Text-steerable visual encoders condition a frozen Vision Transformer on a natural-language prompt and point at the matching image region, producing prompt-conditioned heatmaps and features. We audit t...
arXiv:2606. 14703v1 Announce Type: cross Abstract: How a vision-language model internally solves the task of describing an image is far from obvious.
The paper introduces a perception interface that separates vision from language in vision‑language models. A frozen perception stack detects objects, a deterministic semantic serializer converts the perceived state into text, and a standard text‑only large language model (LLM) answers questions. Experiments on a campus‑robot benchmark show that this serialized interface outperforms a zero‑shot VLM of the same language‑model size, especially as the language model shrinks, and that the advantage persists under paraphrase and different supervision regimes.
FairLens is a benchmark and evaluation framework that measures fairness and validity of vision‑language models (VLMs) in high‑stakes domains such as hiring, legal, and healthcare. It uses over 100,000 face‑image and question pairs covering gender, race, and age, and assesses responses through demographic parity, soundness, demographic association, and bias in free‑text generation. The study finds that VLMs often make unwarranted inferences from faces rather than abstaining, especially in legal and healthcare contexts, and that small parity gaps can still hide unsafe treatment across groups.
The paper investigates how different design choices in visual-head discovery scores affect the attention heads selected in vision‑language models. By systematically varying query type, discovery corpus, and attention source in Qwen3‑VL‑8B‑Instruct, the authors find that the attention source and query type have a larger impact on head selection than the corpus. While necessity scores distinguish discovery settings, steerability does not show a significant advantage over random head sets.
The paper introduces an explainable pedestrian routing method that estimates perceived safety by first generating a natural‑language caption from street‑level images and then deriving risk scores solely from structured features of that caption. Benchmarking nine captioning setups against a CLIP image‑embedding baseline shows comparable performance, and the system was deployed on over 650,000 images across 36 wards in Manchester and Huddersfield. Field validation with 3,669 ratings from 70 participants revealed a modest but statistically significant correlation (r = 0.262) with human judgments, while a stronger supervised benchmark did not translate into better real‑world performance.
IMPLICIT-Bench is a benchmark of 1,831 prompt triplets that tests whether text‑to‑image models can override stereotypical defaults when given explicit requests. Each triplet contains a neutral prompt, a stereotype prompt, and an anti‑stereotype prompt, allowing measurement of a model’s default bias and its ability to follow contradictory instructions. Experiments on Qwen‑Image and Stable Diffusion 3 show that models often default to stereotypes, succeed less often on anti‑stereotype requests, and that prompt rewriting or embedding steering can modestly reduce bias when the relevant stereotype is known.
The paper introduces Shift‑Reliance Alignment (SRA), a method for ranking vision‑language models (VLMs) based on how much they rely on specific cues such as scene context or text. By editing source images to mislead these cues and measuring model responses, SRA predicts which VLMs will be misled by physical sticker attacks and which will hallucinate on unedited photographs. In pre‑registered tests, SRA selects models that reduce error on challenging datasets like Objects365 and OpenImages, outperforming selection by source error and model confidence on the CueDial benchmark.
arXiv:2608.16081v2 Announce Type: replace Abstract: Open-weight and frontier vision-language models (VLMs) perform well on general image understanding, but their ability to interpret fine-grained han...
The paper introduces EmoNet‑Face‑HQ, a benchmark of 2,500 AI‑generated portraits each evaluated by four psychology experts across 40 emotion categories. It shows that most vision‑language models (VLMs) perform no better than random guessing unless fine‑tuned, but that changing the prompt to a simple yes/no per category can raise a VLM’s agreement with experts from chance to near expert levels. When evaluated using probability‑of‑yes scores on the five most‑agreed‑upon categories, eleven open‑weight VLMs achieve a quadratic‑weighted κ of 0.425, below the 0.468 two‑expert agreement, yet their average probability‑of‑yes score (0.553) matches the benchmark’s fine‑tuned model and outperforms it in three cases without any training.
arXiv:2610.15855v1 Announce Type: new Abstract: We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a larg...
arXiv:2608.28609v1 Announce Type: cross Abstract: A personalized agent needs a user memory: a persistent model of who its user is. Today it is almost always text -- transcripts and captions retrieved...