Research map / Vision-language and multimodal / Multimodal LLMs
Same Pool, Different Answer: Stable Best-of-N Selection for Vision-Language Models
Pengfei Zheng
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 |
|---|---|---|---|
| replace length-norm selector with anchored selector (fixed reference, per-token average) | BoN on LLaVA-1.5 7B · SRP stability (500 items) | flip-rate (%): 44.3 -> 24.9 | Improved |
| replace length-norm selector with anchored selector | BoN on LLaVA-1.5 7B · POPE | hallucination rate (%): 9.6 -> 6.2 | Improved |
| add epsilon-margin tie-break to length-norm | BoN + length-norm · POPE | hallucination rate (%): 9.6 -> 9.8 | No clear effect |
| add epsilon-margin tie-break to length-norm | BoN + length-norm · SRP stability (500 items) | flip-rate (%): 44.3 -> 37.1 | Improved |
| use per-token averaging instead of sequence-level summation (ablation) | anchored selector · SRP stability (200 items, dropout) | flip-rate (%): 40.8 -> 24.1 | Improved |
| use fixed reference instead of context-free self-anchoring (PMI) (ablation) | anchored selector · SRP stability (200 items, dropout) | flip-rate (%): 35.4 -> 24.1 | Improved |
| use fixed reference instead of co-moving reference (ablation) | anchored selector · SRP stability (200 items, dropout) | flip-rate (%): 33.6 -> 24.1 | Improved |
| add the three training-time filters (length, prompt-copy, object-existence) (ablation) | DSPA · Object hallucination (judge-est.) | object halluc. rate (%): 23.6 -> 6.2 | Improved |
| add the three training-time filters (ablation) | DSPA · AMBER | accuracy: 77.5 -> 81.7 | Improved |
| use separately aligned RLAIF-V model instead of anchored selector on DSPA pool | length-norm selector · SRP stability (500 items) | flip-rate (%): 44.3 -> 35.1 | Improved |
| apply anchored selector at 13B scale | LLaVA-1.5 13B · SRP stability (avg) | flip-rate (%): 38.5 -> 22.3 | Improved |
| apply anchored selector to Qwen2.5-VL-7B-Instruct | Qwen2.5-VL-7B-Instruct · Object hallucination (judge-est., open-ended) | object halluc. rate (%): 38.4 -> 8.7 | Improved |
Closest papers by meaning
- Majority of the Bests: Improving Best-of-N via Bootstrapping (NeurIPS 2025 · similarity 0.88)
- Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models (ICLR 2024 · similarity 0.85)
- MAFA: Managing False Negatives for Vision-Language Pre-training (CVPR 2024 · similarity 0.85)
- Consistency and Uncertainty: Identifying Unreliable Responses From Black-Box Vision-Language Models for Selective Visual Question Answering (CVPR 2024 · similarity 0.84)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.84)
- Automated Generation of Challenging Multiple-Choice Questions for Vision Language Model Evaluation (CVPR 2025 · similarity 0.84)
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.
