Aggregate Reasoning Evaluation on MedQA, GSM8K Overall
76.9AccuracySFT-Oracle
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT-Oracle2026.05 | 76.9 | 1,383 | |
| HSIR-SFT2026.05 | 74.67 | 1,195 | |
| HSIR-DPOIteration=3, |Train|=19.6K2026.05 | 73.35 | 885 | |
| ReSTEM2026.05 | 72.33 | 1,365 | |
| HSIR-DPOIteration=2, |Train|=23.6K2026.05 | 72.16 | 921 | |
| HSIR-SFTIteration=3, |Train|=38.8K2026.05 | 71.33 | 896 | |
| HSIR-DPOIteration=1, |Train|=17.9K2026.05 | 70.86 | 1,007 | |
| H-GRPO2026.05 | 70.71 | 710 | |
| HSIR-SFTIteration=2, |Train|=36.6K2026.05 | 70.68 | 950 | |
| SFT-Oracle|Train|=8.1K2026.05 | 70.49 | 1,392 | |
| ReSTEMIteration=3, |Train|=53.6K2026.05 | 69.73 | 1,114 | |
| Long2Short2026.05 | 69.48 | 668 | |
| STaR2026.05 | 69.3 | 1,442 | |
| GRPO2026.05 | 69.17 | 974 | |
| HSIR-SFTIteration=1, |Train|=33.6K2026.05 | 69.16 | 1,075 | |
| SFT-Initial2026.05 | 69.05 | 1,540 | |
| ReSTEMIteration=2, |Train|=51.0K2026.05 | 69.02 | 1,160 | |
| CosFn2026.05 | 68.48 | 966 | |
| IRPOIteration=3, |Train|=20.5K2026.05 | 68 | 1,271 | |
| IRPOIteration=2, |Train|=27.5K2026.05 | 67.96 | 1,294 | |
| RFT|Train|=132.2K2026.05 | 67.42 | 1,292 | |
| ReSTEMIteration=1, |Train|=44.7K2026.05 | 67.15 | 1,268 | |
| IRPOIteration=1, |Train|=22.7K2026.05 | 66.43 | 1,359 | |
| ReGenesis|Train|=127.2K2026.05 | 65.59 | 363 | |
| STaRIteration=2, |Train|=6.1K2026.05 | 65.06 | 1,328 | |
| STaRIteration=3, |Train|=6.4K2026.05 | 64.82 | 1,288 | |
| STaRIteration=1, |Train|=5.7K2026.05 | 64.31 | 1,379 | |
| SFT-Initial|Train|=1.0K2026.05 | 62.45 | 1,536 |