John Galeotti

Gautam Rajendrakumar Gare, Jia Shi, Zhiqiu Lin et al.

A popular route to interpretable zero-shot classification asks a large language model (LLM) to describe each class name and prompts CLIP with the resulting descriptors. We show that these descriptors carry little visual evidence of their own: removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5%. The diagnosis is that the descriptors are conditioned on the label rather than on the images, so they describe the concept in general and mislead exactly when the data shifts; an LLM insists that strawberries are red, but every strawberry in ImageNet-Sketch is a colorless line drawing. We therefore select attributes from the target image collection instead: we score a large attribute pool against the images in CLIP's joint embedding space and keep the top-scoring attributes per class. Selected this way, class-name-free attribute prompts reach 23.8% on ImageNet (against 15.5% for LLM descriptors), the gain holds on four shifted ImageNet variants, and reselecting from the LLM's own pool isolates the selection mechanism as the cause. With one image per class, the selected attributes outperform the prompt-tuning method CoOp by 3 points while fitting in under a minute instead of 14 hours, with no learned soft prompt to obscure the decision. Because the attribute set is chosen by the data, it doubles as a readable summary of a dataset, which we use to describe distribution shift in words.

Rakshanda Hassan Abhinandan, John Galeotti, Deva Ramanan et al.

Where should the question go in a vision-language model (VLM) prompt: before the image or after it? Intuition says before: knowing what is asked should tell the model where to look. Yet across visual question answering benchmarks, question-first prompting consistently underperforms the image-first ordering recommended for frontier VLMs, a phenomenon we term the question-first paradox. We trace the paradox to a conflict between two stages of VLM computation. Logit-lens and attention probes show the intuition is half right: a question placed before the image genuinely steers perception, moving image patch representations toward question-relevant concepts. The failure lies downstream. Stranded behind hundreds of image tokens, the question is barely attended by the answer token, which instead commits to image-driven (often wrong) answers; a causal attention knockout confirms that the answer reads the question only when the question follows the image. The diagnosis yields a training-free fix: question echoing, restating the question on both sides of the image so that one copy steers perception while the other is read out at answer time. The same division of labor appears in a fifty-year-old finding on human ``adjunct questions'', where repeating a question before and after a passage aids comprehension more than either position alone. Echoing the image as well brings further gains, restoring the whole-image view a causal decoder otherwise loses. The paradox holds across five open VLMs, costing up to 17.5 group-accuracy points. Echoed prompts close it and surpass the best single-pass ordering on NaturalBench, POPE, Winoground, and open-ended VQAv2, by up to 19 Winoground group-accuracy points, with no training, fine-tuning, or architecture change. The paradox reveals a trade-off between steering perception and preserving question access; echoing resolves it through prompt design alone.