Research map / Vision-language and multimodal / CLIP and zero-shot
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
Andy Zeng, Maria Attarian, brian ichter, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke and 1 more
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 |
|---|---|---|---|
| Compose CLIP image information with GPT-3 caption generation and CLIP reranking | ZeroCap on MS COCO Captions · MS COCO Captions test subset (100 images) | CIDEr: 18.0 -> 44.5 | Improved |
| Add 3 training captions to the prompt (ablation) | Socratic Models on MS COCO Captions · MS COCO Captions test subset (100 images) | CIDEr: 44.5 -> 76.3 | Improved |
| Compose CLIP image information with GPT-3 caption generation | Kreiss et al. on Concadia · Concadia test | CIDEr (caption generation): 11.3 -> 38.9 | Improved |
| Compose CLIP image information with GPT-3 description generation | Kreiss et al. on Concadia · Concadia test | CIDEr (description generation): 17.4 -> 22.6 | Improved |
| Remove VLM image information (ablation) | Socratic Models on Concadia · Concadia test | CIDEr (caption generation): 38.9 -> 40.1 | Improved |
| Remove VLM image information (ablation) | Socratic Models on Concadia · Concadia test | CIDEr (description generation): 22.6 -> 20.6 | Hurt |
| Add the ground-truth image description as input (ablation) | Socratic Models on Concadia · Concadia test | CIDEr (caption generation): 93.8 | Improved |
| Add audio transcripts and language-based transcript reasoning to CLIP retrieval | CLIP via Portillo-Quintero et al. on MSR-VTT Full · MSR-VTT Full | R@1: 40.3 -> 44.7 | Improved |
| Use transcript-based reasoning for videos with long transcripts | CLIP via Portillo-Quintero et al. on MSR-VTT Full · MSR-VTT Full, long-transcript subset | R@1: 41.5 -> 54.9 | Improved |
| Add audio transcripts and language-based transcript reasoning to CLIP retrieval | CLIP via Portillo-Quintero et al. on MSR-VTT 1k-A · MSR-VTT 1k-A | R@1: 58.0 -> 60.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- MAGIC — MAGIC: Multimodal relAtional Graph adversarIal inferenCe for Diverse and Unpaired Text-Based Image Captioning (AAAI 2022) · better on 0 of 1 shared result
- ZeroCap — ZeroCap: Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic (CVPR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.89)
- Mastering Robot Manipulation with Multimodal Prompts through Pretraining and Multi-task Fine-tuning (ICML 2024 · similarity 0.88)
- Generative Multimodal Models are In-Context Learners (CVPR 2024 · similarity 0.88)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.88)
- ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.88)
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.
