Quantized Agents Get a Teacher

🔥 What happened Researchers from Tsinghua and Microsoft unveiled SteerDistill: a full-precision teacher corrects quantized LLM agents in a closed loop by analyzing their own failed rollouts and revising responses on the fly. 💡 Why it matters Until now, 4-bit agents often lost track of task requirements or environment conditions. SteerDistill beats the strongest quantized baseline by 1.67–2.18 points and stays within one point of BF16 — while using 67% fewer training GPU hours than QeRL on τ²-Bench. ⚡ Our take This proves quantization no longer has to be a trade-off: if you want efficient agents, bet on closed-loop teacher steering, not ever-larger models.