FAIR-Calib tames quantization instability in diffusion LLMs

Diffusion Large Language Models (dLLMs) generate text by iteratively refining tokens, but each decision is committed irreversibly. This creates a 'stability lag': early choices remain fragile even after being written, making them vulnerable to post-training quantization (PTQ) errors. When those errors flip a borderline decision at the write frontier, the mistake becomes permanent and amplifies through later iterations. Until now, no calibration method specifically accounted for this cascading vulnerability.

FAIR-Calib (Frontier-Aware Instability-Reweighted Calibration) solves this with a two-stage approach. Stage I uses a full-precision teacher model to estimate a position prior that flags frontier hits and masked-stage reliability. Stage II performs off-policy, layer-wise calibration by minimizing a reweighted hidden-state MSE loss—effectively prioritizing fragile frontier states without costly end-to-end diffusion rollouts. The authors theoretically justify the weighted objective as a surrogate for output KL divergence. Empirically, FAIR-Calib consistently beats SOTA baselines on LLaDA and Dream under W4A4 quantization (4-bit weights and activations), drastically reducing frontier decision flips and post-commit mismatches. The paper has been accepted at ICML 2026.