Flint
Compressing reasoning without breaking it
Flintprint, in dot
- Kind
- Research
- Status
- Earlier work
- Reading
- 1 min
A study on whether reasoning models can learn from compressed versions of their own outputs, reducing the tokens they use without losing the steps needed to solve a problem.
Flint's plate, shown with its construction. One root, two arms of equal reach: a bar that holds its weight, and under it a wedge that runs down to a thread. In the print, the plate is the clear area. Drawn by Plate.
Reasoning models often generate long traces before producing a final answer. Some of this text carries the computation needed to solve the problem, while other parts restate the task, describe a plan, or discuss how to format the response. Flint explores whether a model can learn to produce shorter traces by training on compressed versions of its own outputs.
The process uses self-distillation: the model first generates examples, their reasoning traces are compressed, and the resulting data is used to fine-tune the same model. The final answers are left unchanged. The intention is to reduce the text surrounding the reasoning without removing the steps on which the answer depends.
Crucially, these parts cannot all be treated in the same way. In the Qwen experiments, compressing the entire trace uniformly caused severe looping under greedy decoding: the model generated more tokens while answering far fewer questions correctly. A section-aware approach instead preserves computation and verification, compresses planning more lightly, and removes transitions and restatements.
The results on GSM8K were encouraging: in the tested Qwen and Gemma configurations, section-aware compression reduced reasoning tokens and improved accuracy relative to fine-tuning on uncompressed outputs. This was not a consistent improvement across tasks, however. Some benchmarks regressed, and the results were sensitive to decoding temperature and the prompts used during training.
These experiments suggest that selective compression can make reasoning more efficient, but they do not establish that it preserves every capability. The evaluations are small, and shorter traces alone are not enough to show that a model reasons better. The article discusses the experiments and their limitations, while the code and recorded results are available on GitHub.