SMRA — VibeThinker-3B Reasoning Amplifier

Research question

SMRA asks whether a fixed 3B mathematics model can produce better final answers under a fixed compute budget through independent solution paths, deterministic checks, error-directed retries, key-claim review, and adaptive allocation of additional attempts. It studies system-level performance, not whether the underlying model becomes equivalent to a frontier model.

Experiment design

The staged design begins with S0, one unassisted BF16 answer per problem, and S1, three independent paths with majority selection. Later stages add deterministic verification, error-driven retries, key-claim challenges, and adaptive expansion from three to five or seven candidates. Only S0 and S1 are currently implemented and exposed as completed engineering paths.

Frozen baseline and reproducibility

The v0 BF16 baseline fixes the model revision, prompt hash, generation settings, seed list, dataset manifest, and Git tags. Each run records prompts, raw outputs, token counts, latency, memory, verifier results, and source commit. Frontier Text-20 is scored through a bridge whose answer key remains outside the model-accessible project tree.

Current stage

The local runtime, scheduling, streaming, run logging, batch-evaluation bridge, and desktop interface are working. Formal Frontier Text-20 results remain pending; the next milestone is to finish and review the frozen baseline before adding verification or adaptive stages.

Research activity