Opening the AI black box: program synthesis via mechanistic interpretability
This work addresses the challenge of making AI models more interpretable and trustworthy by synthesizing programs without relying on human training data, though it is incremental as it builds on existing interpretability techniques.
The authors tackled the problem of program synthesis by using mechanistic interpretability to auto-distill learned algorithms from neural networks into Python code, achieving a solution rate of 32 out of 62 algorithmic tasks, including 13 not solved by GPT-4.
We present MIPS, a novel method for program synthesis based on automated mechanistic interpretability of neural networks trained to perform the desired task, auto-distilling the learned algorithm into Python code. We test MIPS on a benchmark of 62 algorithmic tasks that can be learned by an RNN and find it highly complementary to GPT-4: MIPS solves 32 of them, including 13 that are not solved by GPT-4 (which also solves 30). MIPS uses an integer autoencoder to convert the RNN into a finite state machine, then applies Boolean or integer symbolic regression to capture the learned algorithm. As opposed to large language models, this program synthesis technique makes no use of (and is therefore not limited by) human training data such as algorithms and code from GitHub. We discuss opportunities and challenges for scaling up this approach to make machine-learned models more interpretable and trustworthy.