Unverified50% confidenceBenchmarkExact time
SJR comprehensively outperformed multiple training methods including end-to-end SFT, STaR/RFT, and RLFT in experiments.
1
Sources
50%
Confidence
Long-term
Relevance
9/22/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBreakthrough directions explored by various laboratories in the post-training stage include RLHF, chain-of-thought distillation, and enhanced tool-calling capabilities.70% 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66% similarVerified从预训练检查点到可商用产品需经历监督式指令微调（SFT）和基于人类反馈的强化学习（RLHF）两个后训练阶段66% similarVerifiedThe three-stage alignment training paradigm of SFT–reward model–RLHF was first systematically proposed by OpenAI in the InstructGPT paper.65% similarUnverifiedIris adopts an SFT-RL climbing strategy, alternating between supervised fine-tuning and reinforcement learning, where high-quality behavioral trajectories discovered during the RL phase are fed back to the SFT phase as new supervision signals.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/940018API
curl https://kongchang.com/api/v1/knowledge/claims/940018MCP
get_claim(id=940018)