What happened in AI today

Google Research says it is connecting AI agents to physical lab robots so they can design molecules, run experiments, and learn from the results. Its VERID framework is a live digital twin of the lab workcell that tracks well volumes, pipette tips, and robot arm position, and it simulates protocols before the robot touches reagents. In aptamer tests over three design rounds, the agents raised confirmed on-target hit rates to 33.5% for myoglobin and 42.1% for α-thrombin, the latter at 98.5% specificity. The team also reports that its de novo design pipeline produced binders with single-digit nanomolar affinity for fibroblast activation protein alpha, a cancer-linked target.

• Grounding AI agents in biological reality: Building blocks for autonomous wet labs

Google DeepMind researchers describe how biologists used Google Antigravity, the company's agentic development platform, to search UniProt for new antimicrobial peptides. The agent queried databases, checked novelty, compared AlphaFold structures and read the genomic context around a candidate bacteriocin from Streptococcus pseudopneumoniae. Google says the candidate has predicted activity against pathogens involved in respiratory tract infections and is now in wet-lab validation, along with several others found through the same pipeline. Google also introduced a Biomolecular Structure Viewer, a 3D tool inside Antigravity that the researcher and agent can both use and annotate.

• Accelerating the search for antimicrobial peptides with Google Antigravity

Microsoft Research has described Magnetar, a zero-knowledge virtual machine that proves programs written in Rust, Go or C. A zero-knowledge proof lets anyone check in milliseconds that a program ran correctly, without rerunning it or seeing its data. Microsoft says Magnetar is 10 to 39 times faster on average than the fastest existing zkVMs across 15 benchmarks, and its proofs verify in about 12 milliseconds. Its security relies only on standard hash functions such as SHA-256 or SHA-3, with no trusted setup.

• Proof-based general-purpose computing with Magnetar

Google Research published a technical report on a Geospatial Reasoning agent that connects Gemini to Earth AI models such as WeatherNext, Flood Forecasting and AlphaEarth Foundations. The report evaluates the agent on four benchmarks, and Google is releasing two new ones, GeoKTools and GeoExpertQueries. Google says the WHO Regional Office for Africa used the agent early in the 2026 Ebola outbreak in the Democratic Republic of the Congo to find 48 exposed settlements and more than 45,500 at-risk people in minutes, work that normally takes weeks. Google is also offering early access to a deep research update in Ask Google Earth and to an experimental agent-to-agent API for developers.

• Putting Earth AI to work: Geospatial agents for our planet's biggest challenges

Meta built an internal coding scaffold powered by Muse Spark 1.3 and Muse Wonder. It could access public data and a licensed database of five million de-identified electronic health records. Medical researchers in cardiology, epidemiology, infectious disease and internal medicine used it over four weeks to brainstorm hypotheses, write and run code, and draft manuscripts. Meta says domain experts stayed responsible for scientific decisions and conclusions, and it is sharing ten manuscripts that cover topics such as infectious disease, sports medicine, oncology and endocrinology.

• Accelerating Scientific Research in Health Data Science With Muse Models

Ai2 has released HiRO-ACEv2, an update to its High Resolution Output Ai2 Climate Emulator. The system pairs ACE, an AI model at 100 km resolution, with HiRO, a model that sharpens those outputs to 3 km. Ai2 says the full system generates global 3 km climate data at about 1/300th the cost of X-SHiELD, a high-resolution global atmospheric model. The new HiROv2 predicts several fields at once, and Ai2 says it matches X-SHiELD's high-resolution statistics to within 3% when given coarsened X-SHiELD data, even for events that occur once in a million data points. Ai2 notes that winds and pressure around hurricane eyes are still somewhat more lopsided than in X-SHiELD, and it has built a web demo that compares HiROv1, HiROv2 and X-SHiELD output.

• HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost
• HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

Microsoft Research has described StudentSim, a research system that trains a personalized AI simulator for each student from public, de-identified learner records. Researchers can use the simulators to get feedback on AI tutors without running slow, costly human studies. Microsoft says StudentSim beat leading baselines in chess, second-language English writing and math. In chess, it scored 0.5150 on matching a student's responses, against 0.2316 for GPT-5.4. A chess tutor trained with StudentSim as the reward source reached 90.5% accuracy in blind expert review, against 75.7% with no reinforcement learning and 71.6% with GPT-5.4 as the reward. Microsoft says StudentSim is not a product, is not deployed in classrooms and does not replace real learner studies. The paper and code are public.

• Simulating student learning for better AI tutoring

Researchers at Google and Beth Israel Deaconess Medical Center published a study in The Lancet on AMIE, Google's research diagnostic AI chatbot. In a real-world study at BIDMC's ambulatory primary care clinic, 98 patients talked with AMIE before urgent care visits while supervising physicians watched in real time. Google says no conversation needed to be interrupted under predefined safety criteria, and AMIE's differential diagnoses matched the doctors' final diagnoses 90% of the time. Clinicians said the AI summaries helped them prepare for visits in 75% of cases and changed their approach to care in more than half. Google says larger clinical trials are needed.

