Research map / Vision-language and multimodal / Multimodal LLMs

SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Proposes SPARC, a two-stage prompting pipeline that first has a VLM localize question-relevant crops (implicit relevance detection) and then reasons over those crops, enabling asymmetric test-time scaling and cheap LoRA SFT of the perception stage.

Problem

Test-time scaling via unstructured "thinking with images" chains for VLMs entangles perception and reasoning, producing long, brittle contexts that require expensive RL with hand-crafted rewards.

Approach

Decouple visual inference into an Implicit Relevance Detection step that outputs bounding boxes or points for question-relevant regions, followed by a reasoning step conditioned on those high-resolution crops; apply weighted box fusion to self-consistent perception rollouts and fine-tune only the perception stage with a LoRA adapter.

Key results

• On V* with Qwen3-VL 8B, SPARC SFT 256 reaches 91.2% vs 87.0% native and surpasses Qwen3-VL 235B-A22B (91.1%).
• Averaged over V*, HRBench-4K and HRBench-8K, Qwen3-VL 4B SPARC WBF 8 at 256 resolution reaches 55.7 vs 41.7 native and 36.8 for thinking with images.
• On the OOD XLRS benchmark, SPARC with Molmo2 at 256 resolution surpasses full-resolution native performance while using about 0.1% of visual tokens.

Surprising findings

• Training the perception module at low resolution (256px) gave better results than full-resolution training, which was attributed to avoiding trivial distillation.
• Standard end-to-end distillation SFT on thinking-with-images traces underperforms the zero-shot thinking-with-images baseline, suggesting catastrophic forgetting.

Limitations

• Relies on VLMs with native spatial grounding (coordinate or point outputs); text-only VLMs would need an external detector.
• Accuracy on Qwen3-VL is reported as unstable for thinking with images; results rely on greedy decoding and multiple-choice benchmarks.
• Molmo2-4B fine-tuning showed no gain, attributed to a weak Molmo2-8B teacher.

Experimental setup

Datasets: DeepEyes dataset V* HRBench-4K HRBench-8K XLRS GQA MME Realworld

Benchmarks: V* HRBench-4K HRBench-8K XLRS MME Realworld GQA

Metrics: accuracy

Baselines: Native performance Thinking with images DeepEyes ViGoRL-7B Pixel-Reasoner Distillation SFT Bottom-up SPARC

Builds on: Chain-of-Thought Self-Consistency Weighted Boxes Fusion DeepEyes thinking with images

Closest papers by meaning

• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.92)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.92)
• Unleashing Perception-Time Scaling to Multimodal Reasoning Models (ICLR 2026 · similarity 0.91)
• Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning (ICLR 2026 · similarity 0.91)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.91)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.91)