A self-evolving router for multi-agent web queries cut inference compute cost 50.95 percent while beating the strongest routing baseline by 7.23 points
AspireAI community · arXiv (Yitao Xiao, Xu Zhang, Ye Cheng et al.)
Serving web queries at scale with multi-agent systems requires jointly configuring language models, collaboration modes, agent counts, interaction rounds and skills, but that configuration space grows combinatorially while sparse execution rewards provide limited fine-grained supervision for interdependent planning decisions. EvoRouter is a self-evolving multi-agent routing framework driven by insight-guided on-policy distillation. A dedicated expert periodically synthesizes performance-cost feedback from configuration trajectories across tasks into reusable conditional routing insights about capability requirements, returns from resource scaling and interactions among configuration choices; a privileged teacher combines those evolving insights with current execution feedback to give the router dense token-level supervision over its own planning trajectories; and cost-aware reinforcement learning jointly optimises task performance and computational cost. The updated policy then generates new trajectories that support further insight extraction, forming an execution-driven self-improvement loop. Evaluated on 19 in-domain datasets spanning five domains plus two out-of-domain benchmarks, EvoRouter achieves an average in-domain accuracy of 68.69%, outperforming the strongest routing baseline by 7.23 percentage points while reducing inference compute cost by 50.95%. The code is open-sourced. The limitation: the cost saving is measured against routing baselines that are themselves cost-aware to varying degrees, and the accuracy figure is an average across 19 in-domain datasets rather than a single headline task.
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.