A 500-million-persona simulation framework evaluated 12 million households and produced 288 million monthly ledger states in under four hours
AspireAI community · arXiv
S2-MAS is a multi-agent simulation framework built around a resource of 500 million synthetic household personas, combining trainable candidate-scoring policies, reusable financial scenarios and explicit state transitions. Trained with SFT plus RLCD on 2.93 million examples, the YN-mini-MLX policy reaches 74.55% accuracy and 38.37% question-macro F1 on held-out consumer financial-choice tasks, improving on the training-question majority baseline by 2.78 and 6.21 percentage points. A completed application evaluates 12 million distinct households across two purchase-funding conditions: 24 million one-shot choices feed 288 million monthly ledger states in 3 hours 59 minutes on 14 GPUs, and joint increases in purchase goals and approved limits produce heterogeneous consumption, cash and repayment paths, with much of the consumption variation arising from transitions between deferral and purchase. The aggregate fit is imperfect even where the per-household error is good: in a separate 1,060-household retrospective experiment the standalone YN-master-9B has 3.42% lower annual MAE than a deterministic fully rational benchmark, alongside a 19.7% lower predicted population mean than observed. The release is a research prototype, and the paper notes its own limits in the information-timing results.
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.