Model description

The training signal comes from a microsecond-exact verifier (bit-identical to the challenge's Rust simulator). Rather than fine-tuning on textbook examples, a verifier-gated search engine produces near-optimal circuits (~0.54× the Toffoli cost of textbook references), and the model is SFT'd (LoRA) on 24,545 such optimal targets across a 7-family curriculum. The model emits an op-stream in the harness DSL: X qT, CX qC qT, CCX qC1 qC2 qT (Toffoli — the cost lever), SWAP qA qB.

Intended uses & limitations

Intended: a proof-of-concept / research artifact for verifier-grounded circuit synthesis; a generator of small reversible arithmetic/boolean circuits (use best-of-N with the open-source verifier as an inference oracle); a teaching example for neuro-symbolic / tool-use research.

Not intended: a production solver. It reliably solves only the easiest tasks.

Evaluation (honest)

Held-out reversible-circuit synthesis, valid_rate = fraction solved with best-of-16:

Key research finding: a 7B trained identically, plus reinforcement learning (GRPO) and reasoning chain-of-thought, all plateau at the same ~4%. The bottleneck is not data, capacity, RL, or reasoning — it is the small model's inability to reliably execute multi-step symbolic procedures (Gaussian elimination, ripple-carry) for unseen instances. It can narrate the algorithm but makes execution errors. Even a state-externalizing tool (single gate at a time) didn't break this zero-shot — the remaining gap is sequential planning. The honest next directions are tool-use with training, frontier-scale reasoning models, and neuro-symbolic methods.

How to use

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("dennisonb/reversible-circuit-coder-1.5b")
model = AutoModelForCausalLM.from_pretrained("dennisonb/reversible-circuit-coder-1.5b")

Use the system prompt + task format from the repo (proxy/system_prompt.txt, proxy/sample_task.txt), sample best-of-N, and verify each candidate with the open-source proxy verifier (proxy/proxy_env.py).

Training data

24,545 near-optimal circuit targets generated by the verifier-gated search engine over a procedurally generated curriculum (modular adders/multipliers/inverse, controlled add/sub, GF(2) linear maps, S-boxes; widths 2–7). Move/reasoning corpora mined from 275 accepted ECDSA.fail submissions are also in the repo. Datasets are regenerable via the repo's scripts.

🤖 Built autonomously with Claude Code.