💬 NLP

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

This paper introduces SPARC, a modular framework that decouples visual perception from reasoning in vision-language models to enable efficient, asymmetric test-time scaling and significantly improve performance on visual reasoning tasks with reduced token budgets.

Original authors: Niccolo Avogaro, Nayanika Debnath, Li Mi, Thomas Frick, Junling Wang, Zexue He, Hang Hua, Konrad Schindler, Mattia Rigotti

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Niccolo Avogaro, Nayanika Debnath, Li Mi, Thomas Frick, Junling Wang, Zexue He, Hang Hua, Konrad Schindler, Mattia Rigotti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Overthinking" Detective

Imagine you are a detective trying to solve a mystery based on a photograph. The photo is huge, high-resolution, and cluttered with thousands of tiny details.

Current AI models (called Vision-Language Models or VLMs) try to solve this by doing something called "Thinking with Images." It’s like the detective staring at the entire photo, muttering to themselves, zooming in on a shoe, then zooming out, then zooming in on a clock, all while writing a long, rambling diary entry about every single thing they see.

This approach has two big problems:

It’s messy: The detective gets distracted. If they misidentify a blue scarf as green in their diary, that mistake ruins the rest of their reasoning.
It’s expensive: Writing that long diary entry takes a lot of time and computer power.

The Solution: SPARC (Separating Perception And Reasoning Circuits)

The authors of this paper propose a new method called SPARC. Instead of letting the AI ramble, they split the job into two distinct steps, inspired by how the human brain works.

Think of it like a Restaurant Kitchen:

Step 1: The Sous-Chef (Perception)

In a restaurant, the head chef doesn’t chop every vegetable. They have a sous-chef whose only job is to prepare the ingredients.

What the AI does: The AI looks at the question (e.g., "What color is the scarf?") and the full image. Its only job is to find the specific part of the image that matters. It doesn’t try to answer the question yet. It just draws a box around the scarf.
The Analogy: The sous-chef finds the scarf in the pile of clothes and sets it aside on a clean cutting board. They don’t care about the rest of the room.

Step 2: The Head Chef (Reasoning)

Now, the head chef steps in.

What the AI does: The AI is shown only the cropped image of the scarf (and maybe the original low-res image for context). It now uses its brainpower to reason and answer the question: "It is green."
The Analogy: The head chef looks at the isolated scarf on the cutting board. Because the distraction of the rest of the room is gone, the chef can easily identify the color without getting confused by other items.

Why Is This Better?

1. It Saves Money and Time (Efficiency)
In the old method, the AI had to process the entire huge photo while writing its long diary. With SPARC, the AI can look at a low-resolution, blurry version of the whole photo just to find where the scarf is. Once it finds the location, it only zooms in on that tiny spot.

The Result: The paper shows that SPARC can achieve better accuracy while using 200 times fewer tokens (the units of computation) than the old method. It’s like solving the mystery by looking at a postcard to find the suspect, then using a magnifying glass on just their face, rather than examining every square inch of the crime scene with equal intensity.

2. It Prevents "Hallucinations" (Accuracy)
When the AI tries to do everything at once, it often gets confused by biases or irrelevant details (like thinking a stone lamp post is a trash can because it’s shiny). By separating the steps, SPARC ensures the AI looks at the right thing before it tries to think about it.

The Result: In tests, SPARC was more accurate than the standard "thinking with images" method, especially when the images were difficult or out of the ordinary (like satellite images).

3. You Can Tune the Parts Separately (Modularity)
If the AI is bad at finding objects, you can train just the "Sous-Chef" part to be better at spotting things, without messing up the "Head Chef’s" ability to reason.

The Surprise: The authors found that training the "Sous-Chef" on blurry, low-resolution images actually made it better. Why? Because it forced the AI to learn the shape and context of objects rather than just memorizing pixel details. It’s like learning to recognize a friend by their silhouette rather than their exact facial features—it makes the recognition more robust.

Summary

SPARC stops AI models from rambling. It forces them to first look (find the relevant part of the image) and then think (answer the question based on that part). This makes the AI faster, cheaper, and more accurate, much like a professional kitchen where prep work is separated from cooking.

