Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts

arXiv cs.CL Papers

Summary

This paper demonstrates that post-training quantization can silently alter how large language models reason, even when task accuracy is preserved, through a taxonomy-based analysis of 30,000 chain-of-thought outputs across multiple models and benchmarks.

arXiv:2607.09999v1 Announce Type: new Abstract: We show that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Using a six-category failure taxonomy validated by two independent human annotators (Cohen's $\kappa$ = 0.906), we classify 30,000 chain-of-thought outputs from five instruction-tuned LLMs (3B--14B parameters) across three quantization precisions (FP32, FP16, NF4) and four reasoning benchmarks. We find that while accuracy is robust across precisions (maximum 3.1 pp drop), Hollow Convergence (correct answers reached through incomplete or unverifiable reasoning) shows a significant size-dependent shift under NF4, dropping sharply for the two smallest models tested but remaining invariant for models at 12B parameters and above. This effect is also benchmark-specific: GSM8K is categorically immune while LogiQA and ARC-Challenge show the largest shifts. Furthermore, under NF4, Shortcut Collapse rises from 44% to 78% of wrong-answer failures in LLaMA 3.2-3B while Confidence Snowballing collapses from 15.8% to near zero, a qualitative shift invisible to accuracy metrics. Finally, we show Hollow Convergence cannot be reliably detected from surface-level text features (best F1 = 0.53), establishing it as a deployment-relevant failure mode that standard evaluation pipelines cannot catch.
Original Article
View Cached Full Text

Cached at: 07/14/26, 04:21 AM

# Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts
Source: [https://arxiv.org/abs/2607.09999](https://arxiv.org/abs/2607.09999)
[View PDF](https://arxiv.org/pdf/2607.09999)

> Abstract:We show that post\-training quantization can silently alter how large language models reason even when task accuracy is preserved\. Using a six\-category failure taxonomy validated by two independent human annotators \(Cohen's $\\kappa$ = 0\.906\), we classify 30,000 chain\-of\-thought outputs from five instruction\-tuned LLMs \(3B\-\-14B parameters\) across three quantization precisions \(FP32, FP16, NF4\) and four reasoning benchmarks\. We find that while accuracy is robust across precisions \(maximum 3\.1 pp drop\), Hollow Convergence \(correct answers reached through incomplete or unverifiable reasoning\) shows a significant size\-dependent shift under NF4, dropping sharply for the two smallest models tested but remaining invariant for models at 12B parameters and above\. This effect is also benchmark\-specific: GSM8K is categorically immune while LogiQA and ARC\-Challenge show the largest shifts\. Furthermore, under NF4, Shortcut Collapse rises from 44% to 78% of wrong\-answer failures in LLaMA 3\.2\-3B while Confidence Snowballing collapses from 15\.8% to near zero, a qualitative shift invisible to accuracy metrics\. Finally, we show Hollow Convergence cannot be reliably detected from surface\-level text features \(best F1 = 0\.53\), establishing it as a deployment\-relevant failure mode that standard evaluation pipelines cannot catch\.

## Submission history

From: Renuka Oladri \[[view email](https://arxiv.org/show-email/8aee97f3/2607.09999)\] **\[v1\]**Fri, 10 Jul 2026 21:55:05 UTC \(467 KB\)

Similar Articles

Quantized Reasoning Models Think They Need to Think Longer, but They Do Not

arXiv cs.LG

This paper reveals that aggressive post-training quantization of reasoning models leads to increased overthinking errors, where models reach correct intermediate answers but fail to finalize them. A simple logit penalty on overthinking markers reduces chain-of-thought length by 12-23% while improving accuracy, especially for quantized models.

Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations

arXiv cs.CL

This paper presents a comprehensive empirical evaluation of how large language models handle corruptions in chain-of-thought reasoning steps, testing 13 models across 5 perturbation types (MathError, UnitConversion, Sycophancy, SkippedSteps, ExtraSteps) on mathematical reasoning tasks. The findings reveal heterogeneous vulnerability patterns with implications for deploying LLMs in multi-stage reasoning pipelines.