Hazumi News

Hi HN, I've been exploring various applications of formal methods to ML/interpretability and I've been hoping to get more eyes on the approach.

I have been working on a small interpretability project I call Symbolic Circuit Distillation. The goal is to take a tiny neuron-level circuit (like the ones in OpenAI's "Sparse Circuits" work) and automatically recover a concise Python program that implements the same algorithm, along with a bounded formal proof that the two are equivalent on a finite token domain.

Roughly, the pipeline is:

1. Start from a pruned circuit graph for a specific behavior (e.g. quote closing or bracket depth) extracted from a transformer. 2. Treat the circuit as an executable function and train a tiny ReLU network ("surrogate") that exactly matches the circuit on all inputs in a bounded domain (typically sequences of length 5–10 over a small token alphabet). 3. Search over a constrained DSL of common transformer motifs (counters, toggles, threshold detectors, small state machines) to synthesize candidate Python programs. 4. Use SMT-based bounded equivalence checking to either: - Prove that a candidate program and the surrogate agree on all inputs in the domain, or - Produce a counterexample input that rules the program out.

If the solver finds a proof, you get a small, human-readable Python function plus a machine-checkable guarantee that it matches the original circuit on that bounded domain.

Why I built this

Mechanistic interpretability has gotten pretty good at extracting "small crisp circuits" from large models, but turning those graphs into clean, human-readable algorithms is still very manual. My goal here is to automate that last step: go from "here is a sparse circuit" to "here is a verified algorithm that explains what it does", without hand-holding.

What works today

- Tasks: quote closing and bracket-depth detection from the OpenAI circuit_sparsity repo. - Exact surrogate fitting on a finite token domain. - DSL templates for simple counters, toggles, and small state machines. - SMT-based bounded equivalence between: sparse circuit -> ReLU surrogate -> Python program in the DSL.

Limitations and open questions

- The guarantees are bounded: equivalence is only proven on a finite token domain (short sequences and a small vocabulary). - Currently focused on very small circuits. Scaling to larger circuits and longer contexts is open engineering and research work. - The DSL is hand-designed around a few motifs. I am not yet learning the DSL itself or doing anything very clever in the search.

What I would love feedback on

- Are the problem framing and guarantees interesting to people working on mechanistic interpretability or formal methods? - Suggestions for next benchmarks: which circuits or behaviors would you want to see distilled next? - Feedback on the DSL design, search strategy, and SMT setup.

Happy to answer questions about implementation details, the SMT encoding, integration with OpenAI's Sparse Circuits repo, or anything else.