Surrogates picked purely on validation error violated causality, superposition and boundary conditions, while audit-enabled ones passed the checks at comparable error
AspireAI community · arXiv
Agentic discovery can yield mechanics surrogates with low validation error that nonetheless violate physical requirements. Physics-Audited Agentic SciML (PA-SciML) combines large language model agents with physics auditing, where physical requirements encoded as contracts such as superposition, boundary conditions and causality are fixed before candidate search and tested on predicted fields without reference solutions, including through an adversarial search. Audit results guide later proposals and model selection but are reported separately from validation error, and single-edit tests document reusable design changes. Across four solid-mechanics problems of increasing complexity, from parametric elasticity to heterogeneous stiffness fields and cyclic plasticity with combined hardening, the audit reveals causality, zero-input and stress-bound violations in surrogates selected by error alone. In elastodynamics, two surrogates reach nearly equal error, but only the audit-enabled one is causal by construction; audit-enabled surrogates pass the required checks at sampled inputs while matching or exceeding audit-disabled accuracy in most comparisons. The limitation is that the audit checks are only as good as the contracts written, so a physical requirement nobody wrote down is still violated silently, and the evaluation covers four solid-mechanics problems rather than a broad scientific workflow.
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.