A new arXiv preprint introduces LabAgent, a system that uses AI agents to tailor research hubs for scientific discovery. The work frames science as an ongoing, cumulative effort where prior methods are reused and extended, and points to lab staffing changes as a challenge. The abstract available is truncated, so full details of the method and evaluation are not yet clear.
Researchers describe AgentRivet, a system that automatically generates Rivet analysis routines from published journal articles. Rivet is a C++ toolkit used in collider physics to compare new theoretical models against preserved experimental measurements. The work aims to reduce the manual effort of encoding published analyses into reusable software.
A new arXiv preprint introduces OpenAI4S, a system that frames computational research by AI co-scientists as sessions in which code serves as the action taken at each step. The authors argue that long-running studies require persistent computational state and provenance so that workflows remain inspectable, resumable and reproducible.
Researchers describe an Immune World Model that aims to capture immune processes across cellular, tissue, and patient-level scales rather than treating them in isolation. The model was assembled using a governed evolutionary AI Scientist framework, according to the preprint. It is positioned as a tool for forecasting and for generating hypotheses about immune therapies.
A new arXiv preprint describes Atria Dawn Preview, a foundation language model built around agentic capabilities and aimed at scientific research tasks. The authors frame the work around the idea that AI agents are increasingly involved in building their own successors, which they say changes how intelligence is produced and how human researchers fit into that process. The abstract is accompanied by an announcement-type listing indicating a first submission, and no independent evaluations are cited in the provided text.
A new arXiv preprint argues that molecular design should not be reduced to finding candidates that lock into one bioactive shape, since real molecules exist across a range of conformations. The authors propose an ensemble-conditioned method that designs molecules against this distribution of shapes rather than a single structure, aiming to better capture the properties that determine whether a candidate succeeds.
A new arXiv preprint describes a "biology-in-the-loop" framework that chooses which perturbations to test next when experimental budgets are limited. The approach amortizes the cost of adaptive decision-making so that sequential selection can be applied efficiently to CRISPR screens. The authors frame the problem as sequential experimental design for biological discovery under constrained resources.
A revised arXiv preprint proposes a method for learning intrinsic water-quality dynamics that incorporates rainfall as an environmental driver. Rainfall affects water quality through runoff, pollutant transport, dilution and resuspension, processes that mechanistic models describe explicitly but which are hard to capture in purely data-driven approaches. The work aims to combine these perspectives for improved forecasting.
A new arXiv paper addresses the difficulty of scoring open-ended, tool-using LLM agents in chemistry and drug discovery, where conventional benchmarks fall short. The authors propose an evaluation system built on the LLM-as-a-Judge approach and tune it against human expert judgments to improve reliability. The work aims to make automated assessment of agentic scientific workflows more trustworthy.
A revised arXiv preprint introduces a method that uses automatic verification to track how well language models reason about unsolved mathematical problems. The author notes that although large language models now handle sophisticated math and science reasoning, whether they can contribute genuinely new research remains contested and thinly studied. The work aims to give a measurable way to assess progress on that question.
A new study presents an active learning framework designed to connect laboratory-scale results with real decisions about which critical-materials recovery processes to scale up. The method accounts for product requirements, process costs, and scale effects, drawing on records from Pacific Northwest National Laboratory's critical-element recovery database.
Researchers have introduced FrontierChallenge, a benchmark of 300 tasks spanning multiple scientific disciplines that measures whether AI agents can carry out complete research workflows. It goes beyond existing evaluations that score only final answers, standalone programs, or work within a single field, instead assessing capabilities like data processing, coding, and producing research artifacts.
A new machine learning study argues that protein-protein interaction databases carry study and technical biases that skew protein and interaction attributes, letting models succeed by exploiting shortcuts rather than genuine biological signals. The authors propose methods to audit these datasets for such biases and to mitigate them, aiming for models that learn real biology instead of dataset artifacts.
A new arXiv preprint uses Kolmogorov-Arnold Networks, an interpretable neural architecture, to improve theoretical models that predict the masses of atomic nuclei. The authors address the difficulty of learning from limited and highly complex nuclear datasets, aiming to blend physics-based theory with data-driven corrections.
Researchers report a self-evolving scientific agent that designs control strategies for dynamic fluid systems without relying on neural networks. The resulting white-box controllers are explicit and physically reasoned, making autonomous control decisions easier to interpret and diagnose.
A new research paper explores applying reinforcement learning to searches for new physics in particle physics, focusing on anomalies where low-energy measurements deviate from Standard Model predictions. The work targets one of the field's most important open problems: identifying evidence of physics beyond the Standard Model.
Google DeepMind has published a dataset that predicts the effects of roughly nine billion single-letter variations across the human genome. The resource is about one petabyte in size, roughly thirty times larger than the AlphaFold database. It was used to help identify a previously unexplained genetic cause in an epilepsy case.
Google DeepMind has published a dataset called the AlphaGenome Atlas that predicts the possible consequences of roughly nine billion single-letter variations in the human genome. The collection is about one petabyte in size, which the outlet notes is more than 30 times larger than the AlphaFold database. A case involving epilepsy is cited as an example of how the resource was used.
Google DeepMind has published AlphaGenome Atlas, a precomputed resource that scores the predicted effects of roughly nine billion genetic substitutions. The aim is to help researchers decide which variants to prioritize for experimental testing by tying predictions to biological mechanisms, with the DNM1 gene used as an illustrative case.