• Study in The Lancet suggests AI could improve patient-physician relationships.

Google Research and NBER published a three-month randomized trial with 133 patent lawyers at eleven firms. Lawyers given an unreleased Google Labs AI drafting tool scored 0.38 standard deviations higher on a drafting task at 90 days. On a redlining task done without AI, the gain came entirely from senior lawyers with 7 or more years of experience, who beat controls by 0.45 standard deviations. Junior lawyers showed no average improvement, and their scores split into more very low scores and more good ones. The authors note the sample is small and the study ran only three months.

• Does better work always mean better workers?

Nvidia says its teams used supervised fine-tuning, reinforcement learning and a generate-and-refine inference loop to build Nemotron 3 specialists that reached gold-medal level at both the 2026 International Olympiad in Informatics and the International Mathematical Olympiad. At IOI, Nemotron-3-Ultra-CC scored 535.4 out of 600, above the 361.12 gold threshold and the top human score of 498.27. That run was unofficial and unsupervised, was held under the same constraints as human contestants, and was not part of the official ranking. At IMO, a system built on Nemotron 3 Ultra scored 30 out of 42 against a gold threshold of 29, and official IMO graders marked its proofs.

• One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

Ai2 says the research behind Bolmo, its family of fully open byte-level language models introduced last December, has been published in Nature. The paper describes "byteifying," a process that converts an existing subword model into a byte-level one with a relatively short additional training run. Ai2 applied it to Qwen 3 8B and Llama 3 8B to make Bwen 8B and BlamaLlama-B 8B, which it says come close to the original models. Ai2 says Bwen 8B outperforms Bolmo 7B across its aggregate evaluation suite, and it has released new checkpoints and Stage 1 checkpoints on Hugging Face.

• Now in Nature: Retrofitting language models to operate over bytes

OpenAI is releasing a broad set of new mathematical results produced by an internal frontier model. It is publishing them in a GitHub repository with protocols for paper revisions and citations, and says it consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to share them. The repository includes Lean formalizations of many proofs, which a computer can check, and OpenAI says it will add more. It also includes summaries of the model's reasoning and statistics on attempted problems. OpenAI says the average result used the equivalent of roughly three hours of ChatGPT Pro thinking, and that it is working to release the model responsibly.

• Sharing AI progress in mathematics

Google researchers, working with Northwestern Medicine and Jacaranda Health, studied 1,000 mothers in Nairobi, Kenya, and 1,000 in Chicago. Healthcare workers with no ultrasound experience learned a "blind sweep" method in eight hours, moving the probe over the abdomen in a set pattern. Google says its machine learning model estimated gestational age and fetal presentation as accurately as a trained sonographer. The model runs on the device, so it needs no electricity supply or Wi-Fi, and it can tell the operator when to redo a sweep.

• Ask a Scientist: How are researchers using AI to help pregnant women access ultrasounds?

Google published research with health partners on using its Earth AI models, including the Population Dynamics Foundation Model (PDFM), to forecast disease and find at-risk communities. In the ongoing Ebola outbreak in the Democratic Republic of Congo, the WHO regional office for Africa used a prototype Geospatial Reasoning agent to find 48 exposed settlements and more than 45,500 at-risk people in minutes. Google says this normally takes weeks. Other case studies covered dengue in Mexico, cholera in the DRC, cardiovascular deaths in U.S. counties and MMR vaccination near the U.S.-Canada border. In the cholera work, adding PDFM improved identification of outbreak-prone health zones up to eight weeks ahead, Google says. PDFM embeddings are available in Preview as Population Dynamics Insights from Google Maps Platform, and researchers can request no-cost access for select, non-operational research uses.

• Making global public health more proactive with Google Earth AI
• Unlocking Earth AI’s planetary geospatial foundation models for global public health

GitHub launched ReviewBench, an offline benchmark for code review agents. It has 219 pull requests from 187 public open source repositories in 19 languages, modeled on GitHub's analysis of 103.9 million pull requests. Ground truth comes from human reviewers, author follow-up commits, analysis tools and several LLMs, and Claude Sonnet 5 grades submissions. GitHub says senior engineers who did not build the dataset agreed with its true and false-positive labels 96.6% of the time. The research preview is on the ReviewBench website, where users can view the dataset and test their own agents.

• ReviewBench: An open benchmark for AI code review

Microsoft has released ThinkingBox, a benchmark that grades AI agents on the backend records and side effects they leave behind, not on the text they generate. ThinkingBox-Bench covers 507 stateful business workflows, and each task runs 20 times against various models. Microsoft says Claude Opus 5.5 leads at 67.16%, and Kimi-K3 is the strongest open-weights model, within a point of GPT-6-Astra. Kimi-K3 passes only 68 of 507 tasks (13.41%) in all 20 attempts, while Claude Opus 5.5 and Claude Opus 5 each pass 241 of 507 (47.53%). The harness and dataset are on Hugging Face.

• The Agent Said It Was Done. The Database Disagreed.

