Slots, Transitions, Loops: Learning Composable World Models for ARC

2026-06-10

AI Summary

This work addresses the challenge of in-context rule induction from few examples in Abstraction and Reasoning Corpus (ARC) tasks by proposing Loop-OWM, a novel architecture that models visual-symbolic rules as composable transition processes over structured world states, eschewing reliance on language descriptions or program search. The approach introduces color prototype slots and demonstration-conditioned task summaries, integrated within a recurrent transition mechanism featuring dense propagation and slot-conditioned refinement to enable efficient rule learning. Evaluated on the ARC-1 and ARC-2 benchmarks, Loop-OWM substantially outperforms existing recurrent and non-recurrent baseline models, even those with comparable or larger parameter counts.

Abstract

ARC tests in-context rule induction: given a few input-output demonstrations, a model must infer the hidden rule and apply it to a new query. While many approaches express ARC rules through language, code, or symbolic programs, ARC itself is visual-symbolic: rules appear as grid transitions over objects, colors, shapes, and spatial relations. We introduce Loop-OWM, an object-centric world-modeling architecture that learns these rules as composable transitions over structured states. It combines color-prototype slots, demonstration-conditioned task summaries, and a looped transition model with dense propagation and slot-conditioned correction. On both ARC-1 and ARC-2, Loop-OWM outperforms non-looped and looped baselines with comparable or fewer parameters. These results suggest that ARC rules can be learned not only as language descriptions or searched programs, but also as transitions over visual-symbolic world states.

Problem

• ARC
• visual-symbolic reasoning
• rule induction
• world models
• composable transitions

Innovation

• composable transitions
• object-centric world model
• color-prototype slots
• looped transition model
• visual-symbolic reasoning