In scientific discovery tasks where the governing law was altered, familiar priors made agents worse, and good explanations hid quantitative failures
AspireAI community · arXiv (Zhao, He, Liu, Li)
Agents are increasingly used for scientific discovery, but existing benchmarks mix law inference with programming, score data fitting instead of prediction on unseen inputs, and do not separate empirical evidence from what the model already believes. SciGame addresses all three by converting scientific literature into structured cards, each of which yields a pair of dimensionally consistent environments - an Original and a Counterfactual - with the same task interface. It ships two collections: a Historical set of canonical laws, and a live Temporal set curated from papers published after model knowledge cutoffs, evaluated in candidate-guided Identification and open-ended Discovery settings under Prior-only, Fixed and Adaptive protocols. Across five frontier models, three findings stand out. Familiar priors hurt performance once the governing relation is altered. Adaptive experimentation helps the strongest models but harms weaker ones under limited budgets. And good explanations can hide quantitative failures - agents that interpret evidence correctly still fail to recover the exact symbolic form, which the authors name the main bottleneck. The limitation is the benchmark's own scale: it is a controlled card-based construction, not a live laboratory, so it measures reasoning about laws rather than the mess of real experimental work.
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.