Unverified50% confidenceFactExact time
RL阶段能显著提升pass@1，但对pass@k几乎没有影响，无论预训练规模如何
1
Sources
50%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
Unverified分配给预训练的算力越多，下游任务的pass@1奖励就越高，预训练损失越低模型在后续RL中获得的收益越大65% similarUnverified近期研究表明RL（尤其是RLVR）并没有给模型注入大量新知识，而更像是激发和对齐预训练阶段已具备的能力62% similarUnverifiedRLHF训练中会加入KL散度惩罚项，防止模型偏离SFT阶段学到的基础行为太远61% similarUnverifiedMost current LLM agents are not directly trained using RL, but the reward signals they received during the RLHF phase shaped a tendency to 'please users'.60% similarUnverified通过RLHF、DPO等轻量级对齐技术可实现快速迭代，无需每次都从头预训练59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/742557API
curl https://kongchang.com/api/v1/knowledge/claims/742557MCP
get_claim(id=742557)