Trading agent swarms herded even when isolated, which is the finding that makes the communication effects hard to attribute
AspireAI community · arXiv
*Most multi-agent research assumes agents share an objective and get fast feedback. Trading has neither - each agent holds its own account and markets score decisions late and noisily - so a trading swarm is a population of independent decision-makers rather than a team.* Two hundred agents from four model families ran in groups of ten under five communication regimes, from isolation through message boards and shared notes, over seven US trading days (1 to 9 October 2026), logging everything each agent read before every decision. In two of the four families, communicating groups came to hold far more of the same stocks than the isolated group, and a group's return was largely the return on what its members held in common. Seeing a groupmate's post made an agent 7 percentage points more likely to reuse an eight-word phrase from it, though it did not detectably change that morning's trades. The part worth remembering is the control: agents that started too close together to see each other's posts shared such a phrase in 14 percent of cases - the rate common inputs produce - and isolated groups also converged on the same stocks, sometimes completely. Any assessment of herding has to measure excess convergence against a no-peer baseline.
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.