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

Researchers introduce SPARC, a modular framework that decouples visual perception from reasoning in vision-language models to improve test-time scaling efficiency. By separating tasks into explicit visual search and conditional reasoning stages, SPARC achieves significant performance gains on visual reasoning benchmarks while reducing computational token requirements by up to 200×.

SPARC addresses a fundamental limitation in current vision-language models: the inability to efficiently scale inference computations for complex visual reasoning tasks. Traditional monolithic approaches entangle perception and reasoning into unstructured chains, causing perceptual errors to cascade throughout inference and requiring expensive reinforcement learning optimization. This framework introduces architectural modularity inspired by biological sensory-cognitive processing, fundamentally changing how VLMs allocate computational resources.

The development builds on growing recognition that test-time scaling—dynamically expanding compute during inference—remains inefficient for visual tasks compared to language-only models. Previous attempts required hand-crafted reward functions and failed to gracefully handle out-of-distribution scenarios. SPARC's two-stage pipeline enables asymmetric compute allocation, allowing models to prioritize resources based on actual bottlenecks rather than uniform scaling.

The performance improvements are substantial: 6.7-point gains on VQA benchmarks and 4.6-point advantages over competing "thinking with images" approaches, achieved while consuming 200× fewer tokens. This efficiency gain matters significantly for deployment, reducing inference latency and computational costs while improving accuracy. The framework's ability to run global search at lower resolutions and concentrate high-resolution processing on relevant regions creates practical advantages for resource-constrained environments.

Looking forward, this modular approach could influence broader VLM architecture design, encouraging separation of concerns across vision-language tasks. The framework's success with compressed contexts positions it well for edge deployment and real-time applications. Subsequent research will likely explore whether similar decomposition benefits other multimodal reasoning tasks and whether the approach generalizes across different VLM architectures and scales.

• →SPARC's modular design separates visual perception from reasoning, enabling more efficient test-time scaling with asymmetric compute allocation
• →The framework achieves 6.7-point improvements on VQA benchmarks while reducing token requirements by up to 200×
• →Explicit visual search stage mitigates perceptual error cascading and eliminates need for complex reinforcement learning with hand-crafted rewards
• →Compressed context processing through variable-resolution handling reduces overall visual token count and computational overhead
• →The approach demonstrates superior out-of-distribution performance compared to monolithic baselines and existing visual-grounding methods