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

Despite recent successes, *test-time scaling*$-$dynamically expanding the token budget during inference as needed$-$remains brittle for vision-language models (VLMs). Unstructured visual reasoning chains entangle perception and reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers. Reasoning also requires expensive reinforcement learning with hand-crafted rewards. Here, we introduce SPARC (Separating Perception And Reasoning Circuits), a modular framework that explicitly decouples visual perception from reasoning. Inspired by sequential sensory-to-cognitive processing in the brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric compute allocation (e.g., prioritizing perceptual processing under distribution shift), and supports selective optimization (e.g., improving the perceptual stage alone when it is the bottleneck for end-to-end performance). It also accommodates compressed contexts by running global search at lower image resolutions and allocating high-resolution processing only to selected regions, thereby reducing visual token count and compute. SPARC outperforms monolithic baselines and strong visual-grounding approaches across challenging visual reasoning tasks, such as improving Qwen3VL 4B on the $V^*$ VQA benchmark by 6.7 points and surpassing "thinking with images" by 4.6 points in an OOD setting with a 200$\times$ lower token budget.

Artificial intelligence models that analyze both text and images often struggle with complex visual questions because they entangle "looking" and "thinking" into one messy process. When they try to do both simultaneously, small visual mistakes can send the AI down the wrong train of thought, causing it to confidently guess completely wrong answers while wasting significant computing power on long reflection cycles. Inspired by how the human brain processes information in modular blocks, we introduce SPARC. Instead of doing everything at once, SPARC first asks the AI to act like a targeted searchlight, strictly scanning the image to locate the exact coordinates of relevant objects. Once those zoomed-in visual clues are isolated, a separate reasoning step uses only those high-resolution snapshots to logically deduce the final answer. By decoupling perception from reasoning, SPARC drastically reduces mistakes and avoids overwhelming the AI with unnecessary background data. It also allows for allocating more resources to the "look" rather than the "think" process. This streamlined approach allows the model to accurately solve much harder visual puzzles while being 200$\times$ more efficient, making advanced visual AI significantly more affordable, accurate, and reliable.