ReDit: Reward Dithering for Improved LLM Policy Optimization

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Wei, Chenxing, Yu, Jiarui, He, Ying Tiffany, Dong, Hande, Shu, Yao, Yu, Fei
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912667257536512
author Wei, Chenxing
Yu, Jiarui
He, Ying Tiffany
Dong, Hande
Shu, Yao
Yu, Fei
author_facet Wei, Chenxing
Yu, Jiarui
He, Ying Tiffany
Dong, Hande
Shu, Yao
Yu, Fei
contents DeepSeek-R1 has successfully enhanced Large Language Model (LLM) reasoning capabilities through its rule-based reward system. While it's a ''perfect'' reward system that effectively mitigates reward hacking, such reward functions are often discrete. Our experimental observations suggest that discrete rewards can lead to gradient anomaly, unstable optimization, and slow convergence. To address this issue, we propose ReDit (Reward Dithering), a method that dithers the discrete reward signal by adding simple random noise. With this perturbed reward, exploratory gradients are continuously provided throughout the learning process, enabling smoother gradient updates and accelerating convergence. The injected noise also introduces stochasticity into flat reward regions, encouraging the model to explore novel policies and escape local optima. Experiments across diverse tasks demonstrate the effectiveness and efficiency of ReDit. On average, ReDit achieves performance comparable to vanilla GRPO with only approximately 10% the training steps, and furthermore, still exhibits a 4% performance improvement over vanilla GRPO when trained for a similar duration. Visualizations confirm significant mitigation of gradient issues with ReDit. Moreover, theoretical analyses are provided to further validate these advantages.
format Preprint
id arxiv_https___arxiv_org_abs_2506_18631
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ReDit: Reward Dithering for Improved LLM Policy Optimization
Wei, Chenxing
Yu, Jiarui
He, Ying Tiffany
Dong, Hande
Shu, Yao
Yu, Fei
Machine Learning
Artificial Intelligence
Computation and Language
DeepSeek-R1 has successfully enhanced Large Language Model (LLM) reasoning capabilities through its rule-based reward system. While it's a ''perfect'' reward system that effectively mitigates reward hacking, such reward functions are often discrete. Our experimental observations suggest that discrete rewards can lead to gradient anomaly, unstable optimization, and slow convergence. To address this issue, we propose ReDit (Reward Dithering), a method that dithers the discrete reward signal by adding simple random noise. With this perturbed reward, exploratory gradients are continuously provided throughout the learning process, enabling smoother gradient updates and accelerating convergence. The injected noise also introduces stochasticity into flat reward regions, encouraging the model to explore novel policies and escape local optima. Experiments across diverse tasks demonstrate the effectiveness and efficiency of ReDit. On average, ReDit achieves performance comparable to vanilla GRPO with only approximately 10% the training steps, and furthermore, still exhibits a 4% performance improvement over vanilla GRPO when trained for a similar duration. Visualizations confirm significant mitigation of gradient issues with ReDit. Moreover, theoretical analyses are provided to further validate these advantages.
title ReDit: Reward Dithering for Improved LLM Policy Optimization
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2506.18631