Texonom
Texonom/
Engineering
Engineering/Data Engineering/Artificial Intelligence/Machine Learning/Reinforcement Learning/Reinforcement Learning Method/Reward model/
Verifiable Reward
Search

Verifiable Reward

Creator
Creator
Seonglae Cho
Created
Created
2025 Mar 21 11:54
Editor
Editor
Seonglae Cho
Edited
Edited
2025 Apr 30 23:58
Refs
Refs
Language Model RL

RLVR

Programmatic Reward, Verifiable Feedback

Externally provided feedback that can be objectively validated against specific criteria.

For Reasoning Data with
GRPO

Rule-based Verifiable Reward. LLM itself is policy network.
  • Accuracy rewards: Checking if the model’s final answer is correct
  • Format rewards: Incentivizing a structured chain-of-thought (enclose CoT tokens by <think> and </think>)
Verifiable Rewards
Length Reward
Accuracy reward
Format reward
 
 
 
but, it has low
Sample efficiency in larger samples and only finds better reasoning paths within its existing capacity which makes its total problem solving coverage smaller
limit-of-RLVR
LeapLabTHU • Updated 2025 May 1 7:48
www.arxiv.org
https://www.arxiv.org/pdf/2504.13837
arxiv.org
https://arxiv.org/pdf/2501.12948
From Zero to Reasoning Hero: How DeepSeek-R1 Leverages Reinforcement Learning to Master Complex Reasoning
A Blog post by Yihua Zhang on Hugging Face
From Zero to Reasoning Hero: How DeepSeek-R1 Leverages Reinforcement Learning to Master Complex Reasoning
https://huggingface.co/blog/NormalUhr/deepseek-r1-explained
From Zero to Reasoning Hero: How DeepSeek-R1 Leverages Reinforcement Learning to Master Complex Reasoning
arxiv.org
https://arxiv.org/pdf/2503.06639v1
 
 

Recommendations

Texonom
Texonom/
Engineering
Engineering/Data Engineering/Artificial Intelligence/Machine Learning/Reinforcement Learning/Reinforcement Learning Method/Reward model/
Verifiable Reward