SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering

What happened: Researchers introduced SteerDistill, a quantization-aware distillation method where a full-precision teacher guides a quantized LLM agent to revise failed responses, with guidance refined using feedback from student execution. Key numbers: SteerDistill exceeds the strongest quantized baseline by 1.67 to 2.18 Why it matters: The method lets quantized agents run unassisted, without a teacher or steering prompts, while producing shorter successful trajectories than quantized baselines. At W2A16 it achieves the highest average score among tested quantized baselines on Qwen3-8B.

Researchers introduced SteerDistill, a quantization-aware distillation method where a full-precision teacher guides a quantized LLM agent to revise failed responses, with guidance refined using feedback from student execution.

The method lets quantized agents run unassisted, without a teacher or steering prompts, while producing shorter successful trajectories than quantized baselines. At W2A16 it achieves the highest average score among tested quantized baselines on Qwen3-8B.