Claims Brief · Lead record
The published application recites an encoder, a language model and a module between them. What the dependent claims name — and what the title does not limit — is where the scope actually sits.
Claims Brief
US20260205268A1, published July 16, 2026, recites embedding an authentication key inside a generative model's own output, with output quality higher when the key is valid than when it is not. It is classified under cryptography, not machine learning.
Claims Brief
US20260203368A1, published July 16, decomposes an n-bit weight matrix into one-bit matrices and retrieves precomputed partial products from a lookup table indexed by the weights themselves. Three independent claims, one of them a bit-serial hardware system.
Claims Brief
A pending application published July 2, 2026 claims a computing system that resolves an AI agent's natural-language 'semantic entitlement' through a generative model to grant resource access. Here is what the independent claim actually covers — and the fact that it is published, not granted.
Claims Brief
A granted patent — issued, not merely published — covers a speech vocoder trained by mapping audio to Gaussian values through invertible layers, then run in reverse to synthesize voice. Here is what the independent claim requires, and where it lands in the G10L/G06N landscape.
Claims Brief
A patent application published June 25, 2026 is directed to prompting a vision language model to evaluate whether an image contains imperfections and to score its quality. It is a pending application, not a granted patent, classified under CPC G06V 10/82.
Claims Brief
A patent granted June 23, 2026 and assigned to Google LLC claims a federated distillation method that compresses a large server-side speech recognition encoder into a smaller on-device one using principal component analysis. This is an issued patent, not a pending application, and it lands in a recent Google cluster pointed at on-device, privacy-preserving model training.
Landscape Report
A landscape read of who holds attention- and transformer-related neural-network patents, grounded in real granted records across Microsoft, Google, Qualcomm, and others — and a caution about what patent counts do and do not show.
Explainer
Beyond eligibility, AI claims must satisfy 35 U.S.C. 112 — written description, enablement, definiteness, and the means-plus-function rule. An explainer on why machine-learning claims are drafted the way they are.
Explainer
Under US law a patent inventor must be a natural person, so an AI system cannot be named — but AI-assisted inventions are not categorically unpatentable. An explainer on Thaler v. Vidal and the USPTO's 2024 inventorship guidance.
Primer
A reference primer on the Cooperative Patent Classification scheme for artificial intelligence: what G06N covers, how its subgroups divide neural networks, learning methods, and machine learning, and where AI spills into G06F and G06V.
Explainer
An explainer on why so many machine-learning inventions face a 35 U.S.C. 101 challenge, and how the USPTO's two-step Alice/Mayo framework actually evaluates them under MPEP 2106.
Research Brief
An arXiv preprint proposes an estimator for evaluating offline reinforcement learning policies when logged rewards are systematically missing, and tests it on simulated data and MIMIC-III sepsis records.
Research Brief
A new arXiv preprint introduces synthetic datasets built on critical mean-field percolation clusters, arguing that today's toy datasets lack the hierarchical, multi-scale structure that real data has.
Claims Brief
A freshly published application is directed to deciding which experts a downstream transformer block should run, using a perplexity measurement taken mid-network. It is a pending application, not a granted patent.
Research Brief
A new arXiv paper learns task-level knowledge on a simplified agent and transfers it to physically different robots through a semantic-magnitude interface, reporting large reductions in tracking error and a fraction of the usual interaction data.
Research Brief
A new arXiv paper proposes a zero-shot object-goal navigation framework that mines actionable rules from past trajectories, retrieves them with an upper-confidence-bound strategy, and forecasts outcomes before acting — reporting a 10.1% gain in success rate.
Research Brief
A new arXiv paper shows that ternary state space models can be made from a pretrained checkpoint with quantization-aware training and distillation, instead of training from scratch on 150 billion tokens — and surfaces a new instability called zero-ratio collapse.
Research Brief
A new arXiv paper recasts diffusion-policy training as a deterministic boundary-value problem in a Cameron-Martin space, reporting convergence guarantees and a reward-free failure detector, with gains on a manipulation benchmark and a manufacturing-line task.
Research Brief
FPRM, a looped transformer described in a new arXiv paper, halts when its latent state reaches a fixed point — letting the model spend more compute on hard reasoning problems and less on easy ones, with results reported on Sudoku, Maze, state-tracking, and ARC-AGI.
Research Brief
A new arXiv paper proposes LoopWM, a world-model architecture that iterates a single parameter-shared transformer block to refine latent states, treating loop depth as a new scaling axis distinct from model size.
Grant Watch
A G06N 3/084 grant issued June 16 gives Autodesk a method for training machine learning models — built around a graph neural network — to perform tasks on B-rep CAD objects represented as graphs plus 2D UV-grids.
Grant Watch
A grant that issued June 16 hands NVIDIA a method for teaching a deep neural network to read 3D road surfaces using parametrically generated synthetic training data — sidestepping the cost of collecting and labeling real-world drives.
Research Brief
A new arXiv framework uses LLMs to abstract low-level click-and-action logs into high-level activities, reconstructing tasks from browser logs and predicting student dropout from sparse examples.
