Research map / Vision-language and multimodal / Multimodal LLMs
From Easy to Hard: The MIR Benchmark for Progressive Interleaved Multi-Image Reasoning
Hang Du, Jiayang Zhang, Guoshun Nan, Wendi Deng, Zhenyan Chen, Chenyang Zhang, Wang Xiao, Shan Huang, Yuqi Pan, Tao Qi, Sicong Leng
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 stage-wise curriculum with reasoning steps instead of plain fine-tuning | Mantis · MIR in-domain (average) | accuracy (%): 33.71% -> 35.99% | Improved |
| add stage-wise curriculum with reasoning steps instead of plain fine-tuning | LLaVA-Next-Interleave · MIR in-domain (average) | accuracy (%): 34.42% -> 36.58% | Improved |
| add stage-wise curriculum with reasoning steps instead of plain fine-tuning | LLaVA-OneVision · MIR in-domain (average) | accuracy (%): 42.36% -> 47.60% | Improved |
| add stage-wise curriculum with reasoning steps instead of plain fine-tuning | Qwen2-VL · MIR in-domain (average) | accuracy (%): 45.15% -> 51.76% | Improved |
| add stage-wise curriculum with reasoning steps instead of plain fine-tuning | mPLUG-Owl3 · MIR in-domain (average) | accuracy (%): 36.51% -> 40.36% | Improved |
| fine-tune on MIR with plain QA pairs | Qwen2-VL · MIR in-domain (average) | accuracy (%): 40.44% -> 45.15% | Improved |
| add curriculum method over plain tuning | Mantis · MIRBENCH | accuracy (%): 43.10% -> 40.82% | Hurt |
| add curriculum method over plain tuning | LLaVA-Next-Interleave · MUIRBench | accuracy (%): 41.32% -> 40.25% | Hurt |
| add curriculum method over plain tuning | Qwen2-VL · MUIRBench | accuracy (%): 45.51% -> 46.67% | Mixed |
| progress from stage 1 to stage 2 (less reasoning guidance) (ablation) | Qwen2-VL · MIR in-domain (average) | accuracy (%): 38.01% -> 36.73% | Hurt |
| progress through later stages 3, 4, 5 (ablation) | Qwen2-VL · MIR in-domain (average) | accuracy (%): 36.73% -> 41.03% -> 45.97% -> 51.76% | Improved |
| train through all five stages (ablation) | Qwen2-VL · MIR Sequential | accuracy (%): 32.22% -> 36.66% | No clear effect |
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.
- mPLUG-Owl3 — mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models (ICLR 2025) · 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
- Learning Interleaved Image-Text Comprehension in Vision-Language Large Models (ICLR 2025 · similarity 0.92)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.91)
- OCR-Reasoning Benchmark: Unveiling the True Capabilities of MLLMs in Complex Text-Rich Image Reasoning (ICLR 2026 · similarity 0.91)
- Will Multimodal Models Be Dazzled by Multi-Image Visual Puzzles? (CVPR 2026 · similarity 0.90)
- WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.90)
- Enhancing Multi-Image Understanding through Delimiter Token Scaling (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.
