Work

Mitigating Shortcut Learning through Change-Aware Losses for ARC-AGI

Year: 2026
Type: preprint
Abstract: Learning a new task from a handful of demonstrations is a skill often assumed to require the scale of large language models or test-time training on the demonstrations themselves, while purely in-context, non-LLM approaches have received far less attention. We find that small models trained for this setting are prone to shortcuts: since most output pixels in ARC-AGI are identical to the input, a uniform per-pixel cross-entropy biases the model toward leaving cells unchanged and erasing content, rather than learning the transformation. We propose a change-aware loss that up-weights the pixels that change and balances the loss across colors. In controlled comparisons, it raises accuracy on changing pixels from 52.3% to 59.8% at equal training, while also increasing the number of tasks solved. Using this loss, we train the Change-Aware In-Context Learner (CAIL), a 23M-parameter model that achieves 1.25% on the ARC-AGI-2 public evaluation with \$8 worth of training compute and \$0.0003 per task at inference, without using test-time training.
Language: English
Open Access status: green
OpenAlex ID: W7232009963
Cites:
Cited by:
Related to: 10