Research map / Vision-language and multimodal / Multimodal LLMs
Looking Back and Forth: Cross-Image Attention Calibration and Attentive Preference Learning for Multi-Image Hallucination Mitigation
Xiaochen Yang, Hao Fang, Jiawei Kong, Yaoxin Mao, Bin Chen, Shu-Tao Xia
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 |
|---|---|---|---|
| add CAPL (selective cross-image attention + attentive DPO) | Qwen2.5-VL-7B · MUIRBench | accuracy: 58.42 -> 62.00 | Improved |
| add CAPL | GLM4.1V-Base-9B · BLINK | accuracy: 58.17 -> 61.33 | Improved |
| add CAPL | InternVL2.5-8B · BLINK | accuracy: 54.81 -> 55.76 | Improved |
| add CAPL | Qwen2.5-VL-7B · MIBench | accuracy: 72.42 -> 71.06 | No clear effect |
| add CAPL | Qwen2.5-VL-7B · MIRB | accuracy: 52.73 -> 56.55 | Improved |
| add CAPL | Qwen2.5-VL-7B · POPE (single-image) | accuracy: 81.23 -> 82.94 | Improved |
| add CAPL | InternVL2.5-8B · CHAIRi (single-image) | CHAIRi: 6.4 -> 7.3 | Hurt |
| add CAPL | GLM4.1V-Base-9B · CHAIRs (single-image) | CHAIRs: 22.0 -> 18.4 | Improved |
| add selective cross-image attention only (inference-time) (ablation) | Qwen2.5-VL-7B · MUIRBench | accuracy: 58.42 -> 58.96 | No clear effect |
| add attentive DPO training on top of cross-image attention (ablation) | Qwen2.5-VL-7B · MUIRBench | accuracy: 58.96 -> 62.00 | Improved |
| use truncated-attention negatives instead of original-model negatives in DPO (ablation) | GLM4.1V-Base-9B · MUIRBench | accuracy: 59.5 -> 60.6 | Improved |
| raise selection ratio rho toward 1 (ablation) | CAPL · BLINK/MUIRBench | accuracy | Mixed |
Closest papers by meaning
- Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.92)
- Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (ICLR 2026 · similarity 0.92)
- Envision, Attend, Then Respond: Counterfactual Hallucination Mitigation in Large Vision-Language Models (CVPR 2026 · similarity 0.92)
- Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.92)
- Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 2026 · similarity 0.92)
- Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · 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.
