Publisher
research.google
Newest story built on research.google reporting: Google's Geospatial Reasoning agent starts every query with a planning stage (October 14, 2026). research.google reporting informs the Build and Invest desks, most often on AI Agents, Foundation models beyond language, and Geospatial AI. 18 of our current stories draw on it, the earliest from August 16, 2026.
Google Research is widening experimental access to its Geospatial Reasoning agent, which its report tests in four configurations on two Gemini models. Access runs through Ask Google Earth and an A2A service.
Reality
Google released Deep Review publicly on October 14, extending an agentic verifier tested on more than 16,000 manuscripts to Google Docs and presentations. Google says AI-assisted drafting has outrun human capacity to verify.
Reality
David Autor and Tanya Rodchenko's trial of 133 patent attorneys found AI lifted drafting quality for everyone but built stronger judgment only in seniors. Juniors' judgment scores split toward both ends with no average gain, so better drafts no longer show a firm which juniors are learning.
Reality
Google's CAPS workshop report, from more than 50 researchers, proposes a policy engine that checks agent actions against context norms before they run. For teams shipping agents, privacy enforcement would move from manual permission grants to rules a language model generates for each task, a design the authors still present as open research.
Reality
Google Research now trains federated models inside attested TEEs and logs each access policy to the public Rekor log, with Gboard already switched over. The server-side noise step that users once had to take on trust becomes code an outside auditor can check against open source.
Reality
Google Research says the fully fine-tuned version of its Diffusion Controller posted a 90% win rate over the base image model. The lighter add-on meant for closed models is reported only as beating the industry standard for human preference.
Reality
Google Research has published a C++ instance generator for the middle-mile leg, where one shipment changes vehicles several times and can take a week to arrive. Single-day vehicle routing benchmarks do not describe that.
Reality
Google Research describes a model that assembles the interface as the request arrives, held to a game structure of progressively harder challenges, with more than 30 teacher-reviewed samples out for middle and high school STEM.
Reality
Retrieve-for-Train runs reinforcement learning once against a fixed corpus and distils the winning sub-query sets into a lightweight diffusion model. Adopting it puts a retrain schedule on whoever owns the catalogue.
Reality
A Google XR team fine-tuned Gemma-3 models on data generated answer-first and says they match proprietary LLMs on tools they never trained on. The post's cost and pass-rate claims carry no numbers.
Reality
Google's new forecaster is 330m parameters trained on a trillion time points. In zero-shot multivariate mode it anticipates a 20% promotion lift that a univariate model misses. That is most of what an in-house demand model does.
Reality
Google's new 330 million parameter forecaster takes covariates and related series zero-shot in one forward pass, which puts a lot of bespoke per-SKU plumbing up for review, assuming the accuracy transfers to your own series.
Reality
Google Research argues the weeks in geospatial modelling go into curation rather than modelling, and that an agent can absorb them. The engineering worth copying is that the orchestrating model plans over pointers and never reads the data.
Reality
The dual-stream design is the part of this that would transfer to other sensor data. Four things changed at once, and no ablation appears in the write-up.
Reality
Google Research's ME-POIs treats a place's visit rhythm as an input feature rather than a prediction target. What it measures is not quite what the summary advertises.
Reality
The Biomarker Discovery Framework wires statistical checks, adversarial validation and human review into the loop, and a top depression correlate still explains about 6 per cent of rank variance.
Reality
Google Research reports Gemini-3-Pro and GPT-5 encode 95-98% of tested facts yet fail to recall 26-34% of them, moving the fix from pretraining scale toward post-training and inference.
Reality
A controlled evaluation across four benchmarks found centralized coordination lifted financial reasoning 80.9%, while every multi-agent variant tested made strict sequential planning worse.
Reality