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

Problem Statement
Current Vision-Language Models (VLMs) struggle with robust test-time scaling, particularly when dealing with high-resolution visual reasoning tasks. The prevailing paradigm, often termed "thinking with images," involves interleaving textual Chain-of-Thought (CoT) reasoning with visual perceptual actions (e.g., zooming, cropping) within a single, monolithic inference loop. This approach entangles perception and reasoning tokens, leading to several critical issues:

Brittleness and Error Cascades: Unstructured reasoning chains can cause small perceptual errors to cascade into completely incorrect answers.
Computational Inefficiency: Managing long, multi-turn contexts with mixed visual and textual tokens is computationally expensive and prone to degradation in smaller models.
Training Complexity: Optimizing these models typically requires expensive Reinforcement Learning (RL) with hand-crafted rewards.
Lack of Flexibility: Monolithic architectures lack mechanisms to dynamically allocate compute based on task difficulty, such as prioritizing perceptual processing under distribution shift.

Methodology
The authors propose SPARC (Separating Perception And Reasoning Circuits), a modular framework inspired by the hierarchical information processing architecture of the biological brain. SPARC decouples the VLM inference process into two distinct, sequential stages:

Stage 1: Perception (Implicit Relevance Detection - IRD):
- The model acts as a "perceptual circuit," performing explicit visual search to localize question-relevant regions.
- Instead of generating a full reasoning chain, the model outputs spatial coordinates (bounding boxes or points) for relevant image crops.
- This stage is designed to be lightweight and can operate at lower image resolutions to reduce token count.
Stage 2: Reasoning (Prefrontal Cortex Circuit):
- The model acts as a "reasoning circuit," synthesizing a CoT to produce the final answer.
- Crucially, this reasoning is conditioned only on the high-resolution crops identified in Stage 1, rather than the entire global image context.
- This separation allows for independent optimization and asymmetric compute allocation.

Key Contributions

Decoupled Architecture: SPARC explicitly separates visual perception from reasoning, preventing perceptual errors from contaminating the reasoning context and reducing the risk of "mirage reasoning" or hallucinations.
Asymmetric Test-Time Scaling: The framework allows for independent scaling of perception and reasoning. For instance, the authors demonstrate Weighted Box Fusion (WBF) applied to the perception stage. By running multiple low-cost IRD rollouts (e.g., $N=8$ ) and fusing the resulting bounding boxes, the model can improve localization accuracy without incurring the linear compute cost of running multiple full reasoning chains.
Efficient Fine-Tuning: SPARC enables targeted fine-tuning of the perceptual circuit using Low-Rank Adaptation (LoRA) without degrading the base model’s general reasoning capabilities. The authors find that training the perception module on lower-resolution images (256px) acts as a regularizer, improving robustness compared to full-resolution distillation.
Context Compression: By running global search at lower resolutions and allocating high-resolution processing only to selected regions, SPARC significantly reduces visual token counts and compute requirements.

Results
The authors evaluate SPARC on challenging visual reasoning benchmarks, including $V^*$ , HRBench-4K/8K, and the out-of-distribution (OOD) XLRS remote sensing dataset. Key findings include:

Performance Gains: SPARC outperforms monolithic baselines and native "thinking with images" approaches. For example, on the $V^*$ benchmark, SPARC improves Qwen3-VL 4B performance by 6.7 points. In OOD settings, it surpasses "thinking with images" by 4.6 points while using a 200 $\times$ lower token budget.
Efficiency: SPARC achieves superior accuracy with significantly fewer tokens. On the XLRS dataset, Molmo2 operating at 256-pixel resolution with SPARC crops surpasses the performance of the standard model prompted at full resolution, processing only ~0.1% of the visual tokens required for a naive full-resolution forward pass.
Scalability of Perception: Applying self-consistency via WBF to the perception stage (e.g., 8 rollouts) boosts performance by up to 9.3% with minimal additional text tokens, as the expensive reasoning backbone processes only one refined input.
Robustness: SPARC does not suffer from "tunnel vision" on global scene understanding tasks, as the reasoning stage still has access to the global image context alongside the localized crops.

Significance
The paper argues that SPARC represents a shift toward visual context engineering, where a well-structured, modular context containing only necessary high-resolution content offers a more compact and robust representation than long, unorganized multi-modal rollouts. By decoupling perception and reasoning, SPARC provides a more efficient, scalable, and trainable paradigm for VLMs, allowing for targeted optimization of perceptual capabilities without the instability and complexity associated with end-to-end reinforcement learning of monolithic models.