Unverified50% confidenceFactExact time
"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.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA 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.78% similarUnverifiedThe Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.69% similarUnverifiedThe mean-field likelihood estimation relied upon by existing RL methods for diffusion models is biased and has high variance.66% similarUnverifiedThere is no stable mapping between a model's self-evaluation score and its actual accuracy; rephrase the same prompt and the score may drift substantially.65% similarUnverifiedFacts located in the middle of a context are used by the model at a noticeably lower probability than those at the beginning or end — a U-shaped recall curve.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1054045API
curl https://kongchang.com/api/v1/knowledge/claims/1054045MCP
get_claim(id=1054045)