Mitigating Hallucinations in Vision-Language Models through Image-Guided Head Suppression
Open accessAbstract
Despite their remarkable progress in multimodal understanding tasks, large vision language models (LVLMs) often suffer from "hallucination", generating texts misaligned with the visual context.Existing methods aimed at reducing hallucinations through inference time intervention incur a significant increase in latency.To mitigate this, we present SPIN, a task-agnostic attention-guided head suppression strategy that can be seamlessly integrated during inference without incurring any significant compute or latency overhead.We investigate whether hallucination in LVLMs can be linked to specific model components.Our analysis suggests that hallucinations can be attributed to a dynamic subset of attention heads in each layer.Leveraging this insight, for each text query token, we selectively suppress attention heads that exhibit low attention to image tokens, keeping the top-k attention heads intact.Extensive evaluations on visual question answering and image description tasks demonstrate the efficacy of SPIN in reducing hallucination scores up to 2.7× while maintaining F1, and improving throughput by 1.8× compared to existing alternatives.Code is available here.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Recent work for VLMs has explored techniques to repair captioning errors prior to returning them during the generation process or post-hoc verification, but these techniques can induce further errors during correction, rather than providing a higher-precision output that only includes details that are likely correct.
“While recent work for VLMs has explored techniques to repair captioning errors prior to returning them during the generation process (e.g., controlling what objects are mentioned [144]; strategically adjusting model weights [78, 108, 136, 141] or fi ne-tuning [22, 146]; sampling multiple patches [29]; guided decoding [150]; backtracking when uncertain [40, 131]) or post-hoc verification [139, 155]”
From "It's trained by non-disabled people": Evaluating How Image Quality Affects Product Captioning with Vision-Language Models (2026) · cites this paper as [108]
Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 3
- 2024. Shiftaddllm: Accelerating pretrained llms via post-training multiplication-less reparameterization. Advances in Neural Information Processing Systems, 37.
- Xuan Gong, Tianshi Ming, Xinpeng Wang, and Zhi- hua Wei. 2024. DAMRO: Dive into the attention mechanism of LVLM to reduce object hallucination. In Proceedings of the 2024 Conference on Empiri- cal Methods in Natural Language Processing, pages 7696-7712, Miami, Florida, USA. Association for Computational Linguistics. · DOI
- Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023. Mitigating halluci- nation in large multi-modal models via robust instruc- tion tuning. In The Twelfth International Conference on Learning Representations.
Cited by
Sorted by citations
Garg, K., Tang, X., Heo, J., Morgan, D. R., Gergle, D., Sudderth, E. B., & Piper, A. M. (2026) · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3791309 · Cited by 3 · see citing works
Seeing to Ground: Visual Attention for Hallucination-Resilient MDLLMs
Narnaware, V., Gupta, A., Zhai, K., Wang, Z., & Shah, M. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37447-9_23 · Cited by 1 · see citing works
Zhao, P., Xu, J., Zhang, S., Wang, C., & Han, W. (2026) · Information Processing & Management · doi:10.1016/j.ipm.2026.104850 · Cited by 1 · see citing works
ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs
Yao, Z., Fu, Z., Zheng, Z., Li, J., Tu, Y., & Mao, Z. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37602-2_29
Wang, Z., Xu, B., Xia, Y., & Li, P. (2026) · doi:10.1109/cvprf71909.2026.00869
SADI: Spatial-Aware Dynamic Intervention for Mitigating Hallucinations in LVLMs
Xu, Y., Gao, Q., Jun, F., Ding, G., Yang, Z., Lin, S., & Xiao, Y. (2026) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3799682.3841021
- Authors
- Sarkar, S., Che, Y., Gavin, A., Beerel, P. A., & Kundu, S.
- Affiliations
- University of Southern California (US)
- Harvard University (US)
- Intel (United States) (US)
- Venue
- Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing
- Published
- 2025
- Pages
- 12492-12511
- OpenAlex
- W4416036555
- Cited by
- 4
- References
- 0
- Citations by year
- 2026: 4
- FWCI
- 3.00
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Conference on Empirical Methods in Natural Language Processing (EMNLP) (publishedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
- Keywords
- hallucination mitigation, large vision-language models, inference-time intervention, visual question answering, image description generation, attention analysis, latency overhead
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
