SOURCE-LINKED INTELLIGENCE

Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning

arXiv · AI, language, vision and robotics · article · Sep 22, 2026 · UTC

Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that place

Read original source ↗ Open in workspace

recordType

paper

region

Global