From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization

Conference: ACL 2026
arXiv: 2604.19884
Code: None
Area: Model Quantization / Interpretability
Keywords: Post-Training Quantization, Signal Degradation, Computation Collapse, Mechanistic Interpretability, Causal Tracing, Knowledge Recall, PTQ

TL;DR

Through systematic mechanistic interpretability analysis, this paper reveals two qualitatively different failure modes in LLM quantization: signal degradation in 4-bit (complete computational patterns with compromised precision, partially repairable) and computation collapse in 2-bit (functional destruction of key components, requiring structural reconstruction).

Background & Motivation

Background: Post-training quantization (PTQ) is a critical technology for the efficient deployment of LLMs. While 4-bit quantization is widely regarded as the optimal balance between precision and compression, 2-bit quantization typically triggers a catastrophic "performance cliff," where accuracy plummets to near zero.

Limitations of Prior Work: Existing research focuses on three directions: (1) macro-evaluation (measuring the magnitude of performance loss); (2) algorithmic improvement (numerical optimizations such as outlier suppression and rotation matrices); and (3) preliminary mechanistic exploration (layer/component sensitivity analysis). A common limitation is treating quantization damage solely as a "numerical problem" without exploring why internal model mechanisms fail.

Key Challenge: Is the catastrophic failure of 2-bit quantization a cumulative "quantitative change" of 4-bit degradation, or does it represent a qualitative shift? If it is a qualitative shift, it implies that current numerical optimization-based repair strategies are fundamentally misdirected for 2-bit scenarios.

Goal: This study aims to reveal internal mechanism differences in quantization failures through systematic mechanistic interpretability analysis (layer-wise information flow, causal paths, component functions, and representation spaces) and to verify that different failure modes require distinct repair strategies.

Key Insight: Quantization failure is likened to a signal processing problem—is the signal weakened by noise (degradation), or is the computation pipeline itself broken (collapse)?

Core Idea: The failures of 4-bit and 2-bit are essential differences rather than differences in degree. Signal degradation can be recovered through targeted training-free repairs, whereas computation collapse requires structural reconstruction (e.g., fine-tuning). This difference provides the strongest evidence for distinguishing the two modes.

Method

Overall Architecture: Using Llama-3.1-8B as the primary subject, the study systematically compares the internal behaviors of FP16, 4-bit, and 2-bit models on a factual knowledge recall task (Pararel). A five-layer analysis establishes and verifies hypotheses: Macro Phenomena \(\rightarrow\) Layer-wise Probing \(\rightarrow\) Causal Analysis \(\rightarrow\) Component/Representation Verification \(\rightarrow\) Mechanism-oriented Intervention.

Key Designs:

1. Multi-level Knowledge Signal Tracing

   • Function: Traces the existence and causal transmission integrity of knowledge signals within the model.
   • Mechanism: Employs Logit Lens to project hidden states into the vocabulary space layer-by-layer, observing changes in the probability and rank of the correct token. In 4-bit, signals appear in middle-to-late layers but with weakened intensity (degradation); in 2-bit, signals remain near zero (absence). Cross-model causal activation patching further verifies this: injecting "clean" FP16 activations into key positions of the quantized model (last subject token) allows 4-bit models to recover, whereas 2-bit models remain completely non-responsive.
   • Design Motivation: Distinguishing "weakened signals" from "signals never generated" is core evidence for the two-mode hypothesis.

2. Component-level Functional Diagnosis (Attention + FFN Key-Value Memory)

   • Function: Pinpoints which specific components fail and how they fail.
   • Mechanism: Uses normalized entropy (global concentration) + JSD divergence (focus deviation) for attention layers. For FFN layers, it uses the gating Sign Flip Rate (SFR, \(>30\%\) indicates severe instability), Jaccard overlap of Top-1% activated neurons (\(\approx 0.1\) indicates complete activation misalignment), and output cosine similarity (\(\approx 0\) indicates complete semantic deviation). 2-bit quantization shows functional collapse across all indicators.
   • Design Motivation: Attributing macro "signal loss" to specific component failures confirms whether the issue is a loss of precision or a loss of function.

3. Mechanism-Aware Two-Stage Repair vs. System Irreversibility Verification

   • Function: Verifies the fundamental difference in repairability between the two modes.
   • Mechanism: A "Source Protection + Signal Recovery" scheme is designed for 4-bit: protecting the first few layers (8-bit for the first 2 layers in Llama/Mistral, \(\approx 4.25\) avg bits; kurtosis-based selection for Qwen/Gemma, \(\approx 4.1\) avg bits) + peak signal amplification (\(\alpha\)-times logit scale). These strategies and EORA low-rank compensation prove ineffective for 2-bit. A "domino experiment" shows that quantizing only the first 2 layers leads to a drop from \(100\%\) to \(41.65\%\).
   • Design Motivation: Differences in repairability serve as the most direct and powerful practical evidence for distinguishing the two modes.

