Abstract:Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including edge attribution patching and sparse autoencoders, to identify minimal circuits and components supporting SQL generation. Our analysis reveals both the potential and limitations of current interpretability methods, showing how circuits can vary even across similar queries. Lastly, we demonstrate how mechanistic interpretability can identify flawed heuristics in models and improve synthetic dataset design. Our work provides a comprehensive framework for evaluating and advancing interpretability techniques while establishing clear boundaries for their reliable application.
Abstract:Large language models (LLMs) aligned to human preferences via reinforcement learning from human feedback (RLHF) underpin many commercial applications. However, how RLHF impacts LLM internals remains opaque. We propose a novel method to interpret learned reward functions in RLHF-tuned LLMs using sparse autoencoders. Our approach trains autoencoder sets on activations from a base LLM and its RLHF-tuned version. By comparing autoencoder hidden spaces, we identify unique features that reflect the accuracy of the learned reward model. To quantify this, we construct a scenario where the tuned LLM learns token-reward mappings to maximize reward. This is the first application of sparse autoencoders for interpreting learned rewards and broadly inspecting reward learning in LLMs. Our method provides an abstract approximation of reward integrity. This presents a promising technique for ensuring alignment between specified objectives and model behaviors.