UniProbe reduces VLM hallucinations by 55% with minimal latency overhead

HaggaiMaron · x · 2026-08-17

UniProbe is a learnable, token-level hallucination detection method for large Vision-Language Models (VLMs). It inspects the VLM's internals to identify hallucinated response tokens and mitigates them before they reach the user. The method achieves a 55% reduction in hallucinations with just a 1.06x increase in latency.

Original post →

More from Research

Research channel →