Mathematical Reasoning on GSM8K (test) (Accuracy, Delta)
90.07AccuracyGibbs-Weighted Retrieval
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Gibbs-Weighted RetrievalInference Strategy=Gibbs-Weighted, Particle Count (K)=12, Physics Interpretation=Attractor Relaxation (β ∝ E−2), Backbone=Microsoft Phi-3.5-mini-instruct (3.8B parameters)2026.06 | 90.07 | 11.6 | |
| Standard SamplingInference Strategy=Standard Sampling, Particle Count (K)=12, Physics Interpretation=High-Temp Ensemble (β = 0), Backbone=Microsoft Phi-3.5-mini-instruct (3.8B parameters)2026.06 | 84.69 | 6.3 | |
| Greedy DecodingInference Strategy=Greedy Decoding, Physics Interpretation=Point Estimate (β → ∞), Backbone=Microsoft Phi-3.5-mini-instruct (3.8B parameters)2026.06 | 78.4 | — |