Research map / Vision-language and multimodal / Multimodal LLMs
OCTOPUS: Multi‑Agentic Universal Compositional Visual Retrieval
Zhangtao Cheng, Bozhu Zheng, Ting Zhong, Fan Zhou
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 |
|---|---|---|---|
| remove the creation agent (visual and textual proxies) (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 69.32 -> 76.21 (w/o Creator vs full) | Improved |
| add retrieval verification (LLM re-ranking) (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 65.64 -> 76.21 | Improved |
| add self-refining tool in the Perceiver (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 70.57 -> 76.21 | Improved |
| add LLM reasoning (modification) tool in the Perceiver (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 70.68 -> 76.21 | Improved |
| add textual proxies (semantic increment) (ablation) | OCTOPUS · CIRCO (ViT-L/14) | avg mAP (k=5,10): 24.43 -> 28.42 | Improved |
| add visual proxy generation (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 71.45 -> 76.21 | Improved |
| use bidirectional retrieval instead of text-to-vision and vision-to-vision (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 71.49 -> 76.21 | Improved |
| swap GPT-4o-mini for open-source LLM MiniGPT-4 (ablation) | OCTOPUS · WebVid (ViT-L/14) | avg Recall (R@1,R@5): 76.21 -> 70.71 | Hurt |
| swap GPT-4o-mini for GPT-4o (ablation) | OCTOPUS · WebVid and CIRCO (ViT-L/14) | avg Recall / avg mAP: 76.21 -> 76.60; 28.42 -> 28.77 | No clear effect |
| increase visual-proxy weight lambda beyond optimum (ablation) | OCTOPUS gating · WebVid, CIRCO, FashionIQ Toptee | Recall / mAP | Hurt |
| add creator and retriever agents to LinCIR and OSrCIR (ablation) | LinCIR, OSrCIR · WebVid | Recall@k | Improved |
| use OCTOPUS vs supervised/zero-shot baselines on FashionIQ | OCTOPUS · FashionIQ ViT-L/14 | avg Recall@10: 33.26 -> 37.00 | Improved |
Closest papers by meaning
- ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge (NeurIPS 2026 · similarity 0.87)
- Navigating Massive Visual Context in Retrieval-Augmented Generation via Multimodal Memory Graph (ICML 2026 · similarity 0.87)
- Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models (NeurIPS 2026 · similarity 0.86)
- VELOCITI: Benchmarking Video-Language Compositional Reasoning with Strict Entailment (CVPR 2025 · similarity 0.86)
- MATA: A Trainable Hierarchical Automaton System for Multi-Agent Visual Reasoning (ICLR 2026 · similarity 0.86)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.86)
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.
