Research map / Vision-language and multimodal / Multimodal LLMs
ONLY: One-Layer Intervention Sufficiently Mitigates Hallucinations in Large Vision-Language Models
Zifu Wan, Ce Zhang, Silong Yong, Martin Q. Ma, Simon Stepputtis, Louis-Philippe Morency, Deva Ramanan, Katia Sycara, Yaqi Xie
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply ONLY one-layer textual-enhanced decoding on top of regular decoding | LLaVA-1.5 · POPE MS-COCO Random | accuracy: 83.13 -> 89.70 | Improved |
| apply ONLY decoding | LLaVA-1.5 (max token 128) · CHAIR | CHAIR_S: 55.0 -> 49.8 | Improved |
| apply ONLY decoding | LLaVA-1.5 · MME-Hallucination | MME total score: 562.50 -> 635.55 | Improved |
| apply ONLY decoding | LLaVA-1.5 · MMVP | accuracy: 22.67% -> 28.00% | Improved |
| use ONLY instead of two-query contrastive decoding (VCD, M3ID) | LLaVA-1.5 · inference latency | relative latency: x2.01 (VCD) -> x1.07 | Improved |
| select attention heads by text-to-visual entropy ratio instead of ratio of attention sums (ablation) | ONLY textual enhancement · POPE and CHAIR | POPE accuracy: 84.20 -> 84.91; CHAIR_S 23.1 -> 20.0 | Improved |
| use entropy-ratio head selection instead of zeroing visual attention, adding noise, or doubling textual attention (ablation) | ONLY textual enhancement · POPE accuracy | accuracy: 84.26 -> 84.91 (vs zeroing visual attention) | Improved |
| add collaborative term (alpha_1 = 3 instead of 0) (ablation) | ONLY adaptive decoding · POPE and CHAIR_S | accuracy: 88.13 -> 89.70; CHAIR_S 23.5 -> 20.0 | Improved |
| increase alpha_1 beyond 3 (to 4) (ablation) | ONLY adaptive decoding · POPE | accuracy: 89.70 -> 88.37 | Hurt |
| add contrastive term (alpha_2 = 1 instead of 0) (ablation) | ONLY adaptive decoding · POPE and CHAIR_S | accuracy: 86.50 -> 89.70; CHAIR_S 24.8 -> 20.0 | Improved |
| apply intervention at the initial layer instead of other layers (ablation) | ONLY · POPE | accuracy | No clear effect |
| ONLY on InstructBLIP count subset MME | InstructBLIP · MME-Hallucination Count and Position | subset score: Count 77.78 vs VCD 91.67; Position 74.44 vs Regular 79.58 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- HALC — HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding (ICML 2024) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.95)
- Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models (ICLR 2025 · similarity 0.94)
- Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (ICLR 2025 · similarity 0.94)
- Grounding Language with Vision: A Conditional Mutual Information Calibrated Decoding Strategy for Reducing Hallucinations in LVLMs (NeurIPS 2025 · similarity 0.94)
- MoLE:Decoding by Mixture of Layer Experts Alleviates Hallucination in Large Vision-Language Models (AAAI 2025 · similarity 0.94)
- MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (ICLR 2025 · similarity 0.94)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
