Insights from the Inverse: Reconstructing LLM Training Goals Through Inverse Reinforcement Learning
This work provides a new method for understanding and improving LLM alignment, with implications for responsible AI development, though it is incremental as it applies an existing technique (IRL) to a new domain (LLMs).
The paper tackled the problem of interpreting large language models (LLMs) trained with RLHF by using inverse reinforcement learning to recover their implicit reward functions, achieving up to 85% accuracy in predicting human preferences and enabling fine-tuning for improved performance on toxicity benchmarks.
Large language models (LLMs) trained with Reinforcement Learning from Human Feedback (RLHF) have demonstrated remarkable capabilities, but their underlying reward functions and decision-making processes remain opaque. This paper introduces a novel approach to interpreting LLMs by applying inverse reinforcement learning (IRL) to recover their implicit reward functions. We conduct experiments on toxicity-aligned LLMs of varying sizes, extracting reward models that achieve up to 85% accuracy in predicting human preferences. Our analysis reveals key insights into the non-identifiability of reward functions, the relationship between model size and interpretability, and potential pitfalls in the RLHF process. We demonstrate that IRL-derived reward models can be used to fine-tune new LLMs, resulting in comparable or improved performance on toxicity benchmarks. This work provides a new lens for understanding and improving LLM alignment, with implications for the responsible development and deployment of these powerful systems.