arXiv cs.AI· Binnan Liu, Yechi Ma, Tian Xie, Wei Hua·· 12 小时前AI 评分35
TraceViT：用有据可依的轨迹监督训练视觉抽象推理模型
TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning
AI 导读
TraceViT 通过将程序化任务实现改写并验证为语义单调的变换链，把每个解法拆成中间网格状态，用软轨迹对齐只约束顺序，让循环视觉推理器的每次迭代都有任务参考和对象工作区作为依据。该模型在 ARC-AGI-1 上达到 67.8% pass@2，在 ARC-AGI-2 上达到 24.3%。ARC-AGI-1 上的消融实验显示，轨迹监督只有与 grounding 结合时才带来收益，代码与数据将公开。
正文 · 原文
Abstract:The Abstraction and Reasoning Corpus (ARC) tests whether a model can infer an unseen transformation from a few input-output examples and apply it to a new grid. Looped visual reasoners refine predictions over multiple iterations, but conventional training constrains only the final output, leaving intermediate refinements unconstrained. We propose that these refinements should instead follow the transformation step by step. We introduce TraceViT, a looped visual reasoner trained with semantically monotonic transformation chains. We obtain these chains by rewriting and verifying programmatic task implementations, decomposing each solution into intermediate grid states. Each iteration is grounded by a task reference derived from the few-shot demonstrations and an object workspace representing the current grid state. Because these chains may differ in length from the loop, soft trace alignment enforces only their ordering, letting the model allocate iterations freely. TraceViT achieves 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2. Controlled ablations on ARC-AGI-1 show that trace supervision becomes beneficial only when paired with grounding. Code and data will be available at this https URL.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.29586 [cs.CV] |
| (or arXiv:2607.29586v2 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2607.29586 arXiv-issued DOI via DataCite |
Submission history
From: Binnan Liu [view email]
[v1]
Fri, 31 Jul 2026 16:09:44 UTC (10,476 KB)
[v2]
Thu, 5 Nov 2026 07:48:27 UTC (10,638 KB)
来源：arXiv cs.AI · arxiv.org