Meta says mathematicians worked with Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, with no custom research scaffold. The six papers cover probability, differential equations, group theory, optimization, arithmetic physics, and non-associative algebra. Meta says five answer previously open questions, including a counterexample in a group of 384 elements. Each paper was reviewed by a second group of mathematicians and marks which passages AI drafted. Meta also says other teams independently announced solutions to some of the same problems, and the papers acknowledge that work.

• Solving Open Research Problems Together

Google says a flu-forecasting model built with its AI ranked first among 39 models evaluated by the CDC for the 2025-26 season. The CDC's FluSight program collects weekly forecasts of U.S. flu hospital admissions for the current week and three weeks ahead to anticipate demand for care. Google says its forecasts were built using Empirical Research Assistance, an AI tool that creates optimization algorithms for science.

• Google's AI ranks #1 for predicting flu hospitalizations.

Google DeepMind says SynthID Bio embeds a detectable signature into AI-designed protein sequences and predicted 3D structures while preserving biological function. In wet-lab tests across three targets, including VEGF-A and the SARS-CoV-2 spike protein, DeepMind says watermarked binders matched unwatermarked versions on hit rate and binding strength. DeepMind says it is publishing the methods paper, open-sourcing code and lab data, and testing genome watermarking with Stanford and Arc Institute.

• Introducing SynthID Bio

Microsoft Research describes a machine learning pipeline that forecasts space-weather risk for 66,935 substations in the continental US 30 to 60 minutes ahead. It says the system combines solar-wind measurements, forecasts of geomagnetic activity, local geology and grid data to estimate where induced currents could threaten equipment. The team reports detection rates of about 76.5% for major events and 81.2% for severe events, and says further validation with utilities would be needed before operational use.

• Forecasting space weather risks on power grids

Hugging Face says it launched the Open TTS Leaderboard to standardize evaluation as the Hub now hosts more than 8,000 text-to-speech models. The board uses objective metrics instead of human votes, including intelligibility via word and character error rate, speed via batched inference and time-to-first-audio, and speaker similarity for voice cloning. It ranks multilingual performance and streaming latency, with early English leaders including Kokoro-82M, Supertonic-3 and S2-Pro, plus a Listen tab to compare generated audio.

• Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

Microsoft says Quine combines a multimodal world model of biology spanning sequence, structure, cell state and imaging with a harness linking tools, literature and researchers. In work with the Broad Institute on pancreatic cancer, Microsoft says Quine prioritized thousands of compounds for cell-state shifts in a weekend and its top-ranked picks produced the largest intended shifts in wet-lab assays. Microsoft is opening a Quine Fellows program for outside scientists, with access initially limited to fellows and select collaborators.

• Introducing Quine: An AI research system designed for the complexity of biology

Google researchers describe an orchestration layer on top of Gemini and Veo that plans visual continuity instead of chaining prompts, with built-in SynthID watermarking. The work combines a co-director for story strategy, CANVAS for character and scene memory, A2RD for minute-long generation, and VQQA for automatic prompt fixes, including a ten-minute demo. Google reports gains on continuity benchmarks and a peak quality score of 81.4 on its GenAD-Bench test.

• Automating coherent long-form video generation

Microsoft researchers say running robotics AI only on onboard GPUs can limit performance, battery life and model size. In mobile manipulation tests, they found mapping and planning slowed by up to 383% versus a datacenter A100, timely obstacle detection fell 30% with lighter GPUs, and manipulation accuracy fell by half, while large onboard GPUs drained batteries by up to 160%. Microsoft says it built Kubernetes-based tooling to containerize and offload robot inference to edge or cloud GPUs, now added to its open-source Physical AI Toolchain.

• Offloaded inference for real-world physical AI robotics

Nvidia says it joined Google DeepMind, EMBL-EBI and others to release predicted 3D structures for protein complexes from more than 2,800 viruses through the AlphaFold Database. The team used DeepMind's AlphaFold2 with Nvidia's BioNeMo Inference Runtime to predict interacting protein groups at scale, and Nvidia also openly released the structure-prediction pipeline. About 30 percent of the interactions are new to science, according to Nvidia.

• How Open Science Can Help Researchers Prepare for the Next Pandemic

OpenAI introduced MentalHealthBench, an open benchmark co-created with more than 80 licensed mental health experts from 22 countries. It uses synthetic conversations across everyday stress, serious distress and emergencies to test safety, context-seeking, agency and guidance. OpenAI says it is releasing the methods openly so others can run evaluations and build on the work.

• Introducing MentalHealthBench

Anthropic says agents using Claude searched DNA databases and discovered a previously uncharacterized system it calls array-associated reverse transcriptases, or ARTs, in bacteriophages. After about 21 hours of work by roughly 950 agents, Claude spotted a repeat array next to a reverse transcriptase gene that Anthropic says no one had noticed. Lab tests show the array is expressed as distinct short RNAs, though its function is still unknown, and Anthropic released a pre-print for further study.

• Claude discovers a novel enzyme system with CRISPR-like repeats