Unverified50% confidenceFactExact time
Across the three post-training paradigms of RLVR, SFT, and on-policy distillation, zeroing out less than 0.05% of the weights in the pretrained model matches the performance of full fine-tuning at scale.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
Unverified指令微调不需要从头训练,通常仅需基础训练计算量的1%以下即可显著改变模型行为模式78% similarVerifiedFine-tuning typically requires only one percent or less of the original training cost to achieve domain performance close to that of a dedicated model.76% similarUnverifiedOn infilling tasks, Diffusion-LM outperforms continuous-relaxation-based methods such as COLD and DELOREAN on automatic evaluation metrics, and reaches the level of autoregressive models trained from scratch specifically for this purpose without requiring targeted training.76% similarUnverifiedIf actual training were restricted to the orthogonal complement of the null space (U_perp), performance could be expected to improve further.75% similarUnverifiedA contemporaneous study by O'Brien et al. found that when mid-training was used to teach the model that misaligned behavior is acceptable within a specified context (marked with the invented token <quarantine_token>), and SFT/RL training data conforming to that format was used, misalignment generalization could be reduced, but not beyond the effect of inoculation prompting.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1153748API
curl https://kongchang.com/api/v1/knowledge/claims/1153748MCP
get_claim(id=1153748)