Research map / Vision-language and multimodal / Multimodal LLMs
Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training
Cheng Tan, Jingxuan Wei, Zhangyang Gao, Linzhuang Sun, Siyuan Li, Ruifeng Guo, BiHui Yu, Stan Z. Li
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 rationale voting during training (ablation) | MC-CoT Base · ScienceQA | average accuracy: 84.70 -> 90.64 | Improved |
| add answer voting during training (ablation) | MC-CoT Base · ScienceQA | average accuracy: 90.19 -> 90.64 | Improved |
| vote during training instead of only at inference (ablation) | MC-CoT Base · ScienceQA | average accuracy: 84.58 -> 90.64 | Improved |
| combine mean and weighted-mean logits instead of either alone (ablation) | MC-CoT Base · ScienceQA | average accuracy: 90.03 -> 90.64 (mean only); 89.79 -> 90.64 (weighted only) | Improved |
| apply MC-CoT self-consistency training to Multimodal-CoT Base | Multimodal-CoT Base (223M) · ScienceQA | average accuracy: 84.91 -> 90.64 | Improved |
| apply MC-CoT self-consistency training to Multimodal-CoT Large | Multimodal-CoT Large (738M) · ScienceQA | average accuracy: 91.68 -> 93.37 | Improved |
| use FLAN-T5 Large instead of UnifiedQA Large as base model | MC-CoT · ScienceQA | average accuracy: 93.37 -> 94.88 | Improved |
| apply MC-CoT to A-OKVQA multi-choice with base model | Multimodal-CoT Base · A-OKVQA multi-choice | accuracy: 50.6 -> 71.0 | Improved |
| rationale quality gain from MC-CoT versus answer accuracy gain | MC-CoT Base · ScienceQA | RougeL of rationale and answer accuracy: about 1% RougeL, about 6% answer accuracy | Improved |
| answer voting removed: category effects (ablation) | MC-CoT Base · ScienceQA NAT and TXT | accuracy: NAT 91.87 -> 92.36; TXT 92.28 -> 92.42 without answer voting | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BLIP-2 — BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (ICML 2022) · better on 2 of 2 shared results
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
- MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency (ICML 2025 · similarity 0.92)
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
- KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning (AAAI 2024 · similarity 0.89)
- RPTS: Tree-Structured Reasoning Process Scoring for Faithful Multimodal Evaluation (AAAI 2026 · similarity 0.88)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.88)
- Tailoring Self-Rationalizers with Multi-Reward Distillation (ICLR 2024 · 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.
