Code Generation on HumanEval (accuracy, delta, error correction)
89AccuracySST
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SSTcompute=staged compute, imax=4, decoding=greedy argmax2026.04 | 89 | 1.8 | 14.3 | |
| Matched fine-tuned baselinecompute=one forward pass per token, decoding=greedy argmax2026.04 | 87.2 | — | — |