Research map / Vision-language and multimodal / Multimodal LLMs
Corvid: Improving Multimodal Large Language Models Towards Chain-of-Thought Reasoning
Jingjing Jiang, Chao Ma, Xurui Song, Hanwang Zhang, Jun Luo
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 gate attention in GateMixer (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 54.1 -> 55.6 | Improved |
| add L_CReg contrastive alignment loss in pretraining (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 55.0 -> 55.6 | Improved |
| add learnable prefix token embeddings (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 54.8 -> 55.6 | Improved |
| use GateMixer instead of MLP connector (FC_GELU_FC, 2xFC_GELU_FC) (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 54.0 -> 55.6; 53.7 -> 55.6 | Improved |
| train on high-quality CoT data instead of direct-answer data (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 45.7 -> 55.6 | Improved |
| use refined CoTs instead of raw rationales (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 51.2 -> 55.6 | Improved |
| use Corvid-1M instead of LLaVA-665K, Cambrian-1M or Ovis-1M (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 44.1, 51.9, 53.8 -> 55.6 | Improved |
| increase SFT data diversity (CoT-only to Corvid-1M mix with direct and OCR data) (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 48.0 -> 55.6 | Improved |
| add inference-time self-verification over always-CoT inference (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 49.7 -> 55.6 | Improved |
| add inference-time self-verification over always-CoT inference (ablation) | Corvid-o1 · 9 reasoning benchmarks | average accuracy: 56.2 -> 59.9 | Improved |
| use hybrid SigLIP+ConvNeXt vision encoder instead of one encoder (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 54.7 -> 55.6; 52.3 -> 55.6 | Improved |
| raise self-verification weight alpha from 0.0 to 0.7 (ablation) | Corvid-base · 9 reasoning benchmarks | average accuracy: 48.6 -> 55.6 | Improved |
Closest papers by meaning
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
- Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision (ICLR 2026 · similarity 0.90)
- Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought (CVPR 2026 · similarity 0.90)
- Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
- MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency (ICML 2025 · similarity 0.90)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
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.
