SabaPivot/repro-on-the-power-of-approximate-reward-models-for-inference-time-scaling-sequential-monte-carl
0
Reproduction: On the Power of (Approximate) Reward Models for Inference-Time Scaling: Sequential Monte Carlo and Beyond
An open experiment logbook, published with Trackio.
Independent finite-state verification v2
The canonical reproduction now includes a second six-gate audit that replaces formula-only plots with actual hidden-prefix searches, exact finite-particle output laws, exhaustive small-policy enumeration, and a literal Metropolis–Hastings transition matrix.
- Re-run:
python upgrade_smc_v2.py - Results:
outputs_v2/smc_v2_results.json - Integrity manifest:
outputs_v2/SHA256SUMS.json - Result SHA-256:
ac5e2396b6a75e0a148c58de9c106b1644211f1d4d033502328a105d235e8f37
All six evidence gates pass. A higher-scoring public reproduction was consulted to identify missing evidence categories; the implementation and all generated results here were independently produced.