Exploring asynchronous or partially-synchronous Combee variants

Investigate asynchronous or partially-synchronous variants of Combee’s per-iteration parallel execution and aggregation, analogous to asynchronous stochastic gradient descent, to evaluate whether they improve throughput in heterogeneous deployment environments without degrading learning quality.

Background

Combee currently assumes synchronous parallel execution within each iteration when aggregating reflections into context updates via parallel scan. Synchronous coordination can limit throughput or complicate deployment in heterogeneous environments.

The authors suggest exploring asynchronous or partially-synchronous alternatives, inspired by asynchronous approaches in distributed training, as a potential route to higher throughput. This remains untested and is posed as future work.

References

While Combee demonstrates consistent improvements across our evaluation settings, several aspects remain open for future work. Finally, the current design assumes synchronous parallel execution within each iteration; exploring asynchronous or partially-synchronous variants, analogous to asynchronous SGD in distributed training, could further improve throughput in heterogeneous deployment environments and is an interesting direction we leave for future investigation.

— Combee: Scaling Prompt Learning for Self-Improving Language Model Agents  (2604.04247 - Li et al., 5 Apr 2026) in Section: Limitations and Future Work