Research map / Vision-language and multimodal / Cross-modal learning
Same Content, Different Answers: Cross-Modal Inconsistency in MLLMs
Angela van Sprang, Laurens Samson, Ana Lucic, Erman Acar, Sennay Ghebreab, Yuki M. Asano
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 |
|---|---|---|---|
| present question as text instead of image | GPT-4o-mini · MMLU | accuracy: 77.0 -> 84.4 | Improved |
| present question as text instead of image | Phi-4 · MMLU | accuracy: 31.4 -> 47.0 | Improved |
| present question as text instead of image | Phi-4 · SoEBench | accuracy: 17.3 -> 13.3 | Hurt |
| raise rendering resolution from DPI 50 to 200 (ablation) | Haiku-4.5 · REST+ (all questions) | RER: 53.6 -> 76.3 | Improved |
| raise rendering resolution from DPI 50 to 200 (ablation) | GPT-5-mini · REST+ (all questions) | RER: 71.5 -> 83.6 | Improved |
| raise rendering resolution from DPI 50 to 200 (ablation) | Gemini-2.5 Flash Lite · REST+ (all questions) | RER: 70.6 -> 70.9 | No clear effect |
| change font family (DejaVu Sans, Courier New, Cursive) (ablation) | MLLMs · REST+ | image accuracy: within 2% absolute difference | No clear effect |
| use coloured (red or yellow) text instead of black (ablation) | MLLMs · REST+ (DPI@200) | image accuracy: more than 5% relative improvement | Improved |
| OCR first then solve instead of direct image processing (ablation) | Deepseek-Small · MMLU | accuracy: -13.1 | Hurt |
| OCR first then solve instead of direct image processing (ablation) | Qwen-2.5-32B · SoEBench | accuracy: +10.1 | Improved |
| OCR first then solve instead of direct image processing (ablation) | Phi-4 · AI2-ARC | accuracy: +9.0 | Improved |
| OCR first then solve instead of direct image processing (ablation) | MLLMs · REST | accuracy | Mixed |
Closest papers by meaning
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.91)
- XModBench: Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models (ICLR 2026 · similarity 0.91)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.90)
- Evaluating Cross-Modal Reasoning Ability and Problem Charactaristics with Multimodal Item Response Theory (ICLR 2026 · similarity 0.90)
- Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs (NeurIPS 2025 · similarity 0.90)
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 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.
