Tag
CERN and Mistral have signed a statement of intent for a long-term partnership to build a Large Physics Model — the first foundation model for collider physics — trained on billions of simulated LHC proton-proton collisions to support tasks from particle reconstruction to anomaly detection.
The paper proposes a pipeline that automatically generates ontology-grounded, reasoner-verified multiple-choice benchmarks for evaluating LLM logical reasoning in scientific AI, validated on three ontologies (Pizza, PMDco, DOID) across six LLMs.
A research team audits OpenAI's Lean 4 formal proof of 3D Navier-Stokes blow-up, showing that while mathematically valid, the solution physically breaks down at the atomic scale (fluid vaporizing at 0.7 nm) — a case of specification gaming. They propose adding a physics-boundary layer to neuro-symbolic AI systems and open-source their verification scripts.
Claude reportedly solved an important mathematical problem, drawing attention from Scientific American. The achievement highlights advances in AI-driven mathematical reasoning.
论文提出 Fold2Reason，一种通过蛋白质折叠数据对语言模型进行后训练的方法，结果表明结构密集的科学数据可以将通用推理基准的宏平均准确率从45.09%提升至48.33%，并显著优于匹配的随机对照实验。
Terence Tao discusses concerns about solving mathematical problems prematurely using purely AI-powered methods, a sentiment the author believes also applies to programming.
NS-Copilot is an LLM-driven multi-agent system that autonomously supports end-to-end workflows for diverse neuroscience analysis tasks, outperforming baselines on key benchmarks.
Tamarind introduces a molecular AI model router that automatically selects the best model for specific inputs based on benchmarks and input characteristics, rather than relying on average performance.
FirstPass is a large-scale peer review dataset from Nature Communications, covering multiple scientific domains and multi-round dialogues to improve AI models for scientific judgment.
The paper introduces K-Bench 01, a benchmark for evaluating AI agents on real scientific requests, revealing that no model consistently meets the threshold for acceptable performance, with overclaiming as a common failure.
This paper introduces the ALD/E-ImageMiner benchmark for multimodal comprehension of scientific images and discusses future research directions for general-purpose scientific AI, based on insights from the ICDAR 2026 competition.
Inherent Labs introduces Faraday, a 27B-parameter AI Scientist agent trained via long-horizon reinforcement learning to replicate scientific research, outperforming Claude Opus 4.8 and GPT-5.5 on paper replication tasks.
A 27B parameter AI scientist agent named Faraday surpasses Claude Opus 4.8 and GPT-5.5 on paper replication tasks by using a scalable reinforcement learning approach called Replica.
The paper presents REDI, an open-source framework that automates the transformation of raw scientific datasets into AI-ready data through a unified five-stage pipeline, with companion tool SetGo for FAIR compliance, evaluated across multiple scientific domains.
OpenScience is an open-source alternative to Claude Science, supporting multiple AI models and over 250 research skills, with native Atlas integration for reproducible research graphs and user-controlled infrastructure.
NVIDIA announces its Vera CPU will power new supercomputers at Los Alamos National Laboratory, delivering significant performance improvements for agentic AI simulations and scientific workloads.
OpenAI and Molecule.one collaborated to have their AI systems (GPT-5.4 and Maria) autonomously select research areas, generate proposals, and run experiments in organic chemistry, achieving yield improvements for 88% of tested reactions — a first for AI-driven open-ended scientific discovery.
Notes2Skills is a two-stage framework that converts laboratory notes into verifiable skills for AI agents while preserving author uncertainty, enabling safer scientific AI systems.
Anthropic's Claude, a general-purpose AI model without chemistry fine-tuning, outperformed specialized software like ChemDraw and MestReNova in NMR analysis, suggesting that the bottleneck in scientific AI has shifted from model capability to workflow design.
This paper introduces I-SAFE, a post-hoc distributional auditing framework for scientific AI models using Wasserstein Coherence Metrics, which reveals structural differences in model outputs that accuracy-based evaluation fails to capture. Demonstrated on drug-target interaction prediction, the framework is model-agnostic and applicable to any domain with structured inputs and external priors.