IC-416Natural language prompts can steer the texture/shape bias in VLMs in both directions without significantly affecting accuracy, with texture-biased prompts more effective than shape-biased ones; this steering also generalizes to low/high-frequency bias.
Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Muhammad Jehanzeb Mirza, Margret Keuper, Janis Keuper
The paper tests hand-crafted and automatically generated prompts that instruct VLMs to focus on shape or texture. Texture-biased prompts shift the bias more than shape-biased ones, suggesting models default to shape but can be pushed toward texture. Automated prompt search using an LLM as optimizer achieves a 23.3% delta between extremes for InternVL-Chat 1.1. Synonym tests confirm the steering leverages a learned representation rather than being coincidental. The effect is not limited to texture/shape: on a new frequency-cue-conflict dataset, prompts produce statistically significant changes in low-frequency bias for both InternVL-Chat 1.1 and LLaVA-Next 7B. Larger LLMs (InternVL 1.2+ 34B vs 1.1 13B) do not show more steerability.
Evidence
correlational
Key metric
InternVL-Chat 1.1 automated prompt search delta 23.3% between extremes; frequency bias: InternVL-Chat 1.1 LF bias 34.5% neutral, 38.6% search(max), 32.9% search(min); LLaVA-Next 7B LF bias 52.4% neutral, 54.5% search(max), 48.7% search(min); accuracy changes are small or sometimes positive
Caveat
The optimization is done on the cue-conflict test set as a proof of concept, not to claim SOTA shape bias. The authors note that other prompts may have led to better results, and that prompting is a potential source of bias in the study. The effect does not fundamentally change the utilized cue.