AI That Describes Photos Now Stops Making Up Objects

If you've ever used an AI tool to describe a photo, you may have noticed it occasionally invents details. A dog that isn't there, a tree that never appeared. This is called object hallucination, and it's a big problem for vision-language models like LLaVA-1.5-7B, which are used to generate captions. This new research tackles the issue head-on by locating the exact microscopic parts of the model responsible.

Inside these AI systems are "attention heads" — tiny components that act like spotlights, deciding where the model looks when interpreting an image. The researchers found 32 of these spotlights that were particularly prone to misfiring, pointing at objects that didn't exist. Instead of retraining the whole model, they applied a slim adapter — think of it as a small adjustment valve — plus a grounding controller that constantly checks what the model is actually seeing. The result: captions with made-up objects dropped from 37% to 23%, a 38% reduction on a standard test using 400 photos.

The fix is also proven to be genuinely targeted. When they randomly chose different attention heads to adjust, there was no improvement at all. That shows the 32 heads they selected really were the culprits. The improvement even held up when generating longer captions, ruling out the possibility that the model was just cutting stories short to avoid mistakes.

There's a trade-off though. The corrected model is more careful, but also more conservative — it mentions fewer real objects, with recall dropping from 78% to 70%. That means it might leave out some true details. Still, for real-world uses like auto-captioning for visually impaired users, photo search, or visual assistants, fewer lies is a huge win. This is a step toward AI that sees more honestly.