Salvato in:
| Autori principali: | , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2503.14286 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915204107862016 |
|---|---|
| author | Roux, Nicolas Le Bellemare, Marc G. Lebensold, Jonathan Bergeron, Arnaud Greaves, Joshua Fréchette, Alex Pelletier, Carolyne Thibodeau-Laufer, Eric Toth, Sándor Work, Sam |
| author_facet | Roux, Nicolas Le Bellemare, Marc G. Lebensold, Jonathan Bergeron, Arnaud Greaves, Joshua Fréchette, Alex Pelletier, Carolyne Thibodeau-Laufer, Eric Toth, Sándor Work, Sam |
| contents | We propose a new algorithm for fine-tuning large language models using reinforcement learning. Tapered Off-Policy REINFORCE (TOPR) uses an asymmetric, tapered variant of importance sampling to speed up learning while maintaining stable learning dynamics, even without the use of KL regularization. TOPR can be applied in a fully offline fashion, allows the handling of positive and negative examples in a unified framework, and benefits from the implementational simplicity that is typical of Monte Carlo algorithms. We demonstrate the effectiveness of our approach with a series of experiments on the GSM8K and MATH reasoning benchmarks, finding performance gains for training both a model for solution generation and as a generative verifier. We show that properly leveraging positive and negative examples alike in the off-policy regime simultaneously increases test-time accuracy and training data efficiency, all the while avoiding the ``wasted inference'' that comes with discarding negative examples. We find that this advantage persists over multiple iterations of training and can be amplified by dataset curation techniques, enabling us to match 70B-parameter model performance with 8B language models. As a corollary to this work, we find that REINFORCE's baseline parameter plays an important and unexpected role in defining dataset composition in the presence of negative examples, and is consequently critical in driving off-policy performance. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_14286 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Tapered Off-Policy REINFORCE: Stable and efficient reinforcement learning for LLMs Roux, Nicolas Le Bellemare, Marc G. Lebensold, Jonathan Bergeron, Arnaud Greaves, Joshua Fréchette, Alex Pelletier, Carolyne Thibodeau-Laufer, Eric Toth, Sándor Work, Sam Machine Learning We propose a new algorithm for fine-tuning large language models using reinforcement learning. Tapered Off-Policy REINFORCE (TOPR) uses an asymmetric, tapered variant of importance sampling to speed up learning while maintaining stable learning dynamics, even without the use of KL regularization. TOPR can be applied in a fully offline fashion, allows the handling of positive and negative examples in a unified framework, and benefits from the implementational simplicity that is typical of Monte Carlo algorithms. We demonstrate the effectiveness of our approach with a series of experiments on the GSM8K and MATH reasoning benchmarks, finding performance gains for training both a model for solution generation and as a generative verifier. We show that properly leveraging positive and negative examples alike in the off-policy regime simultaneously increases test-time accuracy and training data efficiency, all the while avoiding the ``wasted inference'' that comes with discarding negative examples. We find that this advantage persists over multiple iterations of training and can be amplified by dataset curation techniques, enabling us to match 70B-parameter model performance with 8B language models. As a corollary to this work, we find that REINFORCE's baseline parameter plays an important and unexpected role in defining dataset composition in the presence of negative examples, and is consequently critical in driving off-policy performance. |
| title | Tapered Off-Policy REINFORCE: Stable and efficient reinforcement learning for LLMs |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2503.14286 |