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

Dimensions overall score 7.0

• Large language models exhibit substantial answer-flip rates (28.8%-45.1%) when presented with meaning-preserving perturbations of mathematical reasoning problems.result 95%
  Evidencepartial

  All three models exhibit substantial answer-flip rates (28.8%-45.1%)

  Implicationpartial

  Directly stated in abstract with specific numeric range for three models

  Verificationpartial

  partial

• Number format paraphrasing is consistently more disruptive to LLM reasoning than name substitution perturbations.result 90%
  Evidencepartial

  number paraphrasing consistently more disruptive than name swaps

  Implicationpartial

  Explicitly stated in abstract with clear comparative language

  Verificationpartial

  partial

• The Cascading Amplification Index (CAI) outperforms first divergence layer as a failure predictor for two of three architectures, achieving AUC up to 0.679.result 92%
  Evidencepartial

  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)

  Implicationpartial

  Directly stated with specific metric (AUC) and clear comparison

  Verificationpartial

  partial

• Logit lens analysis reveals that flipped samples diverge from correct predictions at significantly earlier layers than stable samples.result 88%
  Evidencepartial

  Logit lens reveals that flipped samples diverge from correct predictions at significantly earlier layers than stable samples

  Implicationpartial

  Directly stated finding from specific analysis method

  Verificationpartial

  partial

• Activation patching reveals architectural differences 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).result 93%
  Evidencepartial

  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)

  Implicationpartial

  Directly stated with specific numeric results for each model

  Verificationpartial

  partial

• The paper introduces the Mechanistic Perturbation Diagnostics (MPD) framework combining logit lens analysis, activation patching, component ablation, and CAI into a unified diagnostic pipeline.method 90%
  Evidencepartial

  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

  Implicationpartial

  Explicitly stated as a methodological contribution

  Verificationpartial

  partial

• Targeted repair experiments show 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.result 91%
  Evidencepartial

  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

  Implicationpartial

  Directly stated with specific recovery percentages for each failure type

  Verificationpartial

  partial

• The paper proposes a mechanistic failure taxonomy with three categories: localized, distributed, and entangled failures.method 85%
  Evidencepartial

  propose a mechanistic failure taxonomy (localized, distributed, and entangled)

  Implicationpartial

  Explicitly stated as a proposed taxonomy based on diagnostic signals

  Verificationpartial

  partial

Competitive landscape

This research develops a framework to diagnose and fix the surprising fragility of large language models to meaning-preserving text changes, offering a path to more reliable AI.