Reinforcement learning with 40% flipped rewards, or with purely phrase-based reasoning rewards, lifts Qwen-2.5-7B's MATH-500 accuracy from 5% to over 70%.
The accuracy paradox in RLHF: When better reward models don‘t yield better language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
The Climb Carves Wisdom Deeper Than the Summit: On the Noisy Rewards in Learning to Reason
Reinforcement learning with 40% flipped rewards, or with purely phrase-based reasoning rewards, lifts Qwen-2.5-7B's MATH-500 accuracy from 5% to over 70%.