Number 9 of 12 · November 20, 2026

QDA: Closed-Loop LLM Quantization Design with Agents

QDA uses an SI agent to design the compression recipe for each language model, and the abstract says it beats the strongest public recipes on four reasoning models.

What they found

The authors built QDA, a system where an SI agent designs the quantization recipe for each model and precision. Quantization shrinks a model's numbers so it is cheaper to serve. To make testing fast, QDA scores each candidate in minutes. It measures how far the candidate's outputs drift from the original model (output KL divergence) on reasoning traces the agent never sees. It then checks promising ones with long generations to catch failures such as repetition loops. Across four reasoning models, accuracy rose over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint.

Why it matters

Experts today redesign quantization recipes by hand for each new model, which takes time and is often not the best choice. This work suggests an agent can do that job with a fixed search budget. The abstract says every QDA checkpoint runs on existing serving kernels with no inference overhead. The authors say they will release the library, the evaluator and the harness.

How it changes the race

If the results hold, cheaper serving of new models could come faster and with less expert labor. The gains are modest in percentage points, so this is a useful engineering step, not a large jump.

Safety and control

The agent never sees the reasoning traces used for scoring, which guards against it overfitting to its own test. The long-generation checks also catch failures such as repetition loops. The abstract does not discuss wider safety or control questions.

The abstract, by the authors

Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.

Quoted from arXiv:2611.23388. The research and its words belong to its authors.

The authors

Xingyang Li, Tianlang Zhao, Song Han, Ligeng Zhu

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.