AutoMaster

Autonomous Post-Training by Making the Same Model a Better Teacher

The AutoMaster teacher robot shows a skill to a student robot working in a notebook.

What makes a better demonstration?

A teacher can solve a task and still teach a bad habit. AutoMaster studies trajectories, writes natural-language skills to guide the same teacher, and selects demonstrations for supervised fine-tuning. The student is evaluated without those skills.

Project video · 1 min 17 sec

How AutoMaster works

See how an agent turns feedback about teacher behavior into skills and training data.

Open video file

Method

Inside Skill-FT

Two feedback loops connect what the teacher demonstrates with what the student learns.

The inner loop analyzes behavior, designs skills and selects teacher rollouts. The outer loop trains the student and returns evaluation feedback.
AutoMaster’s training and evaluation loops View full figure

Inner loop

Shape the demonstrations

The agent identifies recurring behaviors, designs skills for relevant tasks, and pilots them with the teacher. It uses the rollout evidence to refine the skills and select training examples.

Outer loop

Learn from the student

The student fine-tunes on the selected trajectories. Evaluation without skill prompts reveals which behaviors transfer and which need another round of intervention.

Skills

What the teacher reads

Explore nine skills from AutoMaster research runs. Open an illustration to read the teacher’s instructions and visit the original skill file.

Terminal tasks

Test a numerical claim

When a solver must satisfy an error bound, compare its output with an independent reference. Keep the failing input and the original tolerance while diagnosing the error.

Teacher instruction
“Define the reference and bound before running this block.”
Check the saved result

When a task delivers a file, model, or service, use the same entry point its user will use. Measure the requested property on that final output.

Teacher instruction
“Exercise the finished deliverable.”
Keep exact results during optimization

When faster code must match a reference exactly, first speed up the work around sensitive operations. Check tied rankings and floating-point reductions against the reference.

Teacher instruction
“Measure this conservative implementation against the required performance target.”

Web search

Break out of a stalled search

When repeated queries lead to the same weak candidate, find the guessed detail they share. Drop it and search a clue that could independently identify the target.

Teacher instruction
“Drop that guessed restriction from the next discovery query.”
Check what a distance measures

When distance could rule a candidate in or out, check the endpoints, units, and travel mode. A nearby station may explain a clue while leaving the requested walking distance unknown.

Teacher instruction
“Before using a returned number, identify its two endpoints, what was measured, and its units.”
Search in the source’s own terms

When English summaries omit a local fact, keep one verified name or institution and search for the missing relationship in the language of the relevant record.

Teacher instruction
“Preserve names, dates, and roles.”

Repository construction

Find which object owns the work

When framework objects cooperate, try the smallest real construction and activation sequence. Identify which object starts the work, handles it, and cleans up.

Teacher instruction
“Determine which object owns startup, processing and cleanup.”
Exercise the public API

When a library wraps an external service, call its real public entry point. Record the outbound request at the dependency boundary and check the value returned to the caller.

Teacher instruction
“Use a fresh ledger for the call.”
Keep nested readers in sync

When a parser delegates a nested region, specify which reader consumes the closing marker. Test a nested item followed by a sibling so a misplaced cursor becomes visible.

Teacher instruction
“Define whether the closing marker belongs to the child and keep that convention at every call site.”

Browse all skill files on GitHub

How the skills changed a rollout

Read Figure 5 in the paper

Verifying convergence

execute-numeric-decision Figure 5a. Verifying convergence. Both solvers met the error threshold. The unguided teacher claimed fourth-order accuracy despite observing second-order convergence. The guided teacher switched to a nine-point stencil and verified fourth-order convergence on progressively finer grids.
Case notes

Both solvers met the error threshold. The unguided teacher claimed fourth-order accuracy despite observing second-order convergence. The guided teacher switched to a nine-point stencil and verified fourth-order convergence on progressively finer grids.

Read Figure 5a

Finding the director

recover-search-anchor Figure 5b. Finding the director. Both searches found the same episode. The unguided teacher repeatedly guessed an award. The guided teacher used the episode credits to identify director Meera Menon, then checked the award clue.
Case notes

Both searches found the same episode. The unguided teacher repeatedly guessed an award. The guided teacher used the episode credits to identify director Meera Menon, then checked the award clue.

Read Figure 5b

Checking the sort order

Trace failures to their source Figure 5c. Checking the sort order. A flawed self-test led the unguided teacher to reverse a global sort. The guided teacher traced the failure to the test input, checked the UTF-16 ordering, and kept the required ascending sort. The corrected implementation passed 38 of 38 tests.
Case notes

A flawed self-test led the unguided teacher to reverse a global sort. The guided teacher traced the failure to the test input, checked the UTF-16 ordering, and kept the required ascending sort. The corrected implementation passed 38 of 38 tests.

Read Figure 5c

Results

Gains across three domains

AutoMaster improves mean performance by 1.42–8.99 percentage points over Vanilla SFT across six benchmark–model settings.

Full scores and standard deviations
Mean ± SD (%) across three evaluation runs per checkpoint. Gains are in percentage points.
BenchmarkQwen3.5BaseVanilla SFTAutoMasterGain
Terminal-Bench 2.14B15.73 ± 2.2522.47 ± 1.1228.09 ± 2.25+5.62
Terminal-Bench 2.19B25.47 ± 2.5926.59 ± 1.3035.58 ± 1.30+8.99
BrowseComp4B33.69 ± 0.8243.24 ± 0.5446.13 ± 2.98+2.88
BrowseComp9B35.50 ± 2.7246.85 ± 2.9851.53 ± 1.25+4.68
Doc2Repo4B21.39 ± 2.6132.30 ± 4.0033.72 ± 3.58+1.42
Doc2Repo9B33.42 ± 1.6036.05 ± 3.8940.79 ± 3.21+4.74

Terminal-Bench 2.1 and BrowseComp (Lite-185) report task accuracy. Doc2Repo reports mean test pass rate. Both fine-tuning methods use the same teacher, task pool, 512 training trajectories, and SFT recipe.

Citation

Cite AutoMaster

BibTeX
@misc{chen2026automasterautonomousposttrainingmaking,
  title         = {AutoMaster: Autonomous Post-Training by Making the Same Model a Better Teacher},
  author        = {Guoxin Chen and Shuang Sun and Fanzhe Meng and Xinyu Tang and Ruihua Song and Wayne Xin Zhao},
  year          = {2026},
  eprint        = {2611.26510},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2611.26510}
}