Unverified50% confidenceFactExact time
A recently circulating paper proposes that reinforcement learning (RL) used for reasoning tasks changes only 1% to 3% of tokens, and that most of the gains can be reproduced with far less compute without using RL.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/17/2026
First Seen
Valid until: 1/15/2027
Sources
Related Entities
Related Claims
Unverified"RL changes only 1-3% of tokens" means comparing the model's output distribution before and after RL on the same batch of inputs, where only a very small number of positions show significant shifts in prediction.78% similarUnverifiedThe Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.70% similarUnverifiedRLAIF (Reinforcement Learning from AI Feedback) replaces large-scale human annotation with the model's own critical evaluation, reducing human annotation costs by one to two orders of magnitude compared to RLHF.67% similarUnverified少样本学习在提示词中附带少量输入-输出示例，通常1-5个62% similarUnverifiedApproximately 10–15% of papers receive a score increase after rebuttal, but the number that ultimately reverse a rejection is even smaller.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1054044API
curl https://kongchang.com/api/v1/knowledge/claims/1054044MCP
get_claim(id=1054044)