Symbolic planning on Countdown
40.6Exact-match Accuracy (Ngen=128)DISPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DISPObase_model=LLaDA-8B-Instruct, algorithm=DISPO2026.02 | 40.6 | 37.1 | 39.8 | 39.2 | |
| DISPObase_model=SFT w/ s1k, algorithm=DISPO2026.02 | 37.5 | 32.8 | 40.2 | 36.8 | |
| diffu-GRPObase_model=SFT w/ s1k, algorithm=diffu-GRPO2026.02 | 34.8 | 32 | 42.2 | 36.3 | |
| diffu-GRPObase_model=LLaDA-8B-Instruct, algorithm=diffu-GRPO2026.02 | 33.2 | 31.3 | 37.1 | 33.9 | |
| LLaDA-8B-Instructbase_model=LLaDA-8B-Instruct2026.02 | 20.7 | 19.5 | 16 | 18.7 | |
| SFT w/ s1kbase_model=SFT w/ s1k2026.02 | 20.3 | 14.5 | 23.8 | 19.5 |