Tsinghua ACMMM paper traces short-answer MLLM hallucinations to visual features, not language priors
新智元 · wechat · 2026-09-20
A Tsinghua team (Zhu Jun, Hu Xiaolin et al.) at ACMMM 2026 identifies "visual-origin hallucination": when MLLMs answer Yes/No, object hallucinations stem from visual feature extraction, not language priors.
Findings
- Hallucinating samples show collapsed image-text embedding alignment (cosine 0.158 vs −0.122) and inverted attention patterns (entropy 5.1% too high when object exists, 6.2% too low when absent)
- Causal evidence: noising/downsampling the vision encoder drops POPE accuracy from 0.842 to 0.739; a stronger SigLIP encoder raises it to 0.864
Method
- AHAF: PGD adversarial perturbation flips non-hallucinating images into hallucinating ones, creating perfectly aligned positive/negative pairs — also revealing MLLM visual representations sit dangerously close to the hallucination boundary
- ACFT: adversarial contrastive fine-tuning on these pairs
Results
- With only 0.9% of COCO and zero inference overhead, ACFT beats baselines on POPE/MME across LLaVA, MiniGPT-4, and Qwen2.5-VL; outperforming ordinary contrastive fine-tuning by up to 35.8% on equal data
More from Research
- ICLR paper: high train-test similarity doesn't explain CLIP's OOD generalization; 100M samples suffice — HildeKuehne · 2026-09-20
- Stanford professor questions whether peer review system was misaligned all along — anshulkundaje · 2026-09-20
- Reviewer says real names of reviewers and authors are often visible on OpenReview — AnkaReuel · 2026-09-20
- 30-second voice recording predicts biological aging, finds npj Aging study — segal_eran · 2026-09-20
- 5 Months of Daily Commits: Open ML Repo from NumPy to Transformers — oGauRav · 2026-09-20
- Andriy Burkov Tutorial: Conformal Prediction Adds Mathematically Guaranteed Error Bounds — burkov · 2026-09-20