A new arXiv preprint introduces SPEAR, a method for making vision-language model (VLM) preference labeling more practical in robotic manipulation. The authors note that while VLMs can automate preference labeling, each query carries a substantial inference cost, and the models often abstain from answering. These two issues make straightforward deployment impractical.

SPEAR stands for Selective Preference Elicitation with Adaptive Rewards. As the name suggests, the approach appears to focus on choosing which queries to ask and adjusting rewards over time, rather than querying the VLM indiscriminately. The abstract does not provide full implementation details, but the stated goal is to address the cost and abstention problems that currently limit VLM-guided preference learning.

The paper is available on arXiv under identifier 2610.21621. As a preprint, its claims have not yet undergone peer review, and the abstract is truncated in the source, so the exact mechanism and experimental results are not fully described here.