TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning
TraceViT trains visual reasoners to make each intermediate step match a verified transformation chain, not just the final grid.
The paper targets ARC tasks, where a model must infer a grid transformation from a few examples and apply it to a new case. Its authors build step-by-step grid traces by rewriting and verifying programmatic solutions, then use soft alignment so model iterations preserve the trace order without requiring equal lengths. Trace supervision helped in ablations only when paired with grounding from the demonstrations and the current object workspace. Reported results are 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2. ArXiv · AI/CL/LG's note
The paper targets ARC tasks, where a model must infer a grid transformation from a few examples and apply it to a new case. Its authors build step-by-step grid traces by rewriting and verifying programmatic solutions, then use soft alignment so model iterations preserve the trace order without requiring equal lengths. Trace supervision helped in ablations only when paired with grounding from the demonstrations and the current object workspace. Reported results are 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2. ArXiv · AI/CL/LG's note
score 5