🤖 machine learning

OSAQ: Outlier Self-Absorption for Accurate Low-bit LLM Quantization

The paper proposes OSAQ, a training-free post-training quantization method that leverages the low-rank property of the Hessian's null space to construct an additive weight transformation that suppresses outliers and significantly improves low-bit LLM quantization accuracy without introducing inference overhead.

Original authors: Zhikai Li, Zhen Dong, Xuewen Liu, Jing Zhang, Qingyi Gu

Published 2026-05-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhikai Li, Zhen Dong, Xuewen Liu, Jing Zhang, Qingyi Gu

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 Problem: The "Heavy Backpack"

Imagine a Large Language Model (LLM) like a genius student who knows almost everything. However, this student carries a massive backpack filled with millions of heavy books (parameters). Because the backpack is so heavy, the student moves very slowly and needs a huge room to store it. This makes it expensive and slow to use on regular devices like phones or laptops.

To fix this, engineers try to "shrink" the backpack by summarizing the books into smaller, simpler notes. This is called quantization. Instead of writing every detail with high precision (like 32-bit numbers), they write rougher summaries (like 2-bit or 4-bit numbers).

The Catch: The backpack isn't just full of normal books; it has a few giant, heavy boulders hidden inside (called outliers). These boulders are so heavy that they ruin the attempt to shrink the backpack. If you try to compress the whole bag, these boulders force you to keep the bag huge, or the student starts forgetting things (the model gets stupid).

The Old Solutions: Stretching and Rotating

Previous methods tried to fix this by:

Stretching (Scaling): Trying to pull the boulders apart to make them fit better.
Rotating: Turning the whole backpack so the boulders are in a different spot.

The paper argues that while these tricks help a little, they aren't enough. They are like trying to fit a boulder into a small box by just stretching the box or turning it sideways. The boulder is still there, and it still causes problems.

The New Solution: OSAQ (The "Magic Sponge")

The authors propose a new method called OSAQ (Outlier Self-Absorption). Instead of stretching or rotating, they use a clever trick based on the backpack's internal structure.

1. The "Silent Zones" (The Hessian Null Space)

The researchers discovered that inside the backpack, there are certain directions where the boulders don't actually push back. Imagine the backpack is made of a special foam. If you push on the foam in most directions, it resists. But if you push in specific "silent zones," the foam offers zero resistance.

In math terms, they found a "Null Space" in the model's data structure where changes don't hurt the model's intelligence. It's like finding a secret pocket in the backpack where you can move heavy items around without the backpack feeling the weight.

2. The "Self-Absorption" Trick

OSAQ works like this:

It identifies those "silent zones" (the Null Space).
It takes the giant boulders (outliers) and adds a little bit of "magic foam" to them.
Because this foam is added in a "silent zone," it cancels out the weight of the boulder without changing how the backpack behaves.
The boulder effectively disappears (is "absorbed"), leaving behind a smooth, uniform pile of sand that is easy to compress.

The Analogy: Imagine you have a lumpy pillow with a rock inside.

Old way: Try to squeeze the pillow harder (scaling) or turn it sideways (rotation). The rock still makes it lumpy.
OSAQ way: You realize the pillow has a hidden seam where you can slide a piece of soft clay in. You slide the clay in right next to the rock. The clay fills the gap and makes the rock feel like part of the soft pillow. Now the whole pillow is smooth and easy to squish into a tiny bag.

Why This is a Game-Changer

No Extra Work: The "magic foam" is added directly to the rocks inside the backpack. You don't need to change the straps or the other layers of the bag. It's a "plug-and-play" fix.
Instant Fix: You don't need to retrain the student or spend weeks adjusting the backpack. The math gives you the exact amount of foam to add in one step (a "closed-form solution").
Super Results: When they tested this on 2-bit compression (making the notes extremely tiny), OSAQ made the model 40% smarter (lower perplexity) than the previous best method. It's like shrinking the backpack to the size of a wallet without the student losing any of their genius.

Summary

The paper introduces a way to make AI models smaller and faster by finding "silent zones" in their data where they can smooth out the problematic "boulders" (outliers) without hurting the model's performance. It's a new, efficient way to compress AI that works better than just stretching or rotating the data.

Technical Summary: OSAQ – Outlier Self-Absorption for Accurate Low-bit LLM Quantization

1. Problem Statement

Large Language Models (LLMs) face significant deployment challenges due to their massive parameter scales, which result in high memory consumption and inference latency. Post-training quantization (PTQ), specifically weight-only quantization, offers a solution by reducing model size and alleviating memory-bound bottlenecks. However, a persistent obstacle to accurate low-bit quantization is the presence of systematic outliers in model weights. These outliers, characterized by large magnitudes, expand the numerical range of weights, thereby reducing quantization resolution and degrading model performance.

Existing methods primarily rely on multiplicative transformations (e.g., scaling or rotation) between adjacent layers to mitigate outliers. While approaches like GPTQ (using Hessian information), AWQ (activation-aware scaling), and QuIP (orthogonal rotation) have made progress, the paper argues that relying solely on multiplicative paradigms is fundamentally inadequate for achieving satisfactory performance in ultra-low-bit settings (e.g., 2-bit or 3-bit).

