← News

Models & toolsSources: arXiv · 20 Nov 2026

AutoMaster uses skill-guided fine-tuning to make a fixed model a better teacher

The arXiv paper proposes Skill-FT, which shapes teacher behavior with natural-language skills, and reports gains of 1.42-8.99 percentage points over vanilla SFT on 4B and 9B models.

In plain words

When a large language model generates example solutions to train a smaller model, even correct examples can include bad habits like unsupported assumptions or shallow checking. The authors use written skill descriptions to steer the teacher toward better behavior, then train the target model on those improved examples. Their system, AutoMaster, automates this process with an LLM agent.

What was announced

The paper argues that task success alone does not show which behaviors post-training should reinforce, since successful teacher trajectories can still contain unsupported assumptions, unproductive repetition, or superficial verification. Skill-FT addresses this by using natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tuning a target model on the resulting trajectories. AutoMaster automates the approach: an LLM agent develops the interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation.

Across Terminal, Search, and software engineering tasks with 4B and 9B target models, the authors report that AutoMaster improves mean performance over vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. They also found that keeping skill instructions in the training data during SFT yielded higher mean performance than removing them across all six settings, even though the skills were absent at evaluation and the supervised outputs were identical, which they suggest supports better behavioral internalization.

What it does not show

  • The evidence is the paper's abstract only; the full methodology, experimental details, and peer review status are not available in the supplied passages.
  • The reported gains are the authors' own claims and have not been independently confirmed.
  • The abstract does not specify which tasks or model sizes correspond to which end of the 1.42-8.99 percentage-point range.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.