Vision-language models confidently describe things that aren't in the image. A new paper explains why, and offers two fixes that don't require retraining.
Researchers investigated the root cause of hallucinations in decoder-based vision-language models — the architecture behind most systems that combine image understanding with text generation. Their finding: visual embeddings get bent toward the text representation space to satisfy how attention mechanisms work, and that bend introduces a statistical bias toward language patterns. When the model looks at an image, its own training nudges it to describe what words expect rather than what pixels show. The team quantified this for the first time, showing the bias concentrates in the top principal components of a text subspace that holds across different datasets.
The practical stakes are real. Vision-language models are already deployed in medical imaging and autonomous systems — environments where a confident wrong answer is worse than no answer at all. Prior approaches either tried to close the gap between visual and language representations (which doesn't fix the skew) or patched hallucinations with expensive decoding tricks that add latency without touching the cause. This paper targets the geometric source directly, projecting the biased subspace out of visual representations at inference time — with no added compute over the base model — or via a bias-aware fine-tuning path.
Both methods reduced hallucinations across standard benchmarks including POPE, CHAIR, and AMBER, and improved scores on long-form captioning. The training-free variant is the more immediately deployable result: it slots into existing pipelines without modification.
The finding is a useful corrective to the assumption that bigger multimodal models simply get more accurate — if the geometric bias is baked into the architecture, scale alone won't fix it.