TeamGym found AI agents lose 22.9 points of accuracy when pivotal information is distributed across participants, and a further 19.9 points when they must retrieve it themselves
AspireAI community · arXiv
TeamGym, an arXiv paper submitted October 27, 2026, argues that language-model agents are still trained and benchmarked in isolation even though most consequential decisions involve teams integrating information distributed across participants. The authors build team decision problems from real legal, finance, medical and software-engineering corpora and use pivotal information, a minimal set of facts each necessary and collectively sufficient for a correct final decision, to control how information is distributed and to trace whether agents find it. Across eight language models and twelve source corpora, distributing pivotal information across participants lowered success rate by 22.9 percentage points on average versus handing the same information to one model, which they call a communication gap, and requiring participants to identify and extract pivotal information from source material lowered it by a further 19.9 points, a retrieval gap. Training policies with reinforcement learning on team decision loss improved final-decision quality and induced emergent coordination in participation, speaker selection and communication. The limitation is that the tasks are constructed from documents and multi-agent LLM settings, so the gaps are measured in simulation rather than in deployed human-AI teams.
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.