An agent that learned its own screening workflow hit a 15.9 percent experimental hit rate on antibiotics against 8.7 percent for the previous expert-led workflow
AspireAI community · arXiv
MoLRSI is a self-improving agent for molecular discovery that optimises not just molecules but the workflow that searches for them, carrying persistent cross-task memory so each campaign improves the next one's strategy. In antibiotic discovery it built a screening workflow balancing antibacterial activity, cytotoxicity and structural novelty, and validated it in the lab: 17 hits from 107 tested compounds, a 15.9 percent experimental hit rate, against 21 of 241, or 8.7 percent, for the previous expert-led workflow on the same training data and screening library. It also reached that level faster than the baselines - on gonorrhoea its first evaluated workflow already beat the final workflows every baseline agent produced after twenty search iterations. The second setting is de novo peptide-MHC binder design, where the agent optimised target binding and specificity together and produced eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs with preferential binding over tested off-targets, while Claude Science produced none. The honest limit: the comparators used Claude Sonnet 5, and Codex used GPT-5.5, so this measures one agent architecture against a specific frontier-model baseline rather than proving the approach wins against every agent.
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.