Research map / Vision-language and multimodal / Multimodal LLMs

VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning

ICLR 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.

Introduces VisuRiddles, an abstract visual reasoning benchmark plus a perception-annotated riddle synthesizer, and PAVR, a Qwen2.5-VL-7B model trained with SFT on synthetic perceptual descriptions/CoT followed by GRPO, reaching 46.8 average accuracy on the benchmark.

Problem

MLLMs perform poorly on abstract visual reasoning (AVR) puzzles, and existing datasets lack fine-grained perceptual annotations for training and diagnosis.

Approach

Attribute MLLM AVR failure primarily to fine-grained perception; synthesize riddles with structured perceptual descriptions and CoT, use them for SFT to improve perception, then apply GRPO with answer and format rewards to strengthen reasoning and perceptual strategy selection.

Key results

• PAVR (7B) reaches 46.8 average accuracy on VisuRiddles, versus 24.6 for the Qwen2.5-VL-7B baseline and 33.9 for Gemini2.5-pro.
• SFT alone (PAVR-SFT) reaches 39.5 average; adding GRPO raises it to 46.8.
• On VisuLogic, PAVR reaches 30.6 overall, above the 7B Qwen2.5-VL-7B-Instruct (26.0) and the listed VisuLogic-trained RL models of 7B size (28.0).

Surprising findings

• Scaling model size and CoT prompting do not consistently improve AVR performance; Qwen2.5-VL-72B with CoT scores about the same as without.
• Providing structured perceptual descriptions instead of raw images raises Qwen2.5-VL-72B from about 31% to 74% average accuracy, suggesting perception is the bottleneck.

Limitations

• Limited scale of rule design and icon library restricts diversity of synthetic data.
• Synthetic riddles are deliberately easier than real-world riddles.
• Benchmark is small (about 1,000 questions) and drawn from civil service exams with human-translated English.

Experimental setup

Datasets: VisuRiddles Synthesizer data (synthetic) VisuRiddles benchmark VisuLogic

Benchmarks: VisuRiddles VisuLogic

Metrics: accuracy

Baselines: Qwen2.5-VL-72B Qwen3-VL-235B-A22B-Thinking Gemini2.5-pro GPT-4o o3 GPT-5 Claude-3.7-Sonnet InternVL2.5-78B-MPO

Builds on: GRPO Qwen2.5-VL RAVEN VisuLogic

Closest papers by meaning

• VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs (CVPR 2026 · similarity 0.93)
• VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models (ICLR 2026 · similarity 0.93)
• Don't Show Pixels, Show Cues: Unlocking Visual Tool Reasoning in Language Models via Perception Programs (CVPR 2026 · similarity 0.93)
• Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.92)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.92)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.92)