Binding behaviours to events rather than to fixed agent identities let a multi-agent team reassign roles mid-task, and it was the only method that solved the tasks needing sequential reassignment
AspireAI community · arXiv
Effective multi-agent cooperation requires agents to adopt diverse behaviours as task conditions evolve, and to do so at the right moment, yet current multi-agent reinforcement learning frameworks that facilitate this diversity are limited because they bind fixed behaviours to fixed agent identities, leaving them ill-equipped for tasks where agents need to take on different roles at very specific moments. The authors argue that the missing ingredient for defining those behavioural transitions is events, defined as changes in the state of the system that induce qualitative changes in the task, and introduce a framework that decouples agent identity from behaviour, capturing a continuous manifold from which agents instantiate behaviours in response to events. Two elements underpin it: Neural Manifold Diversity, a formal distance metric that remains well-defined when behaviours are transient and agent-agnostic, and an event-based hypernetwork that generates low-rank adaptation modules over a shared team policy, enabling on-the-fly agent-policy reconfiguration. The construction ensures that diversity does not interfere with reward maximization by design. Empirically the framework outperforms established baselines across benchmarks while exhibiting zero-shot generalization, and it was the only method that solved tasks requiring sequential behaviour reassignment. The limitation is that the hypernetwork must be trained for the specific task family, so the framework buys zero-shot transfer across behaviour assignments rather than across environments, and the added machinery is heavier than a fixed-policy team.
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.