According to the arXiv abstract for AutoSFT, language model agents are increasingly used to automate post-training, but when given the full training stack, they converge on supervised fine-tuning and spend most of their effort revising training data. This suggests that data curation is a major bottleneck in the post-training process.
The paper introduces AutoSFT, an agentic system designed specifically for automated data research in supervised fine-tuning. The abstract indicates that existing systems give the agent one kind of resource, but the sentence is cut off, so the precise limitation is not available from the source.
Because the abstract is truncated, the full scope of the contribution and the details of the proposed system remain unclear. Readers interested in the specifics should consult the full paper on arXiv.