aiqbench analysis
This research introduces SARBLOH, an agent that leverages spatial perception rather than brute-force planning to solve the ARC-AGI-3 benchmark with significantly higher sample efficiency. By converting raw integer grids into geometric briefings, the system reduces token usage by nearly 90% while achieving near-perfect accuracy on held-out tasks involving geometric invariants. The study highlights a critical shift toward perception-driven architectures for fluid intelligence, demonstrating that preserving spatial topology is essential for scalable interactive reasoning.
Transforming Integer Grids to Spatial Intuition: Perception, Not Planning, Governs ARC-AGI-3
arXiv:2611.06793v1 Announce Type: cross Abstract: Sample-efficient fluid intelligence requires agents that rapidly form working models of unfamiliar interactive environments. ARC-AGI-3 tests this across uninstructed visual games scored by Relative Human Action Efficiency, which penalizes exploratory waste quadratically. When frontier models read raw integer grids, serialization destroys spatial topology, and they spend hundreds of actions rediscovering basic boundaries and symmetries. We introduce SARBLOH, an open-weight agent built on the premise that spatial perception, not planning capacity, governs interactive reasoning on ARC-AGI-3. A zero-parameter O(HW) perception engine turns each frame, in 35 ms, into an object inventory with bipartite temporal tracking, a geometric briefing that separates proven invariants from affordance hypotheses, and an annotated canvas. Paired with a code-as-action runtime, the agent writes algorithmic hypotheses instead of low-level moves. Across eight presentations, grounded perception raises the rubric score from 3/28 to 25/28 points while cutting token use by 89.4%. On 23 held-out games, geometric invariants generalize with 100% accuracy, while single-frame affordance guesses fall to 6-50%, so active verification probes remain necessary. On the public set, Run A (briefing off) solved 143 of 183 levels in 8.2 hours at 59.4% single-pass efficiency; Run B (briefing on) reached 63.5%. Within the nine-hour budget on one Kaggle RTX PRO 6000 GPU, SARBLOH scored 16.53 on the hidden Kaggle evaluation, 13.7x the first Milestone winner.
Abstract:Sample-efficient fluid intelligence requires agents that rapidly form working models of unfamiliar interactive environments. ARC-AGI-3 tests this across uninstructed visual games scored by Relative Human Action Efficiency, which penalizes exploratory waste quadratically. When frontier models read raw integer grids, serialization destroys spatial topology, and they spend hundreds of actions rediscovering basic boundaries and symmetries. We introduce SARBLOH, an open-weight agent built on the premise that spatial perception, not planning capacity, governs interactive reasoning on ARC-AGI-3. A zero-parameter O(HW) perception engine turns each frame, in 35 ms, into an object inventory with bipartite temporal tracking, a geometric briefing that separates proven invariants from affordance hypotheses, and an annotated canvas. Paired with a code-as-action runtime, the agent writes algorithmic hypotheses instead of low-level moves. Across eight presentations, grounded perception raises the rubric score from 3/28 to 25/28 points while cutting token use by 89.4%. On 23 held-out games, geometric invariants generalize with 100% accuracy, while single-frame affordance guesses fall to 6-50%, so active verification probes remain necessary. On the public set, Run A (briefing off) solved 143 of 183 levels in 8.2 hours at 59.4% single-pass efficiency; Run B (briefing on) reached 63.5%. Within the nine-hour budget on one Kaggle RTX PRO 6000 GPU, SARBLOH scored 16.53 on the hidden Kaggle evaluation, 13.7x the first Milestone winner.
24 pages, 17 figures, 8 tables. Submitted to the ARC Prize 2026 paper track. Code: this https URL
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2611.06793 [cs.LG]
(or arXiv:2611.06793v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2611.06793 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 330 words
Related coverage
Transformer-Based Token Fusion and Dynamic Graph Planning for Audio-Visual Navigation
arXiv · Sep 16, 2026
Learning from Distributed Eyes: Leveraging Collaborative Perception for Automated Model Adaptation
arXiv · Sep 17, 2026
Edge-Aware and Content-Adaptive Infrared Gas Leak Detection for Industrial Safety Monitoring
arXiv · Oct 1, 2026
AUTOPILOT An Advanced Perception, Localization and Path Planning Techniques for Autonomous Vehicles Using YOLOv7 and MiDaS
arXiv · Oct 6, 2026