Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts
Post-training quantization, specifically to NF4, can silently degrade reasoning quality in LLMs without significantly impacting standard accuracy metrics. The study identifies "Hollow Convergence," where models produce correct answers via incomplete or unverifiable reasoning, a failure mode undetectable by surface-level text analysis. Failure modes are highly dependent on model size and benchmark type; smaller models (3B) show sharp declines in reasoning integrity under NF4, while larger models
Analysis
TL;DR
- Post-training quantization, specifically to NF4, can silently degrade reasoning quality in LLMs without significantly impacting standard accuracy metrics.
- The study identifies "Hollow Convergence," where models produce correct answers via incomplete or unverifiable reasoning, a failure mode undetectable by surface-level text analysis.
- Failure modes are highly dependent on model size and benchmark type; smaller models (3B) show sharp declines in reasoning integrity under NF4, while larger models (12B+) remain invariant.
- Specific qualitative shifts include a rise in "Shortcut Collapse" and a collapse in "Confidence Snowballing" for smaller quantized models, indicating a loss of robust reasoning strategies.
Why It Matters
This research highlights a critical blind spot in current AI evaluation practices: high accuracy scores do not guarantee reliable or safe reasoning, particularly when deploying quantized models for cost efficiency. For practitioners, it underscores the necessity of moving beyond simple accuracy metrics to include deeper semantic and logical validation, especially for smaller models where quantization artifacts may fundamentally alter reasoning pathways.
Technical Details
- Methodology: Analyzed 30,000 chain-of-thought outputs from five instruction-tuned LLMs (3B–14B parameters) across FP32, FP16, and NF4 precisions using a validated six-category failure taxonomy.
- Key Findings on Model Size: Under NF4 quantization, "Hollow Convergence" drops sharply for 3B and smaller models but remains stable for models with 12B parameters or more.
- Benchmark Specificity: Effects vary by domain; GSM8K was immune to these shifts, whereas LogiQA and ARC-Challenge exhibited the most significant degradation in reasoning integrity.
- Failure Mode Shifts: In LLaMA 3.2-3B under NF4, "Shortcut Collapse" increased from 44% to 78% of wrong-answer failures, while "Confidence Snowballing" dropped from 15.8% to near zero.
- Detection Limitations: "Hollow Convergence" could not be reliably detected using surface-level text features, achieving a best F1 score of only 0.53.
Industry Insight
- Rethink Evaluation Metrics: Organizations relying solely on accuracy benchmarks for quantized model deployment risk integrating models with fragile or deceptive reasoning capabilities; implement logical consistency checks alongside accuracy tests.
- Size-Dependent Quantization Risks: Smaller models (under 10B parameters) are disproportionately vulnerable to reasoning degradation under aggressive quantization like NF4; consider retaining higher precision for these sizes in safety-critical applications.
- Domain-Specific Validation: Quantization impacts vary by task complexity; rigorous testing should prioritize logic-heavy benchmarks (like LogiQA) over purely mathematical ones (like GSM8K) to uncover silent failures.
Disclaimer: The above content is generated by AI and is for reference only.