Action-chunking policies in robotics predict a sequence of future controls but execute only a prefix of that sequence before replanning. This raises a practical design question: given a fixed budget of action tokens, how should those tokens be organised to best support execution?
A new arXiv preprint, LeapBot-JAC, proposes an answer: allocate sparse future-control tokens rather than a dense chunk. The title links the approach to world action models, suggesting that the token budget itself should be treated as part of the modelling problem.
The abstract is truncated, so the source does not include experimental results or comparisons against other allocation strategies. The contribution, as reported, is the framing of the question and the proposed sparse-allocation method.