Reasoning post-training on long trajectories has substantially improved the reasoning capabilities of large language models, but scaling such training to large datasets is increasingly expensive. The cost of processing long reasoning traces limits how much data can be used in practice.
The new arXiv paper, DeGMatch, studies target-aware reasoning data selection. The paper introduces DeGMatch, a method based on deconfounded gradient matching, to choose which reasoning examples to include in post-training. By selecting data that is relevant to the target task and adjusting for confounding factors, the approach aims to make data selection more efficient.
The abstract is truncated, so specific experimental results are not available in the announcement. The core idea is to reduce the expense of scaling reasoning post-training by selecting the most useful data rather than using all available trajectories.