Research map / Vision-language and multimodal / Multimodal LLMs
SHIFT: Smoothing Hallucinations by Information Flow Tuning for Multimodal Large Language Models
Sudong Wang, Yunjian Zhang, Yao Zhu, Enci Liu, Jianing Li, Yanwei Liu, Xiangyang Ji
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 SHIFT mutation-layer smoothing under greedy decoding | LLaVA-1.5 · CHAIR | CHAIR_S: 49.6 -> 43.8 | Improved |
| apply SHIFT under greedy decoding | mPLUG-Owl2 · CHAIR | CHAIR_S: 59.8 -> 50.2 | Improved |
| apply SHIFT under greedy decoding | InstructBlip · CHAIR | CHAIR_I: 23.2 -> 19.9 | Improved |
| use SHIFT instead of OPERA with beam search | LLaVA-1.5 · CHAIR | CHAIR_S: 45.4 -> 36.7 | Improved |
| use VCD contrastive decoding with nucleus sampling | LLaVA-1.5 · CHAIR | CHAIR_S: 48.4 -> 56.4 | Hurt |
| use SHIFT instead of ICD with nucleus sampling | mPLUG-Owl2 · CHAIR | CHAIR_S: 62.4 -> 47.2 | Improved |
| apply SHIFT under greedy decoding | LLaVA-1.5 · POPE MSCOCO adversarial | accuracy: 68.40 -> 74.40 | Improved |
| apply SHIFT under greedy decoding | InstructBlip · POPE MSCOCO random | accuracy: 90.23 -> 90.40 | No clear effect |
| use OPERA with beam search | LLaVA-1.5 · POPE MSCOCO random | accuracy: 89.63 -> 89.30 | No clear effect |
| apply SHIFT to nucleus sampling | LLaVA-1.5 · POPE MSCOCO adversarial | accuracy: 67.25 -> 74.13 | Improved |
| apply SHIFT versus vanilla decoding | LLaVA-1.5 · CHAIR average output length | tokens: 101.1 -> 99.8 | No clear effect |
| use SHIFT instead of OPERA | LLaVA-1.5 · CHAIR inference time | time (ms): 16657 -> 2519 | Improved |
Closest papers by meaning
- Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection (CVPR 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Mitigating Hallucinations in Large Vision-Language Models by Adaptively Constraining Information Flow (AAAI 2025 · similarity 0.92)
- MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (ICLR 2025 · similarity 0.92)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.92)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 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.
