🤖 AI

Model-Preserving Adaptive Rounding

This paper introduces YAQA, an adaptive rounding quantization algorithm that leverages theoretical end-to-end error bounds based on Hessian approximations to significantly outperform existing methods like GPTQ and quantization-aware training without adding inference overhead.

Original authors: Albert Tseng, Zhaofeng Sun, Christopher De Sa

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Albert Tseng, Zhaofeng Sun, Christopher De Sa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Shrinking a Giant Library

Imagine you have a massive, incredibly detailed library (this is your Large Language Model, or LLM). It contains billions of books (parameters) written in extremely precise ink (high-precision numbers like 32-bit floating point). This library is so huge that it’s expensive to keep, hard to transport, and slow to search through.

To make it cheaper and faster to use, we want to shrink the library. We want to rewrite the books using simpler, coarser ink (low-precision numbers, like 4-bit integers). This process is called Quantization.

The Problem:
When you rewrite a book with simpler ink, you inevitably lose some detail. If you just round every number to the nearest simple value blindly, the story changes. The "meaning" of the library shifts, and the AI starts giving wrong or weird answers.

Most current methods try to fix this by looking at one book at a time. They say, "Let’s make sure this specific book doesn’t change too much." But this is like editing a novel chapter by chapter without reading the rest of the book. You might fix Chapter 1, but you accidentally break the plot twist in Chapter 10 because you didn’t consider how they connect.

The Solution: YAQA (Yet Another Quantization Algorithm)

The authors introduce a new method called YAQA. Instead of looking at one book (layer) in isolation, YAQA looks at the entire library to see how changing one book affects the whole story.

Here is how YAQA works, broken down into three simple concepts:

1. The "Map of Importance" (The Hessian)

To decide which details to keep and which to throw away, you need a map that tells you what matters most. In math, this map is called the Hessian.

Old Way (GPTQ/LDLQ): These methods use a "local map." They only look at the immediate input to a layer. It’s like a librarian who only cares if the book fits on the shelf, ignoring whether the book is actually useful for the story.
YAQA’s Way: YAQA builds a "global map." It calculates how much the final output of the AI changes when you tweak the weights. It asks, "If I change this number here, does the final answer get worse?"

2. The "Smart Rounding" Strategy

Once you have the map, you need to round the numbers.

The Analogy: Imagine you are packing a suitcase. You have a limited amount of space (low precision).
Old Way: You pack items one by one. You put in a shirt, then pants, then a sock. You don’t think about how the shirt wrinkles the pants.
YAQA: It uses a technique called Adaptive Rounding. It looks at the "global map" and adjusts the packing order. It realizes, "If I squeeze this shirt, I need to loosen the pants to keep the total volume the same." It balances the errors across the whole model so the final result stays accurate.

3. The "Kronecker" Shortcut

Calculating the "global map" for a giant AI model is normally too slow and requires too much memory. It’s like trying to draw a map of every single grain of sand on a beach.

The Trick: YAQA uses a mathematical shortcut called a Kronecker-factored approximation.
The Analogy: Instead of mapping every grain of sand, you map the beach and the ocean separately, then combine them. It’s much faster to calculate, but it still gives you a surprisingly accurate picture of the whole landscape. This allows YAQA to be fast enough to actually use.

Why Is YAQA Better?

The paper claims three main victories for YAQA:

It’s Mathematically Proven to Be Better:
The authors didn’t just guess; they proved it. They created a mathematical "bound" (a guarantee) that shows YAQA will always produce a model closer to the original than the old methods (like GPTQ or LDLQ). It’s like having a guarantee that your shrunken library will be 99% accurate, whereas the old method only promised 90%.
It Keeps the "Soul" of the Model:
They measured how different the quantized model is from the original using a metric called KL Divergence (think of this as "distance from the original personality").
- YAQA reduces this distance by about 30% compared to the best existing methods.
- It even beats Quantization Aware Training (QAT). QAT is a method where you re-train the AI to get used to the low precision. YAQA achieves better accuracy without needing to re-train the model, saving massive amounts of time and energy.
No Extra Cost for Users:
Once the model is shrunk using YAQA, it runs exactly as fast as any other shrunk model. The "smart rounding" happens only during the preparation phase. When you use the AI, there is zero extra overhead.

