HiRO-ACEv2, Deep Review goes public, Earth AI agents

Today's releases share a theme: how you check an agent's work counts for as much as the agent. I only used pages I opened.

Ai2 ships HiRO-ACEv2 for 3 km climate data

Ai2 released the second version of HiRO-ACE, a two-part system. ACE is an AI model that runs at a coarse 100 km resolution, and HiRO sharpens its output to 3 km. The new piece is HiROv2, which predicts several fields at once instead of only rainfall, and which Ai2 trained to do better on extreme events.

The numbers are specific. Given coarsened data from X-SHiELD, a high-resolution physics model, HiROv2 matches the target statistics to within 3%, even for events that occur once in a million data points. It estimates the total energy of tropical cyclones, including storms it never saw in training. Ai2 generated a full year of global 3 km data, and a 10-year run produced hurricane energy comparable to X-SHiELD. The full system costs about 1/300th of running X-SHiELD, and more savings apply if you only need one region.

Ai2 is plain about the gaps. Hurricane winds and pressure around the eye are somewhat lopsided compared with X-SHiELD. Because HiRO learned from X-SHiELD, it only works with the ACE version tuned to match it, so it cannot yet serve a warmer climate or the historical record. A warmer-world dataset is the next planned step. The models, tutorial notebooks and a web demo are available.

Google makes Deep Review public

Deep Review is an agentic system that checks long technical documents. It started as the Paper Assistant Tool, which gave opt-in authors pre-submission feedback at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, covering more than 16,000 manuscripts. It now accepts PDFs and plain text, adds documents such as legal and finance material, and gives each user two free reviews a week, with a Gemini API key route for higher volume and an API. Uploaded documents are not used to train Google models.

The design is a four-stage pipeline: segment the document, budget compute by how dense each section is, verify segments in parallel with full-document context, then synthesize and run grounding checks with tools. That budgeting step is the part I find most practical.

Google's geospatial agent gets a technical report

Google Research described a Geospatial Reasoning agent that connects Gemini to its Earth AI models, including WeatherNext, AlphaEarth Foundations and remote sensing models, plus Maps and Earth Engine data. It plans each query up front, works in a separate workspace so map layers stay out of the context window, and shows its plan and data sources. Users can bring CSVs, GeoJSON or cloud tables.

The report compares four configurations across four benchmarks and releases the development splits of two new ones, GeoExpertQueries and GeoKTools. Access is expanding through Ask Google Earth and an agent-to-agent API, both experimental.

NVIDIA NeMo Gym and a chip-design agent

Siemens used NeMo Gym, NVIDIA's open library for evaluating agents in environments, on its Fuse EDA AI Agent for standard-cell characterization. A verifier checks the final Liberty files and gives zero reward if any gate fails. After revising the agent's skills and MCP tool descriptions based on the failures, the composite score went from about 55% to about 95%, agent turns dropped from 155 to 53, and output tokens fell from about 81K to 29K. The post notes that final-artifact checks caught errors that intermediate checks missed.

What I'd watch next

Whether Ai2 delivers the warmer-climate version, how the Earth AI benchmarks get used by outside teams, and whether other agent builders adopt final-artifact verifiers the way Siemens did.