Teacher-guided corrections help quantized agents recover
SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering · arXiv:2611.13410
In plain words
Tool-using language models can miss a task requirement or an environmental condition, so their actions go off course. SteerDistill is a training method that uses a full-precision teacher to guide a quantized student through a correction, then trains the student to use that behavior without prompts.
In a multi-turn tool-use diagnostic, guidance raised the share of selected off-course turns with a correct action in four tries from 9.6% to 82.3%. That measures turns, not whole-task success.
The research
For judging what low-bit tool-using agents can do, a failure may reflect a missed requirement or environmental condition that can be corrected in context.
SteerDistill's results make recovery through guided corrections a relevant test of capability, while its measured gains remain tied to the evaluated models and tasks.
SteerDistill trains quantized tool-using agents to learn teacher-guided corrections at points where their actions diverge from successful references.
SteerDistill revises a failed student rollout at its first environment transition that differs from a successful full-precision reference. A teacher guides the student through retries and checks each correction locally; training then distills accepted behavior without giving the learner the prompts.
Across five tests measuring tool use, simulated task success, and exact question-answer matching, the unweighted average was 1.67 to 2.18 points above the strongest quantized baseline for three models at W4A16, and within one point of their full-precision references.
The revision process needs a successful teacher reference, the ability to restore the earlier environment state, and a reliable local check. The reported average covers three models on five tests at W4A16; a locally accepted correction does not guarantee that the full task succeeds.
What the evidence shows
- Correct the point of divergence
SteerDistill is a method for training quantized language-model agents to correct actions that diverge from successful full-precision references, then act without teacher guidance. The paper examines failed rollouts in which agents overlook task requirements or environmental conditions.
Its focus is on recovering at the point where the agent's interaction with the environment goes off course.
- Guidance changes retry outcomes
The BFCL V4 multi-turn test measures agents' tool-use actions across turns. In its turn-level Pass@4 measure, the share of selected divergent turns with at least one correct action in four attempts was 9.6% for unguided retries and 82.3% with teacher steering.
Each attempt began from the same pre-turn history and environment state. The change shows that guidance can help the student find a correct action in these selected contexts.
- A broader but bounded gain
Across five tests of tool use, simulated task completion, and question answering, the score average was higher for SteerDistill than for the strongest quantized baseline, QeRL, on each of three named models at W4A16. The average was 1.67 to 2.18 points higher and stayed within one point of the corresponding full-precision reference.
This supports a result for these models and tests, not a claim that every score improved by the same amount.
- The student must make the correction
In a Qwen3-8B comparison on BFCL V4 Overall at W4A16, teacher steering scored 2.38 points above an unguided retry and 1.78 points above teacher takeover. All three began from the same restored history and environment state.
Since the student generated the guided response, this comparison favors steering over direct teacher replacement in this test, while leaving the wider results to establish the method's overall performance.
Deep
- Original paper authors
- Genming Cui, Xiangyu Li, Zixu Hao, Xin Ding, Yunxin Liu
- Submitted to arXiv
- License
- CC BY 4.0