Coding agents that score the same alone diverge sharply when work piles up: two systems 0.6 points apart in isolation delivered 54 and 68 percent of the same tasks
AspireAI community · arXiv
Coding agents are becoming persistent assistants that hold a stream of pending work at once - a build, a data-processing job and a bug fix - and must decide what to take on, how to overlap tasks, and how to stay inside CPU and memory limits. Existing benchmarks evaluate each task in isolation and so cannot measure this. AOS-Bench evaluates an agent system as a workload manager: verified tasks arrive over time under shared CPU and memory limits, with the same tasks run one at a time as a per-system reference.
Across five model-harness pairs, isolated task success does not carry over evenly. On heavy workloads every system loses 13 to 23 points. On medium workloads, systems span 11 points in isolated success but 26 points in stream success, and two systems only 0.6 points apart in isolation deliver 54 percent and 68 percent of the same tasks. Execution-policy interventions trace much of the gap to task intake: serial dispatch leaves 73 percent of heavy-workload tasks unclaimed. More elaborate management does not necessarily help - a fixed concurrency of two lands within four points of the best agent-managed policy on both medium and heavy workloads, and under every policy median CPU use stays below 30 percent of the provisioned cores while tasks wait.
The honest limit: five model-harness pairs is a narrow sample, and the paper does not yet tell a reader which harness to pick - the finding is that the isolated score is simply not the metric that predicts queue behaviour.
https://arxiv.org/abs/2610.33178
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.