Research map / Vision-language and multimodal / Multimodal LLMs
Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Uses attention lens (visual attention ratio plus logit lens) to locate middle-layer visual information processing in LVLMs, shows hallucinated object tokens get weaker visual attention, and proposes a training-free inference-time attention intervention averaging across heads to reduce object hallucination.
Problem
Object hallucination in large vision-language models is poorly understood mechanistically, with prior work focusing on language bias rather than how visual information is processed inside the model.
Approach
Identifies two middle-layer stages (visual information enrichment in layers 5-18 and semantic refinement in 19-26) via attention and logit lens; hallucinated tokens show lower summed visual attention in the enrichment stage and inconsistent multi-head attention; a head-averaged attention amplification in those layers mitigates hallucination without training.
Key results
- The attention intervention reduces CHAIR_S on LLaVA-1.5-7B from 53.0 (greedy) to 25.0 and averages a 19.5% relative CS reduction over the second-best baseline across four models.
- SVAR5-18 detects hallucinated tokens on LLaVA-1.5-7B with AUROC 0.74 and AP 0.88, improving AUROC by 8.82% over internal confidence.
- Real object tokens get significantly higher summed visual attention than hallucinated ones in layers 5-18 (LLaVA-1.5-7B: 1.70 vs 1.25, p<0.01).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Heads Guided Attention Intervention during inference | LLaVA-1.5-7B · COCO 2014 val | CHAIRS (CS): 24.2 -> 25.0 | No clear effect |
| Add Heads Guided Attention Intervention during inference | LLaVA-1.5-7B · COCO 2014 val | CHAIRI (CI): 7.1 -> 6.7 | Helped |
| Add Heads Guided Attention Intervention during inference | LLaVA-1.5-13B · COCO 2014 val | CHAIRS (CS): 33.0 -> 25.8 | Helped |
| Add Heads Guided Attention Intervention during inference | Shikra-7B · COCO 2014 val | CHAIRS (CS): 38.6 -> 23.8 | Helped |
| Add Heads Guided Attention Intervention during inference | MiniGPT-4-7B · COCO 2014 val | CHAIRS (CS): 23.2 -> 21.4 | Helped |
| Apply attention intervention in layers 0-4 instead of layers 5-18 (ablation) | LLaVA-1.5-7B · COCO 2014 val | CHAIRS (CS): 25.0 -> 31.0 | Hurt |
| Apply attention intervention in layers 0-4 instead of layers 5-18 (ablation) | LLaVA-1.5-7B · COCO 2014 val | F1: 76.1 -> 77.1 | Helped |
| Apply attention intervention in layers 19-26 instead of layers 5-18 (ablation) | LLaVA-1.5-7B · COCO 2014 val | CHAIRS (CS): 25.0 -> 53.4 | Hurt |
| Apply attention intervention in layers 27-31 instead of layers 5-18 (ablation) | LLaVA-1.5-7B · COCO 2014 val | CHAIRS (CS): 25.0 -> 53.4 | Hurt |
| Adjust attention in the visual information enrichment range, layers 5-18, rather than using greedy decoding | LLaVA-1.5-7B · COCO 2014 val | CHAIRI (CI): 15.6 -> 6.7 | Helped |
Surprising findings
- Intervening in layers 19-26 or 27-31 increases hallucination rather than reducing it, so only the visual enrichment range is effective.
- Shikra-7B shows extremely high visual attention across nearly all layers, which weakens the attention-based detection metric.
Limitations
- Identifying the exact layer range for the two stages is subjective and manually set per model.
- Attention-based detection is weakened when a model like Shikra has uniformly high visual attention across layers.
Experimental setup
Datasets: COCO 2014 val
Benchmarks: CHAIR AMBER
Metrics: CHAIR_S CHAIR_I F1 AUROC AP
Baselines: Greedy decoding Beam search OPERA VCD PAI InterConf
Builds on: logit lens PAI OPERA VCD CHAIR
Closest papers by meaning
- Multi-Object Hallucination in Vision Language Models (NeurIPS 2024 · similarity 0.95)
- Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.94)
- Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 2026 · similarity 0.94)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.94)
- Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.94)
- Not All Tokens and Heads Are Equally Important: Dual-Level Attention Intervention for Hallucination Mitigation (AAAI 2026 · 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.