Vision-Language Models Suppress Female Representations Under Ambiguous Input

The more pointed finding here is not that VLMs are biased, which is well-documented, but that standard alignment techniques appear to suppress bias conditionally: they work when the model is explicitly cued to apply them, and fail silently when no such cue exists. That conditional suppression is arguably worse than consistent bias, because it creates a false sense of safety in evaluation settings that don't reflect deployment conditions.

The interpretability angle connects directly to 'What Am I Missing? Question-Answering as Hidden State Probing' from the same day, which also argues that internal model states encode information that surface outputs conceal. Both papers are pushing toward the same methodological claim: you cannot trust what a model says about itself, you have to probe what it has actually encoded. That framing also rhymes with the hate speech paper ('Disagreeing Rationales'), which found that evaluation metrics can hide whether models learn robust reasoning or just pattern-match to majority-vote outputs. Taken together, these three papers from the same week suggest a growing consensus that output-level evaluation is structurally insufficient for alignment work.

Watch whether LALS gets adopted as an audit tool in any major model card or third-party evaluation suite within the next six months. If it does, that would confirm the field is moving toward activation-level auditing as a standard rather than a research curiosity. If it stays confined to citations, the methodology may be sound but too costly to operationalize at scale.