Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2024
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866916162925756416 |
|---|---|
| author | Xi, Zhiheng Chen, Wenxiang Hong, Boyang Jin, Senjie Zheng, Rui He, Wei Ding, Yiwen Liu, Shichun Guo, Xin Wang, Junzhe Guo, Honglin Shen, Wei Fan, Xiaoran Zhou, Yuhao Dou, Shihan Wang, Xiao Zhang, Xinbo Sun, Peng Gui, Tao Zhang, Qi Huang, Xuanjing |
| author_facet | Xi, Zhiheng Chen, Wenxiang Hong, Boyang Jin, Senjie Zheng, Rui He, Wei Ding, Yiwen Liu, Shichun Guo, Xin Wang, Junzhe Guo, Honglin Shen, Wei Fan, Xiaoran Zhou, Yuhao Dou, Shihan Wang, Xiao Zhang, Xinbo Sun, Peng Gui, Tao Zhang, Qi Huang, Xuanjing |
| contents | In this paper, we propose R$^3$: Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models. The core challenge in applying RL to complex reasoning is to identify a sequence of actions that result in positive rewards and provide appropriate supervision for optimization. Outcome supervision provides sparse rewards for final results without identifying error locations, whereas process supervision offers step-wise rewards but requires extensive manual annotation. R$^3$ overcomes these limitations by learning from correct demonstrations. Specifically, R$^3$ progressively slides the start state of reasoning from a demonstration's end to its beginning, facilitating easier model exploration at all stages. Thus, R$^3$ establishes a step-wise curriculum, allowing outcome supervision to offer step-level signals and precisely pinpoint errors. Using Llama2-7B, our method surpasses RL baseline on eight reasoning tasks by $4.1$ points on average. Notebaly, in program-based reasoning on GSM8K, it exceeds the baseline by $4.2$ points across three backbone models, and without any extra data, Codellama-7B + R$^3$ performs comparable to larger models or closed-source models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2402_05808 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning Xi, Zhiheng Chen, Wenxiang Hong, Boyang Jin, Senjie Zheng, Rui He, Wei Ding, Yiwen Liu, Shichun Guo, Xin Wang, Junzhe Guo, Honglin Shen, Wei Fan, Xiaoran Zhou, Yuhao Dou, Shihan Wang, Xiao Zhang, Xinbo Sun, Peng Gui, Tao Zhang, Qi Huang, Xuanjing Artificial Intelligence Computation and Language Machine Learning In this paper, we propose R$^3$: Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models. The core challenge in applying RL to complex reasoning is to identify a sequence of actions that result in positive rewards and provide appropriate supervision for optimization. Outcome supervision provides sparse rewards for final results without identifying error locations, whereas process supervision offers step-wise rewards but requires extensive manual annotation. R$^3$ overcomes these limitations by learning from correct demonstrations. Specifically, R$^3$ progressively slides the start state of reasoning from a demonstration's end to its beginning, facilitating easier model exploration at all stages. Thus, R$^3$ establishes a step-wise curriculum, allowing outcome supervision to offer step-level signals and precisely pinpoint errors. Using Llama2-7B, our method surpasses RL baseline on eight reasoning tasks by $4.1$ points on average. Notebaly, in program-based reasoning on GSM8K, it exceeds the baseline by $4.2$ points across three backbone models, and without any extra data, Codellama-7B + R$^3$ performs comparable to larger models or closed-source models. |
| title | Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning |
| topic | Artificial Intelligence Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2402.05808 |