HEAL: MLLMs hallucinate when information distribution drifts in synergy heads

SUAT-SZ · hf · 2026-09-21

This paper presents HEAL (Head-lEvel information disentAnglement and caLibration) for reducing multimodal LLM hallucinations. It first uses causal noise intervention on multi-head outputs to filter causally redundant heads, then disentangles information via counterfactual Difference-in-Differences into four head types. Key insight: hallucinations occur when information distribution drifts from a healthy equilibrium in synergy heads, largely independent of the number or strength of modality-specific heads. HEAL injects dynamic calibration factors into synergy heads' value vectors to steer outputs toward factual evidence, reducing hallucinations across multiple MLLMs with a simple, interpretable mechanism.

Original post →

More from Multimodal

Multimodal channel →