Rubrics to Tokens: Bridging Response-level Rubrics and Token-level Rewards in Instruction Following Tasks

1. 1Identify response-level rubrics for instruction following tasks
2. 2Map response-level scores to token-level rewards
3. 3Implement Rubrics to Tokens framework to bridge response-level and token-level rewards
4. 4Evaluate model performance using the proposed framework

ML researchers and engineers working on large language models can benefit from this framework to improve model performance and alignment with complex tasks

💡 The proposed framework addresses reward sparsity and ambiguity problems in rubric-based reinforcement learning

🤖 Bridging response-level rubrics and token-level rewards for instruction following tasks with Rubrics to Tokens framework