Summary in a Nutshell

The Goal: Shrink AI models to make them cheaper and faster without losing their intelligence.
The Old Problem: Previous methods focused on small, local details, causing the overall model to drift away from its original behavior.
The YAQA Fix: It uses a global view of the model’s error and a smart mathematical shortcut (Kronecker factorization) to round numbers in a way that preserves the model’s overall output.
The Result: A model that is significantly closer to the original high-quality version than any previous method, proven by math and tested on real-world tasks, with no extra cost to run.

Technical Summary: Model-Preserving Adaptive Rounding

Problem Statement
The primary objective of Large Language Model (LLM) quantization is to compress model parameters while preserving the original output distribution. Most existing post-training quantization (PTQ) algorithms, such as GPTQ and LDLQ, minimize the immediate activation error of individual layers as a proxy for end-to-end error. This approach ignores the cumulative effect of quantization errors on subsequent layers, resulting in a poor approximation of the true end-to-end loss. While Quantization Aware Training (QAT) optimizes end-to-end error, it requires significant computational resources. The authors identify a gap in theoretical understanding regarding which Hessian approximations allow for tractable, high-quality adaptive rounding that considers the full model error.

Methodology
The authors introduce Yet Another Quantization Algorithm (YAQA), an adaptive rounding algorithm that directly minimizes the end-to-end KL divergence. YAQA generalizes LDLQ by utilizing a Kronecker-factored Hessian approximation ( $\tilde{H} = H_O \otimes H_I$ ) rather than the layer-wise activation Hessian ( $H_1$ ) used by LDLQ.

Theoretical Framework:
- Structural Nilpotence Degree (SND): The authors define SND to characterize the convergence time of adaptive rounding algorithms. They prove that for a Kronecker-factored Hessian, the SND is the sum of the SNDs of its factors ( $snd(L_O \otimes L_I) = snd(L_O) + snd(L_I)$ ), allowing YAQA to converge in $m+n$ iterative steps, compared to $n$ steps for LDLQ.
- End-to-End Error Bounds: The paper provides the first theoretical end-to-end error bounds for quantization algorithms. Theorem 3.4 establishes that the true second-order error is bounded by the proxy error and the cosine similarity between the approximate Hessian ( $H_O \otimes H_I$ ) and the true Hessian ( $\nabla^2 L(W^*)$ ). This bound motivates the need for Hessians with high cosine similarity to the true Hessian, low incoherence, and low rank.
Hessian Sketches:
To tractably compute the Kronecker-factored Hessian, YAQA employs two sketches derived from power iteration:
- Sketch A: Assumes token independence within a sequence, reducing variance but introducing bias. It computes expectations over individual tokens.
- Sketch B: Computes one round of power iteration on the true Hessian $\nabla^2 L(W^*)$ starting from an identity initialization. This is computed over entire sequences in a single pass, offering a theoretically and empirically superior approximation to Sketch A when sufficient data is available.

Key Contributions

YAQA Algorithm: A new weight-only PTQ algorithm that generalizes local adaptive rounding to optimize end-to-end KL divergence. It is provably better than GPTQ/LDLQ under the derived error bounds.
Theoretical Bounds: The first end-to-end error bounds for quantization algorithms, linking quantization error to the cosine similarity of the Hessian approximation and the structural properties (incoherence, rank) of the Hessian factors.
Hessian Desiderata: Characterization of the properties required for Hessian sketches to enable tractable adaptive rounding and high-quality models, leading to the development of Sketch A and Sketch B.

Results
YAQA was evaluated on Llama, Gemma, and Qwen models across various bit-widths (2, 3, 4 bits).

KL Divergence: YAQA reduces KL divergence by approximately 30% compared to LDLQ and GuidedQuant.
Comparison with QAT: YAQA achieves a lower KL divergence to the original model than Google’s Gemma 3 QAT recipe, despite QAT using significantly more data and compute. This suggests YAQA better preserves the original model’s distribution.
Downstream Performance: YAQA achieves state-of-the-art performance on downstream tasks (perplexity, zero-shot accuracy) among PTQ methods.
Robustness: YAQA outperforms LDLQ even without incoherence processing and when combined with recovery finetuning. Sketch B is robust to the number of sequences used, achieving strong results with as few as 2K sequences.

Significance
The paper claims that YAQA represents a significant advancement in PTQ by providing the first theoretical guarantees on end-to-end quantization error. By shifting the focus from local layer-wise error to global model error via Kronecker-factored Hessians, YAQA achieves superior fidelity to the original model without adding inference overhead. The authors argue that this approach offers a more principled alternative to ad-hoc Hessian approximations used in prior work and outperforms both local adaptive rounding methods and resource-intensive QAT in terms of distribution preservation.