Reversible Circuit LLM
dennisonbertram/reversible-circuit-llm
Síntese ótima de circuitos reversíveis com verifier quântico integrado
Ver no GitHubAgent especializado em otimização de circuitos reversíveis quânticos para criptografia ECDSA, treinado com verificação contínua e busca guiada. Combina síntese de circuitos, minimização de custo Toffoli e validação cripto-algébrica em loop agentico com DPO/GRPO. Desenvolvido autonomamente com Claude Code para resolver o desafio ECDSA.fail secp256k1.
Instalação
npx convez-ai dennisonbertram/reversible-circuit-llm1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Otimização de circuitos reversíveis para operações criptográficas
- Síntese automática de hardware quântico com restrições de corretude
- Pesquisa em computação quântica e criptografia de curva elíptica
Veredito

Projeto técnico robusto que enfatiza honestidade experimental: identifica claramente o gargalo (execução simbólica, não capacidade do modelo) e sugere direções neuro-simbólicas legítimas. Artefato de pesquisa valioso para quem trabalha com síntese quântica e verificação formal, mas demanda expertise em circuitos reversíveis e criptografia.
README
A tradução para o português deste README ainda está na fila.
reversible-circuit-llm — a verifier-grounded specialist for reversible-circuit optimization
Training a small, open-source model to do the kind of work the ECDSA.fail secp256k1 point-addition challenge demands: verifier-guided, cost-minimizing synthesis/optimization of reversible quantum circuits under hard correctness constraints (cost = average Toffoli count × peak qubit width; lower is better; every circuit must be classically correct, reversible, phase-clean, and ancilla-clean).
The thesis: with a microsecond-exact verifier, this is not ordinary fine-tuning — it's search + self-improvement with an LLM as the policy. The repo is the machinery for that.
Headline result
Held-out reversible-circuit synthesis (the model must emit a circuit that passes all four
validity gates), valid_rate = fraction solved with best-of-16 (the verifier is a free inference
oracle):
| model | held-out valid_rate | mean reward |
|---|---|---|
| base Qwen2.5-Coder-1.5B | 0% (writes Python) | −1.00 |
| v1 SFT (bloated MMD targets) | 0% | −0.59 |
| v4 SFT (24.5k optimal targets + bug fix) | 4.8% (solves the easiest band) | +0.07 |
Key finding: a 7B, RL (grpo_v2), and reasoning-CoT (4,310 verified traces) all land at the same
~4% ceiling (solve the easiest band, fail n≥3). So this is not a capacity, data, RL, or reasoning
problem — the real bottleneck 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. The honest next directions are tool-use (let the model call the
verifier/synth.py and offload execution), frontier-scale reasoning models, or neuro-symbolic
methods — not more data or bigger models. Full analysis in eval/EVAL_REPORT.md.
What drove the jump from the PoC:
- A verifier-as-search data factory (
proxy/synth.py): training targets at 0.54× the Toffoli cost of the textbook (MMD) references — the PoC was training the model to imitate bloat. - A fundamental bug fix: the S-box / GF(2) task prompts omitted their truth-table/matrix, making those tasks underspecified (unlearnable, contradictory targets) and capping diversity at ~2,300. Fixing it unlocked 31,718 distinct tasks (14×).
- Best-of-N with the verifier oracle — the honest deployment mode.
What's here
proxy/— the environment.proxy_env.pyis a faithful proxy verifier, bit-identical (800/800) to the real Rust simulator (vendored inproxy_rs/).tasks.pyis the task curriculum (7 families × bands B0–B6 + held-out generalization).synth.pyis the verifier-gated search engine (greedy-delete + peephole + simulated annealing + IDA*-optimal for tiny n).reason_gen.pyproduces verified algorithmic chain-of-thought traces.data/— dataset builders (build_sft.py, generators) + the mined optimization-move corpus (moves_raw.jsonl, 264 accepted ECDSA.fail submissions). Large generated.jsonlare git-ignored (regenerate via the scripts / Modal fan-outtrain/gen_curriculum.py).train/— Modal training apps:modal_app.py(SFT),dpo_app.py(DPO + collapse-proof validity-gated GRPO),gen_curriculum.py(CPU fan-out optimal-target generation),eval_modal.py(GPU-side eval),sft_gemma_standalone.py(Gemma-4 arm).eval/—eval_proxy.py(held-out synthesis, best-of-N),agentic_solve.py(verifier-feedback repair loop),eval_cfg.py(real-challenge move quality),run_harness.py(the real ECDSA.fail harness wrapper). Full results ineval/EVAL_REPORT.md.recon/— the reconnaissance that grounded everything (harness semantics, the ~105 DIALOG knobs, mined reasoning episodes, proxy-env design, training-stack analysis).PLAN.md,artifacts/MODEL_CARD.md— plan and model card.
Reproduce (Modal; the verifier being CPU-cheap is what makes this affordable)
python3 data/build_sft.py # base SFT set from mined corpus
ECDSA_BUILD_HEAVY=1 modal run train/gen_curriculum.py --n-seeds 800 --n-per-family 8 # 30k optimal targets
ECDSA_BUILD_HEAVY=1 modal run train/modal_app.py --stage sft --dataset-path /root/data/sft_v4f.jsonl ...
ECDSA_BUILD_HEAVY=1 modal run train/dpo_app.py --stage grpo_v2 --sft-path /ckpts/sft/qwen-sft-v4 ...
ECDSA_BUILD_HEAVY=1 modal run train/eval_modal.py --model-dir /artifacts/<merged> # best-of-N valid_rate
See train/README.md and eval/EVAL_REPORT.md for the full pipeline and numbers.
License
Apache-2.0. Base models used are Apache-2.0 (Qwen2.5-Coder, Gemma-4). The vendored
proxy_rs/{circuit,sim}.rs derive from the ECDSA.fail challenge harness (see its NOTICE).
🤖 Built autonomously with Claude Code.