Key Experimental Results

4-bit Repair Experiments (Accuracy on Failure Subset):

Model	Baseline (4-bit)	+ Basic Repair	+ Signal Amp (Final)
Llama3.1-8B	0.00%	67.91%	75.19% (\(\alpha=3\))
Mistral-7B	0.00%	66.86%	81.26% (\(\alpha=9\))
Qwen3-8B	0.00%	40.24%	79.88% (\(\alpha=7\))
Gemma2-9B	0.00%	33.85%	64.08% (\(\alpha=2\))

2-bit "Domino Effect" (Llama3.1-8B):

Quantized Layers	Robust Subset	Failure Subset
None (FP16)	100.00%	100.00%
Layer 0	65.47%	15.03%
Layers 0-1	41.65%	5.29%
Layers 0-5	2.51%	0.38%

Representation Space Analysis: * 4-bit: CKA maintains a clear diagonal structure, with activation subspace similarity to FP16 \(>0.8\). * 2-bit: CKA is almost entirely dark (structural collapse), with activation subspace similarity \(\approx 0\). * 4-bit error subspace alignment with signals is \(\approx 0.3\) (resembling random noise). * 2-bit error subspace alignment with signals is \(\approx 0.8\) (directly interfering with core features).

Key Findings: * 4-bit results in a "drop in answer rank" (correct answer remains in Top-5), while 2-bit results in "rank collapse" (dropping to thousands, equivalent to random guessing). * Architecture-dependent degradation: Llama/Mistral exhibit an "early representation bottleneck," while Qwen/Gemma show "uniform degradation." * 2-bit models cannot process signals correctly even when receiving high-precision inputs—the components themselves have failed. * The distinction between the two failure modes is consistent across both GPTQ and AWQ methods.

Highlights & Insights

• Valuable Framework for Qualitative Distinction: This is the first systematic proof that 4-bit and 2-bit are not different degrees on the same continuum but two fundamentally different failure modes.
• Closed Loop from Diagnosis to Repair: Mechanistic analysis directly guides the design of repair strategies, and the variance in repair effectiveness further validates the diagnosis.
• Compelling "Domino Experiment": Demonstrating that quantizing just the first two layers in 2-bit leads to catastrophic collapse—unrecoverable by 30 subsequent FP16 layers—visually illustrates the irreversibility of computation collapse.
• Deep Insight into Error Directions: The high alignment of 2-bit quantization error with the signal subspace implies that noise is not random but systematically destroys the model's core features.

Limitations & Future Work

• The study focuses on weight-only quantization; failure modes in activation quantization remain to be explored.
• Evaluations are anchored in factual recall tasks; performance in complex reasoning tasks needs verification.
• Repair strategies require additional precision overhead (\(\approx 4.1-4.25\) avg bits), and their practicality needs optimization.
• The boundary between the two modes (behavior of 3-bit) warrants further research.
• The cutoff point for failure modes may vary across different model architectures.

Related Work & Insights

• GPTQ (Frantar et al., 2023): The most widely used weight-only PTQ method and the primary quantization baseline in this paper.
• Causal Tracing (Meng et al., 2022): A knowledge localization method extended here for cross-model repair experiments.
• Logit Lens (nostalgebraist, 2020): A tool for intermediate layer decoding.
• SpQR (Dettmers et al., 2023): A mixed-precision method echoed by the source protection strategy used in this paper.
• Insight: Quantization research should move beyond numerical optimization; mechanistic understanding is vital for overcoming performance bottlenecks. Making 2-bit practical requires a shift from "compensation" to "reconstruction."

Rating

• Novelty: ★★★★★ — The systematic distinction and verification of two failure modes is a fresh and significant contribution.
• Experimental Thoroughness: ★★★★★ — The evidence chain is complete, covering four models, multi-level analysis, and multiple metric validations.
• Writing Quality: ★★★★★ — The narrative is extremely clear, progressing logically from phenomena to hypothesis, verification, and intervention.
• Value: ★★★★☆ — Provides a critical diagnostic framework and mechanistic insights for future quantization research.

Related Papers

• [ACL 2026] Two-Stage Regularization-Based Structured Pruning for LLMs
• [ICLR 2026] ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference
• [ICLR 2026] Rethinking Continual Learning with Progressive Neural Collapse
• [ICML 2026] WUSH: Near-Optimal Adaptive Transforms for LLM Quantization
• [ICLR 2026] The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm