reward-mis-specification

标签

Cards List
#reward-mis-specification

Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimes

arXiv cs.AI ↗ · 2026-08-12 缓存

The paper introduces Evaluation-Conditioned Training (ECT), a post-training framework that conditions on natural language descriptions of feedback fidelity to help LLMs generalize to stronger oversight, with proof-of-concept experiments showing improved even-handedness and reduced sycophancy.

0 人收藏 0 人点赞
← 返回首页

提交意见反馈