Research map / Vision-language and multimodal / Multimodal LLMs

VaseVQA-3D: Benchmarking 3D VLMs on Ancient Greek Pottery

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VaseVQA-3D, a 664-model 3D vase VQA dataset built from filtered 2D vase images via TripoSG, and VaseVLM, a Qwen2.5-VL model fine-tuned with LoRA SFT then GRPO using a six-dimension verifiable reward.

Problem

Vision-language models lack training data and domain knowledge for 3D ancient Greek vase artifacts, a long-tail cultural heritage domain, so they perform poorly on vase analysis questions.

Approach

Convert filtered 2D vase images to 3D GLB models with TripoSG, rewrite VaseVQA metadata into captions with GPT-4o, fine-tune Qwen2.5-VL on 360-degree rotation videos with LoRA, then apply GRPO with a reward that scores six archaeological dimensions (fabric, technique, shape, dating, decoration, attribution) by embedding similarity plus penalties.

Key results

• VaseVLM-7B-RL reaches R@1 3.52% vs 3.12% for the best baselines (Gemini-2.5-flash, GPT-4.1 and Claude-4-sonnet at 2.20-3.12%) and lexical similarity 0.276 vs 0.274 for DiffuRank.
• Human expert average of 4.57 for VaseVLM-7B-RL vs 4.07 for DiffuRank (the best baseline), with RL versions outperforming SFT versions by about 0.4 points.
• TripoSG chosen over Hunyuan3D for 3D generation: CLIP-T 0.9594 vs 0.9237 and CD 0.1490 vs 0.1515 on 24 VaseEval models.

Surprising findings

• Gemini-2.5-flash has the best FID (0.325) and R@10 (28.57%) among all methods but a low lexical similarity (0.210), and GPT-4o scores only 0.104 lexical similarity.
• The overall pipeline retains only 2.2% of the original 30,000 images (664 models), with TripoSG generation succeeding on 17.1% of filtered images.

Limitations

• Only 17.1% of filtered images yield usable 3D models, limiting dataset scale.
• Methodology is specifically designed for ancient Greek pottery and may not transfer to other artifacts.
• Reported R@1 values are very low in absolute terms (about 2-3.5%), and the gains over baselines are small in several metrics.

Experimental setup

Datasets: VaseVQA Sketchfab VaseEval reference models

Benchmarks: VaseVQA-3D

Metrics: FID CLIP score R@1 R@5 R@10 lexical similarity PSNR SSIM LPIPS Chamfer distance

Baselines: DiffuRank Cap3D LLaVA3D Gemini-2.5-flash Claude-4-sonnet GPT-4o Qwen2.5-VL-7B InternVL

Builds on: GRPO Qwen2.5-VL TripoSG VaseVQA LoRA

Closest papers by meaning

• 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.84)
• OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.84)
• Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation (CVPR 2025 · similarity 0.84)
• Zero-shot 3D Question Answering via Voxel-based Dynamic Token Compression (CVPR 2025 · similarity 0.84)
• Bridging the Gap between 2D and 3D Visual Question Answering: A Fusion Approach for 3D VQA (AAAI 2024 · similarity 0.83)
• VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.83)