Evaluation-Conditioned Training: A New Post-Training Framework for LLMs

A new arXiv paper (2608.10209) introduces Evaluation-Conditioned Training (ECT), a post-training framework for Large Language Models (LLMs) that conditions each training sample on the fidelity of feedback using natural language. This approach aims to improve performance under imperfect feedback and can be integrated with existing algorithms like SFT and PPO. The paper provides a conceptual framework and discusses its potential to address reward mis-specification, a key limitation in current post-training methods where human annotators and automated reward functions fail to capture desired feedback accurately. The research is significant for AI alignment, offering a method to teach models to generalize to stronger oversight regimes.