2. Methodology: Outlier Self-Absorption Quantization (OSAQ)

The authors propose OSAQ, a novel method that introduces an additive weight transformation guided by the second-order low-rank properties of the Hessian matrix. Unlike previous methods that require inter-layer adjustments, OSAQ operates as a "self-absorption" mechanism that modifies weights within a single layer without affecting neighboring layers.

Key Technical Components:

Low-Rank Consistency of the Hessian:
The method is grounded in the observation that the Hessian of the task loss with respect to layer weights exhibits a pronounced low-rank property across different inputs. Specifically, certain directions in the weight space consistently show vanishing curvature (near-zero eigenvalues). These directions form a stable null space that remains consistent regardless of the input sample.
Loss-Invariant Additive Transformation:
Leveraging the null space, OSAQ constructs an additive perturbation $\Delta W$ such that the transformation $W' = W + \Delta W$ guarantees the second-order perturbation of the task loss remains zero ( $\Delta w^T H_w \Delta w = 0$ ). This ensures that the model's performance on the original task is preserved while the weights are modified.
Null Space Extraction:
The Hessian $H_w$ is approximated and decomposed via eigenvalue decomposition. To handle cases where eigenvalues are not exactly zero, a tail-energy strategy is employed. The null space dimension $K$ is determined by accumulating the smallest eigenvalues until a predefined energy threshold $\gamma$ is reached. The resulting matrix $N$ represents the stable null space.
Coefficient Optimization via Softmax- $\infty$ Approximation:
The goal is to find a coefficient matrix $\beta$ such that $\Delta W = \beta N$ minimizes the numerical range (specifically the $\ell_\infty$ norm) of the transformed weights to suppress outliers.
- Since the $\ell_\infty$ norm is non-differentiable, the authors employ a Softmax- $\infty$ approximation. This involves applying a softmax operation along the output-channel dimension with a temperature coefficient $\tau$ . As $\tau \to 0$ , this approximates the $\ell_\infty$ norm while remaining differentiable.
- The optimization objective includes an $\ell_2$ regularization term to prevent excessive corrections and an anti-shift constraint to penalize uniform translations.
- Crucially, this optimization yields a closed-form solution for the coefficients $\beta$ via a normal equation, eliminating the need for resource-intensive training or iterative procedures.

3. Key Contributions

Novel Paradigm: The paper introduces an additive transformation strategy for outlier suppression, complementing existing multiplicative (scaling/rotation) approaches. It operates independently within a layer, requiring no inter-layer transformations or inference overhead.
Theoretical Insight: It identifies and exploits the low-rank consistency of the Hessian's null space across inputs, providing a stable subspace for constructing loss-invariant weight adjustments.
Efficient Construction: The method derives the optimal transformation coefficients in closed form using a differentiable Softmax- $\infty$ approximation, avoiding iterative optimization or retraining.
Plug-and-Play Compatibility: OSAQ is designed to be integrated with existing PTQ methods (e.g., GPTQ, AWQ) to further enhance low-bit performance.

4. Experimental Results

Extensive experiments were conducted on various LLMs (LLaMA2, LLaMA3, Mistral-Large, Llama-3.1) across language generation and commonsense QA tasks.

Perplexity Reduction: In 2-bit quantization, OSAQ integrated with GPTQ achieved over 40% lower perplexity compared to vanilla GPTQ on LLaMA2-7B. For example, on WikiText2, OSAQ+GPTQ reduced perplexity from 36.8 to 21.2 (and further to 10.6 with coordinate descent iterations) for LLaMA2-7B.
Accuracy Improvements:
- Commonsense QA: On LLaMA3-8B (2-bit), OSAQ+GPTQ improved average zero-shot accuracy from 38.3% to 44.4% (and 51.7% with iterations).
- MMLU: For LLaMA2-7B (4-bit), OSAQ+GPTQ raised the average score from 38.7% to 39.6%.
- Large Models: The method remained effective on massive instruction-tuned models (123B and 405B parameters), outperforming vanilla GPTQ on MMLU and ARC benchmarks.
Robustness: Ablation studies confirmed that the additive transformation has a negligible impact on full-precision (FP16) models and that the null space remains stable across different input batches. The method also demonstrated robustness to hyperparameter choices (e.g., tail-energy threshold $\gamma$ , temperature $\tau$ ).

5. Significance and Claims

The paper claims that OSAQ represents a significant step forward in low-bit LLM quantization by addressing the fundamental limitation of multiplicative-only paradigms. By leveraging the intrinsic geometric properties of the loss landscape (Hessian null space), OSAQ provides a computationally efficient, training-free mechanism to suppress outliers.

The authors emphasize that this approach:

Preserves Task Loss: The transformation is mathematically guaranteed to keep the second-order loss perturbation at zero.
Introduces No Overhead: The transformation is absorbed offline into the weights, requiring no changes to the inference pipeline or additional layers.
Scalability: It is applicable to models ranging from 7B to 405B parameters and integrates seamlessly with state-of-the-art quantization algorithms.

The work suggests that exploring additive transformations based on second-order properties offers a promising, complementary route to achieving accurate low-bit quantization for large-scale language models.