Flint cuts reasoning tokens by 2–3× without losing performance
Flint trains Qwen3.5-4B and gemma-4-12b on shorter versions of their own reasoning. It preserves the parts that calculate and check an answer while removing or shortening filler, narration, and transitions. The resulting models match or outperform the originals, sometimes by a wide margin, while using two to three times fewer tokens.
Applying the same compressed style to the entire caused a serious failure: with , the model looped on 93% of GSM8K problems and achieved only 3% accuracy. It often reached the right answer but failed to stop. The same reached 90% accuracy at 1.0 on a subset drawn from those failures, suggesting that its knowledge remained intact but its ability to finish reliably had broken.
The full findings, trained models, and code are available.
Key points
- Keep calculation and steps while shortening filler and narration.
- The compressed models used two to three times fewer tokens while matching or beating the originals.
- Compressing every part in the same way caused loops on 93% of GSM8K problems.
- Raising restored accuracy from 3% to 90% on a subset of the failed cases.
- The findings, trained models, and code are available.