Automated Mixed Precision Search for Efficient Neural Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2308 papers353 conceptsfrozen 2026-10-04

In the quest to deploy deep learning models on resource-constrained devices, quantization is a primary tool. By reducing the precision of weights and activations from 32-bit floating point to lower bit-widths like 8-bit or even 2-bit integers, we can drastically reduce memory footprints and accelerate inference.

However, standard quantization often treats every layer in a network the same way. This "one-size-fits-all" approach is suboptimal. Some layers are highly sensitive to noise and require high precision to maintain accuracy, while others are remarkably resilient and can be compressed to extreme levels without impact. This realization has led to the development of Mixed Precision Quantization (MPQ), where each layer is assigned an individual bit-width tailored to its specific role in the network.

The Combinatorial Search Problem

The fundamental challenge of mixed precision is not how to quantize, but which bit-width to choose for each layer. If a model has $L$ layers and we have $B$ possible bit-widths (e.g., 2, 4, and 8 bits), the search space is $B^L$. For a modern architecture with hundreds of layers, this results in an exponentially large search space that makes manual tuning or exhaustive search computationally impossible.

As noted by Milos Nikolic in BitPruning, traditional profiling or reinforcement learning methods often suffer from extreme computational overhead, sometimes requiring hundreds of candidate evaluations to find a single effective configuration.

Measuring Layer Sensitivity

To navigate this search space efficiently, researchers focus on identifying the "sensitivity" of different layers. A sensitive layer is one where a small reduction in precision leads to a large increase in the model's total error.

Several metrics have emerged to quantify this:

  • Hessian Information: Zhewei Yao and the authors of HAWQ-V3 use second-order derivative (Hessian) information to identify layers with high curvature, which are more sensitive to quantization noise.
  • Network Orthogonality: In the context of Diffusion Models, Rocco Manz Maruzzelli introduced MPQ-Diff, which uses an orthogonality metric to determine which layers are critical at different stages of the denoising process.
  • Statistical Outliers: Wei Feng in MPQ-DM utilizes Kurtosis to identify weight channels with significant outliers, allocating higher precision to those specific "heavy hitters" to prevent performance collapse.

Once sensitivity is understood, the next step is automating the allocation of bits across the network. Three main strategies have emerged to solve this optimization problem:

  1. Differentiable Search: BoRui Wu and Péter Vajda proposed a Differentiable Neural Architecture Search (DNAS) framework in Mixed Precision Quantization of ConvNets via DNAS. By relaxing discrete bit-widths into a continuous, differentiable space, the optimal precision for each layer can be learned directly via gradient descent during the training process.
  2. Integer Linear Programming (ILP): HAWQ-V3 treats bit-allocation as a constrained optimization problem. By using ILP, it can find a configuration that minimizes model perturbation while strictly adhering to hardware constraints like total model size or latency.
  3. Reinforcement Learning: Lin Ning introduced ADRL, which uses an augmented deep reinforcement learning agent to "play" the search space, learning over time which layer-bit combinations yield the best accuracy-efficiency trade-offs.

Hardware-Aware Optimization

A theoretical reduction in model size does not always translate to a speedup on real hardware. Many mixed-precision methods struggle because switching between different bit-widths (e.g., 4-bit and 8-bit) can introduce significant latency overhead.

To address this, Zhewei Yao and colleagues developed HAWQ-V3, which enables "true" integer-only inference. By using dyadic arithmetic (replacing floating-point divisions with integer multiplications and bit-shifts), they ensure that the mixed-precision model remains hardware-efficient. Similarly, Zhuoyi Zhang explored Flexible 8-bit Formats, allowing layers to switch between integer (INT8) and floating-point (FP8) formats based on their specific range and precision needs.

Beyond Static Models: Dynamic and Generative Precision

The field is now moving toward even more granular control. Yonggan Fu proposed Cyclic Precision Training (CPT), where bit-precision is treated as a dynamic hyperparameter that changes during training, much like a learning rate. This not only speeds up training but can actually improve final model generalization.

In generative AI, frameworks like Qua2SeDiMo by Keith G. Mills use Graph Neural Networks to map model architectures to performance, identifying the specific layers in Diffusion Models that resist extreme quantization. This allows models like SDXL to be compressed to sub-4-bit weights while maintaining high image quality.

By shifting from uniform constraints to automated, layer-wise precision, mixed precision quantization allows us to capture the varying information density across a neural network, pushing the boundaries of what can be achieved on edge hardware.

Papers in this version

Showing 60 of 308 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 238 papers
  2. Version 1 · 2026-10-04 · 308 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_mixed_precision_search_space_complexity_9c76d256,
  title = {mixed-precision search space complexity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-precision-search-space-complexity/rdr_9c76d256e146408ab2ba54ee67059390},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_automated_mixed_precision_search_for_eff_9c76d256_v1,
  title = {Automated Mixed Precision Search for Efficient Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-mixed-precision-search-for-efficient-neural-networks/rdr_9c76d256e146408ab2ba54ee67059390/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}