Parallel Language Models - HumanEval: leaderboard
Parallel Language Models (2026): post-trained 1.3B and 3B research checkpoints with explicit beta and decoding configurations. The 3B autoregressive result uses the same PLM checkpoint in AR inference mode, not a separately trained AR model. GSM8K was used to select learning rates and epochs (Appendix A7.SS3), so its reported score is a development metric; other task scores are kept distinct. Throughput is excluded. Score (%); higher is better.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | PLM-3B Beta1 Relaxed | 32.3 | |
| 2 | PLM-3B Beta1 AR Decoding | 23.2 | |
| 3 | PLM-1.3B Beta0.5 Relaxed | 17.7 | |
| 4 | PLM-1.3B Beta2 Relaxed | 15.8 | |
| 5 | PLM Study AR-1.3B | 12.8 | |
| 6 | PLM-1.3B Beta1 Relaxed | 10.4 |
Interactive version: theaggregate.ai/benchmark?slug=parallel-language-models-humaneval · How It Works · Data refreshed daily, snapshot 2026-10-21.