Research map / Vision-language and multimodal / CLIP and zero-shot
Zero-Shot Rankability: Revealing Latent Ordinal Structure in Multimodal Large Language Models via Language
Nam Hyeon-Woo, Moon Ye-Bin, Sohwi Lim, Kwon Byung-Ki, Tae-Hyun Oh
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 attribute-specific conditional prompt to the image embedding (ablation) | Qwen2.5-VL-7B text-driven rank axis · Average over 8 datasets | SRCC: 0.544 -> 0.734 | Improved |
| add attribute-specific conditional prompt (ablation) | Qwen3-VL-8B text-driven rank axis · Average over 8 datasets | SRCC: 0.558 -> 0.710 | Improved |
| add attribute-specific conditional prompt (ablation) | Qwen3-VL-4B text-driven rank axis · Average over 8 datasets | SRCC: 0.587 -> 0.743 | Improved |
| add attribute-specific conditional prompt (ablation) | Qwen3-VL-2B text-driven rank axis · Average over 8 datasets | SRCC: 0.544 -> 0.724 | Improved |
| add conditional prompt (abstract aesthetics) (ablation) | Qwen2.5-VL-7B text-driven rank axis · KonIQ-10k | SRCC: 0.349 -> 0.763 | Improved |
| add conditional prompt (recency) (ablation) | Qwen2.5-VL-7B text-driven rank axis · HCI | SRCC: 0.234 -> 0.571 | Improved |
| add conditional prompt to the supervised linear probe embedding (ablation) | Qwen2.5-VL-7B linear probing · Average over 8 datasets | SRCC: 0.799 -> 0.839 | Improved |
| apply I0T modality gap-reduction to frozen CLIP-RN50 | CLIP-RN50 zero-shot rank axis · KonIQ-10k | SRCC: 0.466 -> 0.661 | Improved |
| apply I0T modality gap-reduction to frozen CLIP-ViTB/32 | CLIP-ViTB/32 zero-shot rank axis · KonIQ-10k | SRCC: 0.502 -> 0.663 | Improved |
| apply I0T modality gap-reduction to frozen CLIP-CNX | CLIP-CNX zero-shot rank axis · KonIQ-10k | SRCC: 0.508 -> 0.738 | Improved |
| replace cross-attention MLLM (IDEFICS1) with projector-based (IDEFICS2) | IDEFICS zero-shot rank axis · KonIQ-10k | SRCC: 0.338 -> 0.718 | Improved |
| replace cross-attention MLLM (IDEFICS1) with projector-based (IDEFICS2) | IDEFICS zero-shot rank axis · UTKFace | SRCC: 0.774 vs. 0.751 | No clear effect |
Closest papers by meaning
- Revisiting Multimodal Positional Encoding in Vision–Language Models (ICLR 2026 · similarity 0.89)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.88)
- HoPE: Hybrid of Position Embedding for Long Context Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- NaViL: Rethinking Scaling Properties of Native Multimodal Large Language Models under Data Constraints (NeurIPS 2025 · similarity 0.87)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.87)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.87)
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.
