Letting the main agent write the collaboration programme instead of following a fixed one gained 10.90 points on six benchmarks with the same backbone
AspireAI community · arXiv
Most multi-agent systems are locked into a predefined collaboration structure. DynaFlow puts free-form executable workflow synthesis into the main agent's action space, so the agent decides when collaboration is worth it and writes a task-specific programme that coordinates sub-agents through sequential or parallel execution, structured information exchange and verification. Execution results and traces then feed back into the next action. Only the main agent is trained, through supervised fine-tuning on distilled trajectories plus refinement on examples filtered for unsupported premises, biased verification and information-flow defects; every sub-agent stays frozen, which separates orchestration learning from executor capability.
Across six benchmarks spanning mathematical reasoning, scientific question answering and multi-hop retrieval, DynaFlow reaches 88.08% overall accuracy in Forced mode, beating the strongest evaluated baseline by 10.90 percentage points on the same Qwen3.5-27B backbone. Workflow analysis shows more explicit use, verification and refinement of upstream outputs after training.
The honest limit: the paper trains a single 27B backbone, so it does not establish that free-form orchestration survives at frontier scale.
https://arxiv.org/abs/2610.34467
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.