Diagnosing Corruption-Induced Reliability Failures in Vision-Language Models

Technical Summary: Diagnosing Corruption-Induced Reliability Failures in Vision-Language Models

Problem Statement
Current evaluations of Vision-Language Models (VLMs) under visual corruption rely heavily on top-1 accuracy. However, the authors argue that accuracy is an insufficient metric for reliability because it fails to capture changes in the underlying prediction structure. A model may maintain the correct answer while losing distributional support (silent degradation), or it may switch from a wrong to a correct answer through unstable, overconfident shifts. Furthermore, standard benchmarks often include low-discriminative samples where answers remain stable regardless of corruption, masking model differences. The paper identifies a "visual quality paradox" where mild corruptions sometimes improve top-1 accuracy, yet it remains unclear whether these gains reflect reliable correction or unstable behavior.

Methodology
To address these gaps, the authors propose BENCH-C, a controlled diagnostic testbed, and the Robustness Alignment Score (RAS), a metric for measuring distribution-level reliability.

1. BENCH-C Construction:

   • Candidate Pool: The authors aggregate 4,000 samples from six diverse multiple-choice benchmarks (e.g., SEEDBENCH2+, MMBENCH) covering text-rich, spatial, and fine-grained reasoning tasks.
   • Corruption Protocol: They apply 19 single-factor corruptions (grouped into photometric, blur, noise, weather, and digital artifacts) across five ordered severity levels, following the ImageNet-C convention.
   • Diagnostic Selection: Instead of random sampling, BENCH-C selects samples that are sensitive to corruption. It filters for samples where predictions vary across severity levels (measured by Gini impurity of the answer trajectory) and ensures semantic diversity via greedy selection on image and text embeddings.
   • Final Dataset: The resulting testbed contains 849 diagnostic questions and 80,655 corrupted image-question pairs.

2. Prediction-Structure Diagnostics:

   • The authors analyze the option-level probability distribution $p$ rather than just the argmax answer.
   • Uncertainty Shift ($\Delta S$): Measured as the change in normalized entropy, indicating whether the distribution becomes flatter (less decisive) or sharper.
   • Confidence-Correctness Alignment Shift ($\Delta C$): Measured as the change in the absolute difference between the model's top-1 confidence and the correctness of that prediction.
   • Robustness Alignment Score (RAS): A composite metric defined as $RAS = -\Delta C - [\Delta C]^+[-\Delta S]^+ - [-\Delta S]^+[\Delta C]^+$. This score penalizes "overconfident degradation" (worse alignment with lower entropy) and "hesitant support" (better alignment with higher entropy), while rewarding improved alignment. RAS ranges from -2 to 1.

3. Attribution Framework:

   • Source Attribution: Changes are attributed to "originally correct" ($Q_c$) vs. "originally wrong" ($Q_w$) samples.
   • State Attribution: Samples are further categorized into four states based on clean vs. corrupted correctness: $Q_{cc}$ (correct $\to$ correct), $Q_{cw}$ (correct $\to$ wrong), $Q_{wc}$ (wrong $\to$ correct), and $Q_{ww}$ (wrong $\to$ wrong).
   • Trajectory Analysis: The authors track whether changes are transient or persistent across the five severity levels.

Key Results
Experiments on 13 VLMs reveal that corruption-induced reliability failures are more structured than accuracy suggests:

• Accuracy vs. Structure Mismatch: High clean accuracy does not guarantee robustness. For instance, Qwen2.5-VL had the highest clean accuracy but the largest accuracy loss and most negative RAS. Conversely, models with lower clean accuracy sometimes exhibited better structural robustness (RAS).
• Hidden Degradation: On originally correct samples ($Q_{cc}$), models often maintained the correct answer but lost distributional support (negative RAS), indicating silent degradation.
• Counterintuitive Gains: Apparent accuracy gains often stem from "wrong-to-correct" transitions ($Q_{wc}$). However, these gains frequently coexist with significant degradation on originally correct samples. The paper notes that stronger corruptions can make originally wrong answers appear more correct (positive structural shift) while simultaneously destabilizing originally correct predictions.
• Severity Trends: As corruption severity increases, the damage to originally correct samples becomes more negative, while the structural improvement on originally wrong samples becomes more positive. This suggests that severe corruption induces prediction-structure instability rather than robust correction.
• Corruption Types: Blur was identified as particularly damaging to originally correct samples, likely due to the loss of fine-grained contours and local grounding cues.

Significance and Claims
The paper claims that BENCH-C and RAS provide a necessary shift from evaluating "final answers" to evaluating "prediction structure." The significance lies in:

1. Revealing Hidden Failures: Identifying that models can maintain correct answers while losing reliability (silent degradation) or improve accuracy through unstable mechanisms.
2. Attributing Reliability Changes: Distinguishing between damage to reliable behaviors and corrections of unreliable ones, showing that aggregate accuracy gains can mask underlying structural costs.
3. Guiding Robustness Training: The authors suggest that robustness training should not merely reward answer invariance but should aim to preserve confidence-correctness alignment and distributional support, particularly for fine-grained visual evidence.

The work concludes that robustness evaluation must move beyond aggregate metrics to include source, state, and severity-trajectory analyses to accurately diagnose where and how reliability changes occur under visual degradation.