Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning

Zhiheng Xi, Wenxiang Chen, Boyang Hong, Senjie Jin, Rui Zheng, Wei He, Yiwen Ding, Shichun Liu, Xin Guo, Junzhe Wang, Honglin Guo, Wei Shen, Xiaoran Fan, Yuhao Zhou, Shihan Dou, Xiao Wang, Xinbo Zhang, Peng Sun, Tao Gui, Qi Zhang, Xuanjing Huang
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:54030-54048, 2024.

Abstract

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. Notably, in program-based reasoning, 7B-scale models perform comparably to larger models or closed-source models with our R$^3$.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-xi24a, title = {Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning}, author = {Xi, Zhiheng and Chen, Wenxiang and Hong, Boyang and Jin, Senjie and Zheng, Rui and He, Wei and Ding, Yiwen and Liu, Shichun and Guo, Xin and Wang, Junzhe and Guo, Honglin and Shen, Wei and Fan, Xiaoran and Zhou, Yuhao and Dou, Shihan and Wang, Xiao and Zhang, Xinbo and Sun, Peng and Gui, Tao and Zhang, Qi and Huang, Xuanjing}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {54030--54048}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/xi24a/xi24a.pdf}, url = {https://proceedings.mlr.press/v235/xi24a.html}, abstract = {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. Notably, in program-based reasoning, 7B-scale models perform comparably to larger models or closed-source models with our R$^3$.} }
Endnote
%0 Conference Paper %T Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning %A Zhiheng Xi %A Wenxiang Chen %A Boyang Hong %A Senjie Jin %A Rui Zheng %A Wei He %A Yiwen Ding %A Shichun Liu %A Xin Guo %A Junzhe Wang %A Honglin Guo %A Wei Shen %A Xiaoran Fan %A Yuhao Zhou %A Shihan Dou %A Xiao Wang %A Xinbo Zhang %A Peng Sun %A Tao Gui %A Qi Zhang %A Xuanjing Huang %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-xi24a %I PMLR %P 54030--54048 %U https://proceedings.mlr.press/v235/xi24a.html %V 235 %X 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. Notably, in program-based reasoning, 7B-scale models perform comparably to larger models or closed-source models with our R$^3$.
APA
Xi, Z., Chen, W., Hong, B., Jin, S., Zheng, R., He, W., Ding, Y., Liu, S., Guo, X., Wang, J., Guo, H., Shen, W., Fan, X., Zhou, Y., Dou, S., Wang, X., Zhang, X., Sun, P., Gui, T., Zhang, Q. & Huang, X.. (2024). Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:54030-54048 Available from https://proceedings.mlr.press/v235/xi24a.html.

Related Material