| name | illusion-of-equivalency-quantization-effects-llms |
| description | Shows that post-training quantization evaluation via accuracy/perplexity fails to capture behavioral changes. Introduces correctness agreement metric. Reveals non-linear breakpoints at low bit-widths. Query/key projections more sensitive than value/output. Activation: quantization, LLM deployment, behavioral change, correctness agreement, post-training quantization. |
| metadata | {"arxiv_id":"2607.08734","published":"2026-07-09","authors":"Baha Rababah, Cuneyt Gurcan Akcora, Carson K. Leung","tags":["quantization","llm-deployment","behavioral-change","correctness-agreement","post-training-quantization"]} |
The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs
Overview
Post-training quantization is widely used to deploy LLMs in resource-constrained settings, yet its evaluation relies almost exclusively on accuracy and perplexity. This paper shows that these metrics fail to capture behavioral changes induced by quantization, introducing "correctness agreement" as a decision-level metric that reveals the illusion of equivalency between base and quantized models.
Key Innovations
Correctness Agreement Metric
- Decision-level metric measuring overlap in correct predictions between base and quantized models
- Independent of absolute accuracy
- Reveals behavioral divergence even when task performance appears preserved
Behavioral Divergence Under Quantization
- Divergence emerges under moderate quantization (not just extreme)
- Multiple models and quantization schemes tested (8-bit to 2-bit)
- Standard metrics (accuracy, perplexity) miss these behavioral changes
Layer-Wise Sensitivity Analysis
- Quantization analyzed as structural operator on attention weights
- Statistical and distributional measures of layer-wise distortions
- Non-linear breakpoints identified at low bit-widths
- Query and key projections consistently more sensitive than value and output projections
Methodology
- Quantization: Apply multiple schemes from 8-bit to 2-bit
- Correctness Agreement: Measure prediction overlap between base and quantized
- Layer Analysis: Quantify distortions per layer using statistical measures
- Sensitivity Ranking: Compare Q, K, V, O projection sensitivity
- Breakpoint Detection: Identify non-linear behavioral change points
Implications
- Accuracy and perplexity are insufficient for quantization evaluation
- Behavioral evaluation reveals hidden degradation in quantized models
- Layer-wise sensitivity analysis guides targeted quantization strategies
- Protecting Q/K projections while relaxing V/O can preserve behavior
Pitfalls
- Correctness agreement is binary and may miss nuanced behavioral differences
- Analysis limited to attention weights — other components (FFN, embeddings) not analyzed
- Specific models and quantization schemes may not generalize
- No proposed mitigation strategy — purely diagnostic
Activation Keywords
quantization, LLM deployment, behavioral change, correctness agreement, post-training quantization, attention weight distortion, layer sensitivity, Q/K projections, evaluation metrics
Paper Reference
arXiv:2607.08734 - "The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs" (Jul 2026)