Two Worlds, One Challenge: How AI Solves Unfamiliar Problems — Evidence from ARC-AGI-2 and ARC-AGI-3
Year: 2026
Type: preprint
Abstract: What evidence connects an adaptation procedure to better problem solving? We examine two bounded studies of existing ARC solvers rather than equate intervention execution with success. In a Qwen3-based port of NVARC for static ARC-AGI-2, an updating configuration exceeded a zero-learning-rate configuration by 21.39 and 20.98 rounded public-score points in initial and repeated evaluations. In interactive ARC-AGI-3, recovery telemetry establishes that context-rebuilding interventions executed and were followed by environment actions, but not that they improved outcomes. A retained LP85 trace reveals 962 actions with no visible board change, compared with 18 state-changing actions to completion in a different configuration. We propose reporting configuration, intervention execution, subsequent behavior, and task outcome separately. These studies do not identify a universal causal mechanism for adaptation. Version and provenance Author manuscript v2.1a, not a peer-reviewed journal publication. This is a new, independent record for the combined paper, separate from both technical archives. Manuscript content corresponds to GitHub commit 177f555f0cb5284a0521c74b047a4f3435a5e0b4. The PDF is an export of the Word manuscript with its header corrected from WORKING DRAFT v2 to MANUSCRIPT v2.1a; scientific content is unchanged. No new experiments were conducted for this deposit. Separate technical archives ARC-AGI-2: 10.5281/zenodo.23895859 ARC-AGI-3: 10.5281/zenodo.23869166 LP85 supporting material Public LP85 selected-case evidence supplement at immutable commit aab389457620273417a8345b161f1c9028ac55d1. This supplement is referenced, not included as a newly licensed part of the paper. Rights Copyright © 2026 Tsitsino Dzotsenidze. All rights reserved. No additional reuse license is granted. MIT-0 applies only to the separately identified author-owned methodological prose and pseudocode in the LP85 supplement, not to this manuscript. Referenced third-party works retain their own rights.
Author: Dzotsenidze, Tsitsino
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225317239