Memory and collaboration architectures that evolve separately drift apart, so aligning them through shared skills and counterfactual trajectories gave the best results on ALFWorld and WebShop
AspireAI community · arXiv
LLM-based multi-agent systems have shown real capability, but their memory and collaboration architectures often stay static, which constrains continued adaptation. Recent work evolves memory or collaboration based on interaction experience, treating them separately - and that is where the problem sits. Updates to these interdependent components are not necessarily aligned: an evolved workflow may require experience the memory does not provide, while evolved memory may encode guidance the workflow cannot use.
CoEvolve is a skill-centred framework for aligned memory-collaboration evolution. It represents reusable local collaboration patterns as skills that link evolving collaboration decisions to their supporting memories, then compares original and counterfactual trajectories to identify local evolution signals, which coordinate updates to the skill library, skill memory and executor memories across rounds. Experiments on the ALFWorld and WebShop benchmarks confirm CoEvolve achieves the best overall performance among the evaluated baselines, and ablation studies validate the contribution of both aligned co-evolution and counterfactual-guided evolution.
The honest limit: the evidence rests on two benchmark environments, ALFWorld and WebShop, which are the standard proving grounds but far simpler than a real deployment. The paper reports best-against-baselines ordering rather than large absolute margins, and the mechanism is a framework that must be wired into an existing multi-agent system rather than a drop-in component.
https://arxiv.org/abs/2610.33539
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.