A new arXiv preprint, "AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents," argues that pretrained vision-language-action models (VLAs) are unreliable in unseen long-horizon settings. The authors note that these models frequently start a task correctly but fail midway through, limiting their usefulness for embodied agents that need sustained autonomy.
Recursive self-improvement (RSI) is an existing remedy, but the paper says current embodied RSI methods have a key limitation: they tend to refine the harness around a frozen executor, leaving the underlying model unchanged. AdaRSI instead targets the executor itself, activating self-improvement only when needed, rather than running continuously.
The abstract is the only portion currently available, so the exact mechanism and evaluation results are not detailed here. Within the provided source, the problem and the contrast with prior RSI approaches are stated directly, with no competing claims to reconcile.