Reinforcement Learning (GECCO 2024)

Zeyuan Ma*, Jiacheng Chen*, Hongshu Guo, Yining Ma, Yue-Jiao Gong.

• We explore about how to make a trade-off between exploration and exploitation in Black-Box optimization through learn-based method. In this work, we carefully designed a framework which is based on transfromer-based model and leverage exploration-exploitation related feature tailored for black-box optimization scenario to resolve this problem.