Research map / Vision-language and multimodal / Multimodal LLMs
Improving Reasoning in Vision-Language Models via Perception Verified Self-Training
Sourabh Sharma, Sonam Gupta, Sadbhawna Sadbhawna
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 PerceptEval caption filtering to answer filtering (stage 1) (ablation) | Stage-1 self-training · M3CoT Lang. Sci. | accuracy (%): 49.76 -> 60.66 | Improved |
| add PerceptEval caption filtering to answer filtering (stage 1) (ablation) | Stage-1 self-training · M3CoT Commonsense | accuracy (%): 62.42 -> 69.45 | Improved |
| add two-stage curriculum with medium samples to dual-filter stage 1 (ablation) | Ours · M3CoT Lang. Sci. | accuracy (%): 60.66 -> 64.93 | Improved |
| add curriculum learning instead of mixing medium samples from the start (ablation) | Ours · M3CoT Lang. Sci. | accuracy (%): 57.34 -> 64.93 | Improved |
| add curriculum learning instead of mixing medium samples from the start (ablation) | Ours · M3CoT Commonsense | accuracy (%): 65.05 -> 72.74 | Improved |
| add FG-CLIP similarity to OCR-only PerceptEval (ablation) | PerceptEval · M3CoT Commonsense | accuracy (%): 60.44 -> 72.74 | Improved |
| add OCR similarity to FG-CLIP-only PerceptEval (ablation) | PerceptEval · M3CoT Lang. Sci. | accuracy (%): 56.87 -> 64.93 | Improved |
| add FG-CLIP similarity to OCR-only PerceptEval (ablation) | PerceptEval · M3CoT Lang. Sci. | accuracy (%): 61.13 -> 64.93 | Improved |
| add OCR similarity to FG-CLIP-only PerceptEval (ablation) | PerceptEval · M3CoT Commonsense | accuracy (%): 69.89 -> 72.74 | Improved |
| use perception-verified self-training instead of R3V | LLaVA-v1.5-7B · M3CoT Lang. Sci. | accuracy (%): 50.24 -> 64.93 | Improved |
| use perception-verified self-training instead of R3V on OCR hallucination split | LLaVA-v1.5-7B · HallusionBench OCR split | accuracy (%): 55.24 -> 60.14 | Improved |
| use perception-verified self-training instead of R3V on out-of-domain data | LLaVA-v1.5-7B · MMMU Science val | accuracy (%): 33.78 -> 36.48 | Improved |
Closest papers by meaning
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.93)
- Enhancing Large Vision Language Models with Self-Training on Image Comprehension (NeurIPS 2024 · similarity 0.92)
- Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · similarity 0.91)
- Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.91)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.91)
- Sherlock: Self-Correcting Reasoning in Vision-Language Models (NeurIPS 2025 · similarity 0.91)
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.
