Reflection-Reinforced Self-Training for Language Agents

Self-training can potentially improve the performance of language agents without relying on demonstrations from humans or stronger models. The general process involves generating samples from a model, evaluating their quality, and updating the model by training on high-quality samples. However, self-training can face limitations because achieving good performance requires a good amount of high-quality samples, yet relying solely on model sampling for obtaining such samples can be inefficient. In addition, these methods often disregard low-quality samples, failing to leverage them effectively. To address these limitations, we present Reflection-Reinforced Self-Training (Re-ReST), which leverages a reflection model to refine low-quality samples and subsequently uses these improved samples to augment self-training. The reflection model takes both the model output and feedback from an external environment (e.g., unit test results in code generation) as inputs and produces improved samples as outputs. By employing this technique, we effectively enhance the quality of inferior samples, and enrich the self-training dataset with higher-quality samples efficiently. We perform extensive experiments on open-source language agents across tasks, including multi-hop question answering, sequential decision-making, code generation, visual question answering, and text-to-image generation. Results demonstrate improvements over self-training baselines across settings. Moreover, ablation studies confirm the reflection model's efficiency in generating quality self-training samples and its compatibility with self-consistency decoding.