Research map / Vision-language and multimodal / Multimodal LLMs
On Large Multimodal Models as Open-World Image Classifiers
Alessandro Conti, Massimiliano Mancini, Enrico Fini, Yiming Wang, Paolo Rota, Elisa Ricci
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 |
|---|---|---|---|
| ask the model to be more generic (ablation) | LMMs · fine-grained, very fine-grained and DTD | correct but generic share: +9.1% | Improved |
| ask the model to be more specific (ablation) | LMMs · fine-grained, very fine-grained and DTD | prediction correctness | Hurt |
| use domain-specific prompts (e.g. flower, car) (ablation) | LMMs · fine-grained, very fine-grained and DTD | correct and specific share: 12.5% | Improved |
| use domain-specific prompt on LLaVA-OV 0.5B (ablation) | LLaVA-OV 0.5B · fine-grained datasets | correct and specific share: +29% | Improved |
| use domain-specific prompt on Qwen2VL 7B (ablation) | Qwen2VL 7B · fine-grained datasets | correct and specific share: +15% | Improved |
| add test-time chain-of-thought reasoning prompts (ablation) | Qwen2VL · OW classification | correct and specific share: up to +13% | Improved |
| use multi-turn LlamaV-o1 prompt (ablation) | InternVL2 and Qwen2VL · OW classification | wrong and generic share: +1.7 | Hurt |
| replace base models with reasoning-tailored versions (InternVL2.5, Qwen2.5VL) (ablation) | InternVL2 and Qwen2VL · OW classification | correct and wrong-similar predictions: +6.6% correct, +3.7% wrong but specific | Mixed |
| prompt for multiple labels (list objects, caption, describe) (ablation) | LMMs · OW classification | text inclusion | Improved |
| compare LMMs against contrastive OW baselines | LMMs vs CaSED/CLIP retrieval · 10 datasets | metric/group wins: 11/16 | Improved |
| compare Qwen2VL 2B on FGVCAircraft | Qwen2VL 2B · FGVCAircraft | text inclusion: 25.6 vs 4.6 | Improved |
| allow multi-label tags for wrong predictions | LMMs · OW classification | wrong predictions matching an image tag: 30% to 60% | Improved |
Closest papers by meaning
- A Concept-Based Explainability Framework for Large Multimodal Models (NeurIPS 2024 · similarity 0.90)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.89)
- Matryoshka Multimodal Models (ICLR 2025 · similarity 0.88)
- Multimodal Ensembling for Zero-Shot Image Classification (AAAI 2024 · similarity 0.88)
- MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models (ICLR 2025 · similarity 0.87)
- Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · 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.
