Research map / Vision-language and multimodal / Multimodal LLMs
Delineating Knowledge Boundaries for Honest Large Vision-Language Models
Junru Song, Yimeng Hu, Yijing Chen, Huining Li, Qian Li, Lizhen Cui, Yuntao Du
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 |
|---|---|---|---|
| align with CPO on Visual-Idk preference pairs (few-shot) | LLaVA-1.5 7B · V-Idk test | Truthful rate: 57.9 -> 67.3 | Improved |
| SFT-only refusal tuning (ablation) | LLaVA-1.5 7B · V-Idk test (zero-shot) | IK-IK rate: 45.8 -> 34.4 | Hurt |
| SFT-only refusal tuning (ablation) | LLaVA-1.5 7B · V-Idk test (zero-shot) | Truthful rate: 47.6 -> 67.1 | Improved |
| use preference optimization (ORPO) instead of SFT (ablation) | LLaVA-1.5 13B · V-Idk test (few-shot) | Truthful rate: 67.3 (SFT) -> 76.6 (ORPO) | Improved |
| use DPO instead of SFT/CPO/ORPO (ablation) | LLaVA-1.5 7B · V-Idk test (zero-shot) | Truthful rate: 47.6 -> 57.4 | No clear effect |
| add few-shot demonstrations at inference (ablation) | LLaVA-1.5 7B base · V-Idk test | Truthful rate: 47.6 -> 57.9 | Improved |
| scale model from 7B to 13B (ablation) | LLaVA-1.5 ORPO · V-Idk test (zero-shot) | Truthful rate: 60.0 -> 66.2 | Improved |
| apply V-Idk-trained ORPO to ScienceQA | LLaVA-1.5 7B · ScienceQA (few-shot) | Truthful rate: 56.1 -> 63.8 | Improved |
| apply DPO to OOD scenarios (ablation) | LLaVA-1.5 7B · VizWiz-Unans / PMC-VQA | Truthful (refusal) rate: falls to 82.8% on VizWiz | Hurt |
| SFT on blurry VizWiz-Unans (ablation) | LLaVA-1.5 7B · VizWiz-Unans | Truthful (refusal) rate: 95.6 (+11.3 over baseline) | Improved |
| CPO on medical PMC-VQA (ablation) | LLaVA-1.5 7B · PMC-VQA | Truthful (refusal) rate: 93.9 | Improved |
| ORPO alignment changes refusal Logprob for unknown questions | LLaVA-1.5 7B · Visual-Idk probing | Logprob: -0.80 -> -0.63 | Improved |
Closest papers by meaning
- MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions (AAAI 2026 · similarity 0.89)
- Unveiling the Tapestry of Consistency in Large Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (ICLR 2025 · similarity 0.89)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.88)
- A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models (ICLR 2026 · similarity 0.88)
- Consistency and Uncertainty: Identifying Unreliable Responses From Black-Box Vision-Language Models for Selective Visual Question Answering (CVPR 2024 · similarity 0.88)
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.
