IC-278Claude-3.5 Sonnet and GPT-4o exhibit a memorization failure mode, outputting the same answer regardless of visual parameter changes in the problem

Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, Huan Zhang

SourceDynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models

The paper identifies a phenomenon where VLMs provide the same answer to structurally identical questions with different visual parameters, indicating reliance on a memorized default rather than reading the specific input. For Claude-3.5 Sonnet, altering the period of a sinusoidal function (e.g., from 2π to π or 4π) does not change its response, which consistently remains 2π; this accounts for 4.1% of its 171 incorrect answers. For GPT-4o, a shifted absolute value function f(x)=|x-a| is consistently judged not differentiable at x=0 for 7 of 10 variants, with 90% repetition consistency on the failing variant, despite the sharp corner no longer being at x=0.

Evidence
correlational
Key metric
Claude-3.5 Sonnet: 4.1% of 171 incorrect answers are memorization cases. GPT-4o: 90% repetition consistency on variant 9 of seed question 78; incorrect answer for 7 of 10 variants of the shifted absolute value function.
Caveat
The memorization phenomenon was identified qualitatively and the 4.1% figure is for Claude-3.5 only; a systematic study across all models and all question types is noted as future work. A newly designed prompt explicitly instructing the model to avoid memorized answers did not resolve the issue.
Model
Claude 3.5 Sonnet, GPT-4o
Concepts
Shortcut, Failure mode
Datasets
DynaMath [eval]
Related findings
IC-276, IC-277
Extraction
automatic-extraction