File size: 1,878 Bytes
74c97a3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | ---
license: mit
---
> a simple yet effective postSFT method that enhances long CoT reasoning without requiring new long CoT responses
>
# Introduction
- Here, we show the results of open-source reasoning LLMs before and after ThinkPO.
## Accuracy
| Models | Dataset | SFT | Ours (+ThinkPO) | Improv. (%) |
|:--------:|:--------:|:--------:|:--------:|:--------:|
|DeepSeek-R1-Distill-Qwen-7B (Deepseek) |MATH500 | 87.4 | 91.2 | 4.3% |
|| AIME | 56.7 | 43.3 | -23.6% |
|| GPQA | 47.0 | 49.5 | 5.3% |
|| GSM8K | 87.2 | 87.6 | 0.5% |
|| Olympiad | 58.6 | 58.6 | 0.0% |
|Bespoke-Stratos-7B (Bespoke)| MATH500 | 84.0 | 82.8 | -1.4% |
|| AIME | 20.0 | 23.3 | 16.5% |
|| GPQA | 37.9 | 43.4 | 14.5% |
|| GSM8K | 92.9 | 93.3 | 0.4% |
|| Olympiad | 44.1 | 48.5 | 10.0% |
## Average Response Length
| Model | Dataset | SFT | Ours (+ThinkPO) | Improv. (%) |
|:--------:|:--------:|:--------:|:--------:|:--------:|
|DeepSeek-R1-Distill-Qwen-7B (Deepseek) | MATH500 | 2577 | 3021 | 17.2% |
|| AIME | 11419 | 12875 | 12.8% |
|| GPQA | 4895 | 5604 | 14.5% |
|| GSM8K | 619 | 668 | 7.9% |
|| Olympiad | 7196 | 7383 | 2.6% |
|Bespoke-Stratos-7B (Bespoke)| MATH500 | 5696 | 6404 | 12.4% |
|| AIME | 19858 | 20079 | 1.1% |
|| GPQA | 5968 | 7301 | 22.3% |
|| GSM8K | 1404 | 1755 | 25.0% |
|| Olympiad | 11140 | 12204 | 9.6% |
--- |