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.
More from Multimodal
- Fan-Made Video Pits Dario Against Altman, Made With Seedance 2.5 — teortaxesTex · 2026-09-21
- Qwen-Image-2.1 trending on Hugging Face — Qwen · 2026-09-21
- Google Gemini can now edit your videos: 10 prompts that do it all — nikola_mr64990 · 2026-09-21
- TypeSafe's JEV model fully open with $5 free credit, powers 1-second 3D scene generation — tinyfool · 2026-09-21
- Redditor recreates TaleSpin as chillwave with Minimax video and YuE2 music — the_bollo · 2026-09-21
- Can Minimax H3 turn anime into live action? A parent's reference-to-video attempt fails — yes_u_suckk · 2026-09-21