reversible-circuit-8b-tool API & Inference Endpoint

What it does

Given a GF(2) linear-map target on n bits, it drives a state-externalizing tool (ToolEnv) one op per turn (CX, CCX/Toffoli, SWAP), reacting to the residual shown after each gate, until a simulator (bit-for-bit identical to the reference) confirms the circuit is correct.

Honest evaluation (held-out, 40 tasks/band, best-of-5)

Table with columns: Band, n, solve rate
Band	n	solve rate
B1	3	95%
B2	4	92.5%
B3	5	40%
B4	6	5%
Overall		~58%

Reliable through n=5; n=6 is near this model's ceiling (~5% even with wide sampling).

What we learned (and what did NOT work — stated plainly)

The tool removes the real bottleneck. Without it, a 1.5B and a 7B model one-shot-synthesize identically (~4.8%) — the limiter is symbolic execution, not capacity. With the tool, scale then matters (a trained 1.5B caps at n=4; this 8B reaches n=5).
A self-harvest "flywheel" (expert iteration on the model's own verified solutions) did NOT improve held-out capability — a clean negative result. base ≈ iter-1 ≈ iter-2 (~58% best-of-5). An earlier apparent "n=6 cracked 0→7.5%" was a best-of-2 sampling artifact (this base already solves n=6 at ~5% with enough attempts). SFT on a model's own correct outputs re-teaches what it already does; it cannot push the frontier.
Measurement discipline was the real lesson: under-sampled evals manufactured two phantom "wins" that an adequately-sampled, fixed held-out set erased.

This checkpoint is the SFT base (the strongest model in the study). The flywheel iterations did not beat it, so the base is what's shipped.

Intended use & limitations

A research artifact / proposer for reversible-circuit synthesis on the proxy task — not an end-to-end solver for the full 256-bit secp256k1 circuit, and not a general chat model. Use the base Qwen3-8B for general tasks.

Reproduce

Code, data factories, eval harness, and the complete process log: https://github.com/dennisonbertram/reversible-circuit-llm

What it does

Honest evaluation (held-out, 40 tasks/band, best-of-5)

Table with columns: Band, n, solve rate
Band	n	solve rate
B1	3	95%
B2	4	92.5%
B3	5	40%
B4	6	5%
Overall		~58%

Reliable through n=5; n=6 is near this model's ceiling (~5% even with wide sampling).

What we learned (and what did NOT work — stated plainly)

The tool removes the real bottleneck. Without it, a 1.5B and a 7B model one-shot-synthesize identically (~4.8%) — the limiter is symbolic execution, not capacity. With the tool, scale then matters (a trained 1.5B caps at n=4; this 8B reaches n=5).
A self-harvest "flywheel" (expert iteration on the model's own verified solutions) did NOT improve held-out capability — a clean negative result. base ≈ iter-1 ≈ iter-2 (~58% best-of-5). An earlier apparent "n=6 cracked 0→7.5%" was a best-of-2 sampling artifact (this base already solves n=6 at ~5% with enough attempts). SFT on a model's own correct outputs re-teaches what it already does; it cannot push the frontier.
Measurement discipline was the real lesson: under-sampled evals manufactured two phantom "wins" that an adequately-sampled, fixed held-out set erased.

This checkpoint is the SFT base (the strongest model in the study). The flywheel iterations did not beat it, so the base is what's shipped.

Intended use & limitations

Reproduce

Code, data factories, eval harness, and the complete process log: https://github.com/dennisonbertram/reversible-circuit-llm

reversible-circuit-8b-tool

README

What it does

Honest evaluation (held-out, 40 tasks/band, best-of-5)

What we learned (and what did NOT work — stated plainly)

Intended use & limitations

Reproduce

Explore FriendliAI today

README

What it does

Honest evaluation (held-out, 40 tasks/band, best-of-5)

What we learned (and what did NOT work — stated plainly)

Intended use & limitations

Reproduce