Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations

• URL: http://arxiv.org/abs/2604.01639v1
• Date: Thu, 02 Apr 2026 05:30:20 GMT
• Title: Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations
• Abstract summary: Large language models demonstrate strong performance on mathematical reasoning benchmarks, yet remain surprisingly fragile to meaning-preserving surface perturbations.<n>We systematically evaluate three open-weight LLMs, Mistral-7B, Llama-3-8B, and Qwen2.5-7B, on 677 GSM8K problems paired with semantically equivalent variants.<n>All three models exhibit substantial answer-flip rates (28.8%-45.1%), with number paraphrasing consistently more disruptive than name swaps.
• Score: 2.7946918847372277
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models demonstrate strong performance on mathematical reasoning benchmarks, yet remain surprisingly fragile to meaning-preserving surface perturbations. We systematically evaluate three open-weight LLMs, Mistral-7B, Llama-3-8B, and Qwen2.5-7B, on 677 GSM8K problems paired with semantically equivalent variants generated through name substitution and number format paraphrasing. All three models exhibit substantial answer-flip rates (28.8%-45.1%), with number paraphrasing consistently more disruptive than name swaps. To trace the mechanistic basis of these failures, we introduce the Mechanistic Perturbation Diagnostics (MPD) framework, combining logit lens analysis, activation patching, component ablation, and the Cascading Amplification Index (CAI) into a unified diagnostic pipeline. CAI, a novel metric quantifying layer-wise divergence amplification, outperforms first divergence layer as a failure predictor for two of three architectures (AUC up to 0.679). Logit lens reveals that flipped samples diverge from correct predictions at significantly earlier layers than stable samples. Activation patching reveals a stark architectural divide in failure localizability: Llama-3 failures are recoverable by patching at specific layers (43/60 samples), while Mistral and Qwen failures are broadly distributed (3/60 and 0/60). Based on these diagnostic signals, we propose a mechanistic failure taxonomy (localized, distributed, and entangled) and validate it through targeted repair experiments: steering vectors and layer fine-tuning recover 12.2% of localized failures (Llama-3) but only 7.2% of entangled (Qwen) and 5.2% of distributed (Mistral) failures.

