Research map / Vision-language and multimodal / Multimodal LLMs
See Only When Needed: Context-Aware Attention Intervention for Hallucination-Free LVLMs
Yuqing Lei, Wenbo Lyu, Yingjun Du, Xiantong Zhen, Cees Snoek, Ling Shao
Read the paperCheck your idea against 137,217 papers
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 CAI (token- and entropy-gated attention tilt) to LLaVA-1.5 | LLaVA-1.5 · CHAIR max 128 | CHAIR_S: 56.6 -> 39.6 | Improved |
| Attention-only CAI (lambda=1) without contrastive decoding (ablation) | LLaVA-1.5 · POPE/CHAIR/MME | POPE avg: 74.81 -> 77.13 | Improved |
| Uniform attention amplification (PAI†) instead of selective (ablation) | LLaVA-1.5 · POPE/CHAIR/MME | POPE avg: 74.81 -> 74.57 | No clear effect |
| Add contrastive decoding on top of CAI† (ablation) | CAI · POPE avg | accuracy: 77.13 -> 83.67 | Improved |
| Apply CAI to InstructBLIP | InstructBLIP · CHAIR max 128 | CHAIR_S: 54.2 -> 52.8 | No clear effect |
| Apply CAI on MME | LLaVA-1.5 · MME | score: 588.33 -> 660.00 | Improved |
| Apply CAI to Qwen3.5-4B softmax layers rather than GLA gates (ablation) | Qwen3.5-4B · CHAIR max 128 | CHAIR_S: 35.0 -> 29.8 | Improved |
| Tune start layer ls and entropy threshold gamma (ablation) | CAI · POPE A-OKVQA random | accuracy | Mixed |
| Tune contrastive coefficient lambda (ablation) | CAI · POPE A-OKVQA random | accuracy | Mixed |
Closest papers by meaning
- 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 0.93)
- Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.93)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.93)
- Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (ICLR 2025 · similarity 0.93)
- Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention (CVPR 2025 · similarity 0.92)
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.
