Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Deferred-write commit path for TreeWY

Determine whether implementing a deferred-write commit path for TreeWY, analogous to ReplaySSM’s deferred state materialization, closes the observed chain-throughput gap between TreeWY and ReplaySSM, and test this implementation hypothesis directly.

Background

The paper compares TreeWY with ReplaySSM on Qwen3.5-35B-A3B using otherwise identical serving sweeps. Although both methods remove per-position recurrent-state snapshots and free comparable KV-cache headroom, ReplaySSM achieves higher chain throughput and lower per-token cost than TreeWY across the reported loads.

The authors hypothesize that the difference arises from scheduling and, specifically, from ReplaySSM deferring state writes on verification steps while TreeWY materializes the accepted state on every commit. They explicitly state that this attribution has not been validated because they have not implemented or tested a deferred-write commit path for TreeWY. Establishing whether that implementation change closes the performance gap is therefore left unresolved.

References

That margin is not the solve (the solve is shared): our working hypothesis is that it is the scheduling around it, since they defer writing state on a verify step where we materialize the accepted state on every commit, but this is an attribution, not a validated claim: we have not prototyped a deferred-write commit path ourselves to check that it actually closes the gap.

— TreeWY: Speculative Verification for Gated DeltaNet Hybrids  (2608.20961 - Ghantasala, 21 Aug 2026) in Appendix, Section “ReplaySSM comparison” (Appendix~\ref{app:replayssm})