Prompt Priming Masquerading as Visual Attribution

F-02 · measurementBenchmark ablation.

Question
An apparent visual-misattribution effect in detector-grounded VLM benchmarks.
Method
Marker × prompt ablation isolating the visual channel from the text channel.
Result
The effect was prompt-text priming. Wording shifted results 9.4pp; visual markers ≤0.2pp.
Implication
Benchmark conclusions can flip when input channels are not isolated.
prompt text 9.4pp visual marker ≤0.2pp
Shift in benchmark outcome, in percentage points, by input channel.