Research map / Vision-language and multimodal / Multimodal LLMs

ARC Is a Vision Problem!

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Frames ARC puzzles as image-to-image translation on a canvas, training a from-scratch ViT (18M) on ARC data with test-time training and multi-view voting, reaching 54.5% (single) and 60.4% (ensemble) on ARC-1.

Problem

Abstract reasoning on the ARC benchmark is usually treated as a language problem; the paper asks whether ARC's visual, few-shot, many-task nature can be addressed with vision models trained from scratch.

Approach

Represent each ARC grid on a fixed-size canvas with scale and translation augmentation, process it with a standard vision backbone (ViT or U-Net) as per-pixel classification conditioned on a task token, and adapt per test task with test-time training and multi-view majority voting.

Key results

• VARC (ViT-18M, single model) reaches 54.5±0.7% pass@2 on ARC-1 and 8.3% on ARC-2, about 10 points above TRM (44.6%) with a similar small scale.
• Ensemble of an 18M ViT and a 55M U-Net with test-time training reaches 60.4% on ARC-1 and 11.1% on ARC-2, matching the reported average human score of 60.2 on ARC-1.
• Visual priors (2D positional embedding, 2x2 patchification on a 64x64 canvas, translation and scale augmentation) cumulatively add 27.7 points in the ablation.

Surprising findings

• Test-time training run independently per task outperforms joint TTT across all test tasks by about 10 points.
• Without any offline training, test-time training alone reaches 26.4% on ARC-1, suggesting some tasks are solvable tabula rasa.

Limitations

• Scaling beyond the 18M-scale regime overfits; the 66M ViT has higher training accuracy but lower test accuracy.
• ARC-2 results are modest (8.3% single, 11.1% ensemble), and models are trained only on ARC-1 data for ARC-2.
• Relies on per-task test-time training, adding inference-time compute per task.

Experimental setup

Datasets: ARC-1 training set RE-ARC ARC-1 evaluation set ARC-2

Benchmarks: ARC-1 ARC-2

Metrics: pass@2 accuracy pass@1 accuracy pass@k

Baselines: HRM TRM ViT-ARC Deepseek R1 Claude 3.7 8k o3-mini-high GPT-5 Grok-4-thinking Bespoke (Grok-4)

Builds on: Vision Transformer U-Net test-time training RE-ARC RoPE

Closest papers by meaning

• Think Visually, Reason Textually: Vision-Language Synergy in Abstract Reasoning (CVPR 2026 · similarity 0.91)
• Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale (ICML 2026 · similarity 0.89)
• What Makes a Maze Look Like a Maze? (ICLR 2025 · similarity 0.88)
• Slot Abstractors: Toward Scalable Abstract Visual Reasoning (ICML 2024 · similarity 0.88)
• What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning (AAAI 2026 · similarity 0.88)
• Bongard in Wonderland: Visual Puzzles that Still Make AI Go Mad? (ICML 2025 · similarity 0.88)