| name | sparc-vl-perception-reasoning |
| title | SPARC: Separating Perception And Reasoning Circuits for VLMs |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2602.06566 |
| keywords | ["VLM","Perception","Reasoning","Test-Time Scaling","Two-Stage Pipeline","Efficiency"] |
| description | Improve vision-language model reasoning efficiency by decoupling perception (identifying task-relevant image regions) from reasoning (generating explanations), enabling asymmetric compute allocation. Reduces token overhead while improving accuracy through separate optimization of visual grounding and semantic reasoning stages. |
SPARC: Perception-Reasoning Separation for Efficient VLM Inference
Vision-language models (VLMs) struggle with efficiency when reasoning about complex visual scenes. Traditional approaches interleave perception (identifying relevant image regions) and reasoning (generating explanations) tokens, leading to token bloat and redundant processing. The model wastes capacity re-examining images for each reasoning step rather than establishing visual context upfront.
SPARC decouples VLM inference into two distinct stages mimicking the biological visual system. Stage 1 identifies task-relevant image regions with self-consistency (multiple rollouts), creating high-confidence visual grounding. Stage 2 generates reasoning using only the refined crops, reducing token overhead. This separation enables asymmetric compute allocation—perception scales independently through rollout aggregation while reasoning remains compact and efficient.
Core Concept
Standard VLM reasoning embeds visual tokens throughout reasoning: [visual_tokens, reasoning_tokens_1, visual_tokens, reasoning_tokens_2, ...]. This causes redundancy and token explosion.
SPARC structures inference as two explicit stages:
Stage 1 - Perception: Given image + question, generate coordinates of task-relevant regions (bounding boxes or points). Run this k times with self-consistency, aggregating results.
Stage 2 - Reasoning: Crop image to relevant regions identified in Stage 1, feed only these crops to the model for reasoning and answer generation.
This separates concerns: perception learns "what to look at," reasoning learns "how to reason given visual context." Each stage optimizes independently via LoRA fine-tuning, and the boundary between stages is explicit and controllable.
Architecture Overview
- Stage 1 - Perception Circuit: Generate k candidate region detection rollouts (bounding boxes or point coordinates), apply Weighted Box Fusion (WBF) to aggregate overlapping detections into high-confidence crops
- Stage 2 - Reasoning Circuit: Take merged crops from Stage 1, feed to reasoning module for chain-of-thought generation and final answer
- Asymmetric Compute: Perception stage scales via multiple rollouts; reasoning stage processes single merged crop context
- Fine-tuning Independence: Use LoRA to optimize perception circuit separately from reasoning backbone without coupling their parameter spaces
- Context Engineering: Maintain structured format (region coords, crops, reasoning chain, answer) rather than entangling tokens
Implementation
The implementation requires three components: region detection, box fusion, and two-stage VLM inference.
First, implement region detection in Stage 1:
