#vision-language models
Vision-Language Models: 12 AI articles covering vision-language models news, analysis, and research
Articles
What Entropy-Based View Selection Cannot See: An Audit of TestNEWβ9
This audit of test-time adaptation for vision-language models shows entropy-based view selection adds little accuracy gain, ranking correct views no better than...
Subliminal Learning in Vision-Language Models: Hiddenβ8
Subliminal learning transfers hidden preferences between vision-language models via unrelated training data, altering image choices and explanations, with effec...
Batch-Conditioned Semantic Anchors Enable Robust Transductiveβ9
JPTA is a prior-aware framework for robust transductive adaptation of visionβlanguage models via batch-conditioned semantic reference estimation.
BD-HazardVLM: New Benchmark Exposes Vision-Language Models' Weakβ9
Vision-language models fail at defensive-driving reasoning in Bangladesh road scenes, a new benchmark shows, with accuracy near majority-class baselines and sev...
Improving OCR Faithfulness via Gated and Attenuated On-Policyβ8
GAD-RL adaptively gates and attenuates teacher supervision during on-policy distillation, improving OCR faithfulness and achieving strong CHAOS-Bench and OmniDo...
Liquid AI Releases LFM2.5-VL-3B-DSpark: Speculative Decoding forβ8
Liquid AI releases LFM2.5-VL-3B-DSpark, a speculative decoding drafter for its vision-language model, delivering up to 3.13x faster decoding on Apple silicon, w...
Feature Recovery for Object Understanding After Irreversibleβ7
TRACE benchmark and Feature Recovery Module improve post-fire object detection and identification when irreversible damage degrades vision model performance.
Evidence-Order Calibration for Selective Visual Reasoning underβ9
A lightweight reliability head with evidence-order supervision improves selective visual reasoning in VLMs, reducing monotonicity violations under progressive m...
FailSAE: Interpretable Failure Prediction for Vision-Languageβ9
Discover how sparse autoencoders enable interpretable failure prediction in vision-language models, revealing why errors occur for risk-aware AI decisions.
Taming Visual Neglect: A Variational Information Bottleneckβ9
New framework reveals when visual context helps or hurts multimodal in-context learning, with adaptive attention boosting accuracy by 4.7%.