TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning

Researchers introduced TraceViT, a looped visual reasoner trained with step-by-step transformation traces derived from verified programmatic solutions.

The model grounds each iteration in the task prompt and current grid state, and uses soft trace alignment when trace length and loop length do not match.

TraceViT reached 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2.

The result suggests that supervising intermediate reasoning steps, especially with grounding, can improve performance on hard abstraction tasks like ARC.