Async-RBench: nine agents all scored worse when results arrived one at a time instead of in a batch
AspireAI community · arXiv (Detao Zeng et al.)
Agents increasingly run work asynchronously, with a main agent continuing while subagents finish and the task state changing before results land, so the main agent has to decide what to update, what to preserve and which checks to redo. Async-RBench provides 200 executable tasks across eight scenarios, with a scenario controller that regulates result delivery using predefined task-state conditions and an Event Response Contract that specifies what each response must change, preserve, avoid and re-verify. Its Dynamic Replanning Score weights process and outcome equally and is reported separately from task success. Across nine agents, DRS ranged from 19.2 to 45.1, with the lowest scores on duplicate reports and subagent failures. The clearest finding: under the authors' execution protocol all nine agents obtained lower task success with asynchronous delivery than with batched delivery, 26.0 percent against 29.7 percent on average, and prompting three agents to rebuild their plan after every result improved DRS by only 2.6 to 3.5 points. The honest caveat: the delivery schedule is controlled by the benchmark's scenario controller rather than by real workloads, and the results are specific to the nine agents and execution protocol the authors evaluated.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.