AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge

Yueqi Song, Tianyue Ou, Yibo Kong, Zecheng Li, Graham Neubig, Xiang Yue

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge graph reasoning with LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
enable extended thinking mode (ablation)Claude-3.7-Sonnet · VisualPuzzlesaccuracy (%): 48.3 -> 48.2No clear effect
enable extended thinking mode (ablation)Claude-4.5-Sonnet · VisualPuzzlesaccuracy (%): 48.4 -> 49.4No clear effect
use thinking variant of Gemini-2.0-flash (ablation)Gemini-2.0-flash · VisualPuzzlesaccuracy (%): 45.0 -> 42.2Hurt
use reasoning model QvQ-72B instead of Qwen2.5-VL-72B (ablation)Qwen2.5-VL-72B · VisualPuzzlesaccuracy (%): 42.3 -> 37.8Hurt
use reasoning model o1 instead of GPT-4o (ablation)GPT-4o · VisualPuzzlesaccuracy (%): 41.3 -> 51.8Improved
scale up model size (ablation)open MLLMs · VisualPuzzlesaccuracy (%): LLaVA-OV 7B 29.4 vs 72B 30.8No clear effect
scale up model size (ablation)open MLLMs · MMMUknowledge and reasoning accuracyImproved
higher knowledge accuracy (correlation with reasoning accuracy)open MLLMs · MMMU vs VisualPuzzlesPearson correlation: 0.8 (MMMU) vs 0.4 (VisualPuzzles)Mixed
branching and re-validation reasoning patternsClaude-3.7-Sonnet-Thinking · VisualPuzzlescorrelation with accuracy: 0.01 and 0.12 vs 0.07 and 0.19 on MMMUNo clear effect
enforce Answer-First versus Option-First prompting (ablation)Kimi-K2.5 · VisualPuzzles vs MMMUaccuracyMixed
best model versus human 5th percentileall evaluated MLLMs · VisualPuzzlesaccuracy (%): 57.0 (o4-mini) vs 57.5 (human 5th pct)Hurt
perform VisualPuzzles instead of MMMU for knowledge needbenchmark design · knowledge concept checklistaverage knowledge questions: 3.9 -> 1.1Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy