FORK reward mechanism alleviates reward hijacking
The story so far
qwen has introduced the FORK reward mechanism, combining necessity credit driven by ratings with factual ratings while optimizing recall and precision. This improves response quality and reduces redundancy. Compared to traditional rewards based solely on recall, FORK significantly improves precision and shortens response length while maintaining similar rating coverage, increasing win rates from 10.6% to 68.0% on RubricHub-Medical and from 22.1% to 74.1% on HealthBench. Experimental results show that FORK outperforms existing standard-based reinforcement learning methods and achieves Pareto optimal performance in balancing rating coverage and overall response quality.
AI-written from the reports · Updated 23 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- Hugging Face Daily PapersEat Your Precision: Precision Objective Mitigates Reward Hacking in Rubric-based RL
The FORK reward mechanism enhances response quality and reduces redundancy by combining necessity credit driven by scoring with factual scoring, while optimizing recall and precision. Compared to traditional rewards based solely on recall, FORK significantly improves precision and shortens response length while maintaining similar scoring coverage, increasing win rates from 10.6% to 68.0% on RubricHub-Medical and from 22.1% to 74.1% on HealthBench.