Related papers

• Sensitivity, Causality, and Repair Dissociate: A Layer-Wise Analysis of Perturbation Robustness and Its Scaling [0.0]
  We show that when a language model fails on surface-perturbed input, "which layer is responsible" has three natural operationalizations.<n>We identify two propagation regimes - spike-and-suppress and late-accumulation.<n>A fixed-harness layer sweep across four models (3.8-8B) confirms the core prediction on chain-of-thought GSM8K.
  arXiv  Detail & Related papers  (2026-08-04T15:48:55Z)
• Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts [0.0]
  We show that post-training quantization can silently alter how large language models reason even when task accuracy is preserved.<n>We classify 30,000 chain-of-thought outputs from five instruction-tuned LLMs across three quantization precisions.<n>We find that while accuracy is robust across precisions, Hollow Convergence shows a significant size-dependent shift under NF4.
  arXiv  Detail & Related papers  (2026-07-10T21:55:05Z)
• Modeling Normal Is All You Need: Joint Latent Clustering for Anomaly Detection in Multimodal Cyber-Physical Systems [47.027290803102666]
  Normal behaviour is the union of many imbalanced, curved, thin-fringed operating regimes.<n>We model the normal law with a jointly learned latent representation plus explicit Gaussian-mixture mode clustering.<n>We evaluate under a deliberately fair protocol: raw point-wise metrics with no point adjustment, a trivial-detector difficulty split, prevalence-matched F1, and train-normal-only calibration.
  arXiv  Detail & Related papers  (2026-07-07T10:10:00Z)
• Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models [2.1291393112295935]
  We study how correctness is assembled inside aligned language models.<n>Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip.<n>We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B)
  arXiv  Detail & Related papers  (2026-07-06T03:51:24Z)
• LLM Doesn't Know What It Doesn't Know: Detecting Epistemic Blind Spots via Cross-Model Attribution Divergence on Clinical Tabular Data [2.1443570696048906]
  Large language models (LLMs) are increasingly applied to structured clinical data.<n>We compare Qwen 2.5 7B and XGBoost on a prediction task via attribution divergence analysis.
  arXiv  Detail & Related papers  (2026-06-17T18:49:44Z)
• Rethinking Molecular OOD Generalization via Target-Aware Source Selection [10.102858221967791]
  SCOPE-BENCH is a benchmark built on cluster-level partitioning in an explicit physicochemical descriptor space.<n>POMA is a framework that formulates knowledge transfer as a retrieve-compose-adapt pipeline.<n> Evaluations show that prediction errors of state-of-the-art 3D molecular models surge by up to 8.0x on SCOPE-BENCH with a mean of 5.9x.
  arXiv  Detail & Related papers  (2026-05-13T16:09:46Z)
• Anatomical Heterogeneity in Transformer Language Models [0.0]
  Current transformer language models are trained with uniform computational budgets across all layers.<n>We challenge this assumption through empirical analysis of SmolLM2-135M.
  arXiv  Detail & Related papers  (2026-03-19T16:59:17Z)
• Semantic Invariance in Agentic AI [2.7821684674538347]
  Large Language Models increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems.<n> deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations.<n>Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension.
  arXiv  Detail & Related papers  (2026-03-13T17:08:44Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Breaking Semantic Hegemony: Decoupling Principal and Residual Subspaces for Generalized OOD Detection [10.596067236901968]
  We investigate the Simplicity Paradox in state-of-the-art (SOTA) models.<n>Models exhibit keen sensitivity in distinguishing semantically subtle OOD samples but suffer from severe Geometric Blindness.<n>We propose D-KNN, a training-free, plug-and-play geometric decoupling framework.
  arXiv  Detail & Related papers  (2026-02-05T06:32:33Z)
• Scaling Trends for Multi-Hop Contextual Reasoning in Mid-Scale Language Models [0.0]
  We present a controlled study of multi-hop contextual reasoning in large language models.<n>We show that multi-agent systems show the inverse pattern, achieving up to 80% on reasoning tasks where rule-based methods fail.
  arXiv  Detail & Related papers  (2026-01-06T20:18:55Z)
• Think Before You Prune: Self-Reflective Structured Pruning for Reasoning Language Models [31.422773877490613]
  Reasoning LLMs (RLMs) deliver strong multi-step reasoning through chain-of-thought generation.<n>RLMs' large model sizes and lengthy decode-time outputs make them costly to deploy and unsuitable for resource-constrained settings.<n>We introduce RESP, a structured pruning framework that aligns pruning decisions with the model's reasoning dynamics.
  arXiv  Detail & Related papers  (2025-12-01T20:27:05Z)
• ShortcutBreaker: Low-Rank Noisy Bottleneck with Global Perturbation Attention for Multi-Class Unsupervised Anomaly Detection [59.89803740308262]
  ShortcutBreaker is a novel unified feature-reconstruction framework for MUAD tasks.<n>It features two key innovations to address the issue of shortcuts.<n>The proposed method achieves a remarkable image-level AUROC of 99.8%, 98.9%, 90.6%, and 87.8% on four datasets.
  arXiv  Detail & Related papers  (2025-10-21T06:51:30Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• Seeing What's Not There: Spurious Correlation in Multimodal LLMs [47.651861502104715]
  We introduce SpurLens, a pipeline that automatically identifies spurious visual cues without human supervision.<n>Our findings reveal that spurious correlations cause two major failure modes in Multimodal Large Language Models (MLLMs)<n>By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.
  arXiv  Detail & Related papers  (2025-03-11T20:53:00Z)
• Benchmarking Reasoning Robustness in Large Language Models [76.79744000300363]
  We find significant performance degradation on novel or incomplete data.<n>These findings highlight the reliance on recall over rigorous logical inference.<n>This paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps.
  arXiv  Detail & Related papers  (2025-03-06T15:36:06Z)
• "Knowing When You Don't Know": A Multilingual Relevance Assessment Dataset for Robust Retrieval-Augmented Generation [90.09260023184932]
  Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) output by leveraging external knowledge sources to reduce factual hallucinations. NoMIRACL is a human-annotated dataset for evaluating LLM robustness in RAG across 18 typologically diverse languages. We measure relevance assessment using: (i) hallucination rate, measuring model tendency to hallucinate, when the answer is not present in passages in the non-relevant subset, and (ii) error rate, measuring model inaccuracy to recognize relevant passages in the relevant subset.
  arXiv  Detail & Related papers  (2023-12-18T17:18:04Z)
• ERNIE-SPARSE: Learning Hierarchical Efficient Transformer Through Regularized Self-Attention [48.697458429460184]
  Two factors, information bottleneck sensitivity and inconsistency between different attention topologies, could affect the performance of the Sparse Transformer. This paper proposes a well-designed model named ERNIE-Sparse. It consists of two distinctive parts: (i) Hierarchical Sparse Transformer (HST) to sequentially unify local and global information, and (ii) Self-Attention Regularization (SAR) to minimize the distance for transformers with different attention topologies.
  arXiv  Detail & Related papers  (2022-03-23T08:47:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.