An arXiv preprint argues that recursive transformers are sensitive not just to the number of shared parameters, but to how those parameters are arranged across depth. The authors point out that existing static parameter-sharing schemes fall into two extreme reuse topologies, one being layer-wise repetition.
Because the available abstract is truncated, the second topology and the proposed alternative are not described in the text. The core claim, however, is clear: the structure of parameter sharing is an independent design axis that deserves attention alongside parameter count.
This suggests that future work on recursive models may need to consider intermediate or structured sharing patterns rather than relying on the two existing extremes. The paper is available on arXiv under ID 2610.24998.