SPARC：用于VLM测试时间缩放的分离感知和推理电路

Niccolo Avogaro* 1 2 Nayanika Debnath* 1 2 Li Mi1 Thomas Frick2 Junling Wang1 Zexue He3 Hang Hua3Konrad Schindler1 Mattia Rigotti2 1. Introduction Abstract Despite recent successes,test-time scaling—dynamically expanding the token budget duringinference as needed—remains brittle for vision-language models (VLMs). Unstructured visualreasoning chains entangle perception and rea-soning, leading to long, disorganized contextswhere small perceptual mistakes may cascadeinto completely wrong answers. Reasoning alsorequires expensive reinforcement learning withhand-crafted rewards. Here, we introduce SPARC(Separating Perception And Reasoning Circuits),a modular framework that explicitly decouplesvisual perception from reasoning. Inspired by se-quential sensory-to-cognitive processing in thebrain, SPARC implements a two-stage pipelinewhere the model first performs explicit visualsearch to localize question-relevant regions, thenconditions its reasoning on those regions to pro-duce the final answer. This separation enablesindependent test-time scaling with asymmetriccompute allocation (e.g., prioritizing perceptualprocessing under distribution shift), and supportsselective optimization (e.g., improving the percep-tual stage alone when it is the bottleneck for end-to-end performance). It also accommodates com-pressed contexts by running global search at lowerimage resolutions and allocating high-resolutionprocessing only to selected regions, thereby reduc-ing visual token count and compute. SPARC out-performs monolithic baselines and strong visual-grounding approaches across challenging visualreasoning tasks, such as improving Qwen3VL 4Bon theV∗VQA benchmark by 6.7 points and sur-passing “thinking with images” by 4.6 points inan OOD setting with a200×lower token budget.arXiv:2602.06566v3 [cs.CV] 24 Jun 2026 Multimodal Vision-Language Models (VLMs) have becomethe de facto standard in visual reasoning and perception (Liet al., 2025). VLMs are architectures that combine visualand textual inputs. By aligning a vision backbone with anLLM (Huang et al., 2023), they extend the impressive NLPcapabilities of LLMs to the vision realm (Alayrac et al.,2022; Chen et al., 2023b; Hua et al., 2025; Li et al., 2023;Chen et al., 2023a; Liu et al., 2023; Zhu et al., 2024; Penget al., 2024; Achiam et al., 2023; Karlinsky et al., 2025).Among the capabilities that VLMs inherit from LLMs isChain-of-Thought (CoT) reasoning (Wei et al., 2022), atest-time compute mechanism to iteratively generate theoutput step-by-step, which can be optimized via Reinforce-ment Learning and has been popularized by models likeChatGPT-o1 (OpenAI, 2024) and DeepSeek-R1 (Guo et al.,2025). Works like ViGoRL (Sarch et al., 2025) and Deep-Eyes (Zheng et al., 2026) have demonstrated that multi-modal chain-of-thought reasoning, obtained by interleavingpure text CoT reasoning with image content, can be explic-itly grounded to the relevant visual evidence in the image viaa multi-turn workflow that calls appropriate image analysistools. In this so-called “thinking with images” paradigmfirst introduced in the OpenAI ChatGPT-o3 report (OpenAIResearch, 2025), the model alternates between reasoningsteps and perceptual actions (like selecting a region of in-terest in the image). Such grounded multi-modal CoTs canyield significantly better performance in visual reasoningtasks, especially when it comes to high-resolution percep-tion where one must repeatedly focus attention on small butdecisive image details. A core issue of “thinking with images”, and multi-modalCoT reasoning in general, is that learning is considerablymore complex than for standard, text-only reasoning: theLLM must acquire the ability to manage multi-turn conver-sations and tool calls that repeatedly mix visual and reason-ing tokens within the context window (Sarch et al., 2025;Su et al., 2025a; Zheng et al., 2026; Kumar et al., 2025).This is not only computationally expensive, but also morebrittle, particularly for smaller models whose performancerapidly degrades when faced with long token sequences due to visually heavy contexts and extended reasoning chains(Tian et al., 2026), which amplify VLMs’ difficulty withfine visual details (Rahmanzadehgervi et al., 2024) and theirtendency to fall into bias-driven “mirage reasoning” (Voet al., 2026; Asadi et al., 2026). Furthermore, a monolithicapproach is inflexible and lacks a mechanism to adapt theallocated compute to the difficulty of the vision task: whento terminate the response is left to the LLM. results than native “thinking with images”; moreover, weshow that it has a number of interesting properties. First, when using the two-step pipeline with an efficientinitial IRD step it becomes possible to scale perception attest time independently from reasoning. As an example,employing self-consistency over eight roll-outs of the IRDstep, with a sharedKV-cache, creates only a few additionaltext tokens and an additional crop, but boosts performanceof the ful