MLLM hallucinations track synergy-head drift; HEAL recalibrates the mix at decode time

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

Meng'en Qin, Junye Chen, Jucheng Liu, Youlu Xing, Song Wang, Ruize Han

cs.CV, cs.CL, cs.LG

2026-09-06

HEAL finds hallucinations track synergy-head visual share falling from 0.43 to 0.28. On LLaVA-1.5-7B, POPE F1 rises 85.4→87.7 and CHAIR_S drops 51.0→36.7.

What problem this solves

Multimodal LLMs invent objects that are not in the image. Inference-time fixes split into decoding tricks (VCD, OPERA) and attention-weight hacks that boost “visual heads.” Weights are a proxy. Heads also interact. Pushing one modality too hard recreates a familiar swing: extra vision truncates the reply, extra language grows the hallucination.

The sharper question is what actually moves inside a head when a hallucinated token is produced.

Method

HEAL has three stages.

The finding: hallucinated tokens are not short on visual or language heads in count or strength. The visual–language mix inside synergy heads drifts off equilibrium. Head roles also phase-shift: synergy heads lean language on language-centric tokens; some language heads become synergy heads on visually grounded tokens.

Results

On LLaVA-1.5-7B with Gaussian masking, synergy-head visual:language ratios are about 0.43:0.54 on correct tokens and 0.28:0.62 on hallucinated ones.

MethodPOPE F1CHAIRS ↓MME Total
Beam Search85.451.0565.34
EAH85.736.4603.99
MemVR87.146.6648.30
LocoRE86.938.4656.66
HEAL87.736.7669.76

CHAIRS is a hair behind EAH (36.7 vs 36.4) with higher recall and longer captions, and MME is best in the table. Plug-and-play on other nets: LLaVA-NeXT-7B CHAIRS 29.9 → 24.6; Qwen2.5-VL-7B 27.2 → 23.3; InternVL-7B 46.6 → 39.2. Qwen3-VL-8B MMHal hallucination rate 17.5 → 16.6. LLaVA-Bench on LLaVA-1.5 goes 72.5 → 75.2.

Causal checks run both ways. Lowering the visual share on a previously correct answer can induce a hallucination; raising it on a hallucinated answer can suppress one. Too large an α crushes language and yields broken grammar, a U-shaped curve. Zero masking, uniform noise, and cross-sample swaps still agree on head types 92.13%–95.36% of the time, and the drift direction holds.

Cost on LLaVA-1.5-7B: 5.02 tokens/s (base 17.25, EAH 3.17), wall-clock 70.74 s vs 11.36 s, extra 0.26 GB. LLaVA and InternVL use α=0.5 on easy suites and 0.6 on harder ones; Qwen families sit one notch lower.

Why it matters

The mechanistic rewrite is that hallucination is synergy-head mix drift, not a shortage of visual heads, and the patch needs no training. Anyone sitting on LLaVA, Qwen-VL, or InternVL who wants fewer object hallucinations can drop it in. α still has to be chosen per model and task: smaller for coarse existence, larger for spatial reasoning.

It cannot recover visual evidence that the encoder already dropped. That is a training and architecture problem; twisting attention heads will not grow new pixels.

Limitations

α and the update interval are empirical, with no closed-form rule, so a new model needs a sweep. One global α is applied to every synergy head, even though the paper notes that heads do not share a uniform preference. Throughput is about 3.4× slower than the base (17.25 → 5.02 tokens/s), which hurts long captions. CHAIRS does not beat EAH, so “balance beats uni-modal boosting” is a tie on the object-hallucination metric. The MMHal drop on Qwen3-VL is 17.5 to 16.6, a small absolute move. Thresholds change how many heads land in each bin; the drift direction is stable, but the synergy-head set is still a function of those thresholds.

Terms

Source

Related papers

All paper explainers