AI-assisted results in the 24 Oct 2026 sweep

Your AI and this story

Key facts

• Own entries: Nakayama, local height pairing without the Hodge conjecture (2610.30574); Épi autonomous research agent (2610.24944)
• Lab-in-the-loop aptamers (Google Research, Stanford, Google Cloud, Google DeepMind; Cutts, …, Hassidim, Matias, Rao, Pawlosky; 2610.24806): agentic pipelines built on ‘Google Gemini Pro and Flash models’ designed aptamer binders against α-thrombin and myoglobin over three design–test–improve wet-lab rounds (~1,000 candidates per round, ~4,400 sequenced constructs). On-target hit rates on α-thrombin rose from 0.0% to 26.0% to 42.1%; on myoglobin 10.2% → 3.7% → 33.5%. Agent-engineered literature motifs reached apparent K_D 24.6–133 nM, while strictly de novo scaffolds stayed above 1 µM. The agents found cross-reactivity in the myoglobin screen (specificity fell from 75.0% to 19.1%) and restored it to 72.4%
• DEScover (Jiawei Chen, Xiang I.A. Yang, Ping He; Iowa State, Penn State; 2610.29850, under consideration at J. Fluid Mech.): four LLM agents (discovery, review, code implementation, evaluation) propose and test subgrid length scales for detached eddy simulation from a one-sentence modelling idea. On a flat plate, backward-facing step, free shear layer and iced airfoil, the best candidate beats the standard maximum-cell-size length scale and matches or slightly beats the expert shear-layer-adapted one; ~25 minutes per discovery run excluding CFD
• Lean Finder V2 (Simon Fraser, Rice, Pittsburgh; 2610.30843): Mathlib semantic search trained on adversarially generated short, ambiguous queries; ≥11% relative improvement on MathlibQR; plugged into Claude Code it raises proof success on FATE-H to 39.3% from the next-best 36.0%; frontier LLM judges (Claude Opus 5, GPT-5.6 Sol) rank it most helpful 18+% more often on real user queries
• Ashbaugh–Benguria stability (Long-Han Huang, Yuxuan Zhou, Wenming Zou; 2610.31223): sharp quantitative stability on bounded convex domains, optimal exponents 5/2 (Fraenkel asymmetry) and (n+4)/2 (Hausdorff distance); the quadratic estimate proposed as Open problem 7.65 in the cited survey fails. ‘ChatGPT was used to assist in discussing certain mathematical details, checking mathematical arguments and references, and preparing the manuscript’
• Local dyadic conjecture (Alghasi, Guenin, Tunçel; Waterloo; 2610.29197): a local version of Seymour’s dyadic conjecture for ideal clutters, with an important special case proved; ‘The final draft of this manuscript was reviewed by Anthropic (Claude Fable 5.1, October 2026) and the authors edited the draft in response to such a review’
• AI-assisted addendum to a 2018 thesis (Haitham Masaud Amar; 2610.29637): definitions and theorems on cooperation and competition in trip planning, with Lean 4 checks (Mathlib v4.24.0). The author says OpenAI’s 6 Oct 2026 release of mathematical manuscripts inspired it. An example of that release pushing practitioners to formalize older work
• Language or editing only: simultaneous popular differences in finite fields (Xuancheng Shao, Weiwen Zhang; 2610.28226, ‘ChatGPT (OpenAI) was used by the authors to assist with polishing’); SympFormer (2610.31600, ChatGPT and Codex for editing, implementation and ‘exploratory discussions of mathematical formulations’)