Decentralised LLM collaboration improved when agents maintained beliefs about what each other had said
AspireAI community · arXiv
Recent work has explored optimising LLM collaboration through multi-agent reinforcement learning, but many approaches still rely on predefined workflows, which limits efficiency and scalability. Decentralised LLM collaboration is hard because the agents execute independently with limited communication, share a joint objective, and must infer useful information from the communication traces they do receive. BCMAAC is a Belief-Conditioned Multi-Agent Actor-Critic method that maintains beliefs over contextual information and uses them to facilitate training. Experiments on Minecraft game-playing, web searching and tool-using tasks show that BCMAAC improves task performance and collaborative behaviour compared to naive actor-critic baselines, with the improvements particularly pronounced in multi-turn tasks with richer contextual information, where belief modelling provides a clear advantage. The paper was accepted at the NeurIPS 2026 workshop on PTA, From Pretrained Representations to Acting Agents. The limitation is that this is still multi-agent RL with LLM policies, so it inherits that method's core weakness: the training signal is expensive to obtain, and the belief representation is only as good as the communication traces the agents produce for it to condition on.
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.