The Abstraction and Reasoning Corpus (ARC) and its variants test whether models can solve completely new reasoning tasks, yet they often conflate genuine rule learning with shortcut exploitation.
Vision ARC (VARC) augments a pre‑trained backbone with a trainable task embedding that is meant to capture the transformation rule of each task. In the original VARC, test‑time training (TTT) updates both the backbone and the embedding jointly.
We introduce a two‑step TTT protocol: first fine‑tune only the task embedding (Embed‑TTT), then freeze it and fine‑tune the backbone.
Across ARC‑AGI‑1, ConceptARC, and two controlled datasets with known rules, Embed‑TTT consistently yields higher‑quality embeddings that align better with the underlying rules, improve embedding‑based retrieval, and enable accurate linear probing of known rules.
Optimizing only the task embedding (less than 0.01% of model parameters) already solves a non‑trivial fraction of ARC‑AGI‑1, ConceptARC, and Mini‑ARC tasks, while the full two‑step pipeline further boosts final performance.
The embeddings recover the geometric structure of parametric rules and exhibit compositional capabilities that allow rule‑wise interpolation but not extrapolation, supporting a clearer separation between rule induction and rule execution in ARC‑like evaluations.
These results motivate benchmarks that more distinctly separate in‑distribution from out‑of‑distribution rules.
Review