TL;DR of the paper, generated automatically.

Quantized LLM agents can regain most of their lost task success.

• Method: a full-precision teacher steers student retries, then the student trains unassisted.
• Results: at W4A16, average score stays within one point of BF16.
• Versus QeRL: success rate 1.84 points higher on τ²-Bench, 67% fewer GPU hours.