A recent study proposes a new training framework called Evaluation-Conditioned Training (ECT). This method aims to improve the performance of Large Language Models (LLMs) under imperfect feedback signals. The basic idea of ECT is to use natural language to condition each training sample on the fidelity of the feedback signal and then elicit the desired behavior by conditioning the LLM on a high-fidelity monitor in deployment. This approach can be used as an add-on to existing algorithms such as SFT and PPO. The researchers first provide a conceptual framework for ECT and discuss its potential advantages. Then, they motivate ECT in the context of the eliciting latent knowledge (ELK) problem. Finally, the researchers evaluate ECT on two proof-of-concept experiments: increasing even-handedness in news article generation and reducing sycophancy on an arithmetic task. In each setting, the researchers use imperfect feedback, rewarding bias and agreement with the user, respectively. The results show that ECT improves the targeted behavior relative to direct training. Blogger's Review: Evaluation-Conditioned Training provides a new perspective on the training of Large Language Models, especially when faced with imperfect feedback signals. The potential advantages of this approach lie in its ability to improve the model's generalizability and robustness, making it worth further research and exploration.