USPTO Patent Applications - AI & Computing (G06N)

Quantum Computation Method and Information Processing Apparatus

USPTO published patent application US20260094038A1 by Fujitsu Limited covering a quantum computation method using Molmer-Sorensen (MS) gates. The invention generates multiple equivalent quantum circuits by converting two-qubit gates using MS gates with varying parameter values, executes computations across these circuits, and outputs averaged results. Application 19412895 was filed December 9, 2025.

Methods and Systems for High-Fidelity ZZ-Rotation of Qubits

The USPTO published patent application US20260094037A1 for a method and system performing high-fidelity ZZ-rotation quantum gates on two qubits. The invention involves applying rotation pulses and echo pulses with specific phases and amplitudes to control detuning and unwanted rotations. The application was filed September 29, 2025, and names Ilya Gurwich, Itsik Cohen, Ron Melcer, and Shlomi Kotler as inventors.

Hybrid Classical-Quantum Computer for Chemical System Simulation

Quantinuum GmbH filed USPTO patent application US20260094036A1 for a hybrid computing system combining classical and quantum processors to simulate chemical systems. The invention uses pre-entangler and variational circuit algorithms to generate quantum ansatz values for electron orbital simulations. Application filed September 13, 2023; published April 2, 2026.

Quantum Kernel Data Generation and Classification System

USPTO published patent application US20260094035A1 for a quantum kernel-based data generation and classification system. Inventors include Shungo Miyabe, Noriaki Shimada, Sudeep Ghosh, Jae-Eun Park, and Abhijit Mitra. The system employs quantum kernels to generate new datasets from input data and classify information, with potential applications in quantum computing and machine learning.

Method for Accelerating LLM Inference Procedures

USPTO published patent application US20260094028A1 by MEDIATEK INC. disclosing a method for accelerating large language model inference through draft token generation, rule-based determination, and matching operations. The invention aims to improve computational efficiency of LLM inference procedures using a two-stage drafting and matching approach. The application was filed September 26, 2025.

Neural Network Device with Non-Linear Operations

Samsung Electronics Co., Ltd. filed US Patent Application US20260094027A1 with the USPTO for an electronic device executing neural network models with non-linear operations. The invention covers updating weights between neural network nodes based on reference node values, with inventors including Chang-Woo Shin, Carlos Cristiano De Jesus Alcantara, Anes Ju, and Kitae Park. The application was published on April 2, 2026.

Selective Interaction with a Portion of Content by a Generative Response Engine

OpenAI OpCo, LLC filed patent application US20260094016A1 for a generative response engine capable of dynamically determining and displaying generated content in appropriate user interface formats. The system enables content display in content frames separate from traditional conversational interfaces, with intelligent assessment of content type to determine optimal display modes. The application was published on April 2, 2026, naming 13 inventors including Bryan Ashley, Lee Byron, and others.

Knowledge Graph Generation Method for AI Systems

Fujitsu Limited has filed a patent application (US20260094018A1) with the USPTO for a knowledge graph generation method enabling AI systems to process multiple input documents through cause-result relationships. The application was published on April 2, 2026, with inventors including Tatsuru Matsuo.

RBAC System Patent for Generative AI Development

USPTO published Truist Bank's patent application for a role-based access control system designed for generative AI development. The system authenticates user credentials, filters permissions based on user access levels, and regulates prompts and user inputs to a knowledge domain framework used in AI model creation. Application US20260093839A1 was filed April 9, 2025 and published April 2, 2026.

Machine Learning Model Pruning System

The USPTO published patent application US20260094048A1 for a machine learning model pruning system developed by Amazon Technologies, Inc. The invention describes methods for optimizing neural network weights by strategically pruning connections and minimizing loss functions through batch processing techniques. This is a publication of a patent application filed on September 27, 2024.

Transfer Learning AI/ML Patent for Beam Management in Telecommunications

VIAVI Solutions Inc. filed a patent application (US20260094001A1) for transfer learning systems enabling AI/ML-based beam management in telecommunications networks. The invention allows pre-trained neural network models to be applied across different frequency bands to predict optimal beam configurations, reducing training requirements for 5G and future wireless systems. Application No. 18904464 was filed October 2, 2024.

AI System for Dynamically Generating Knowledge Assessment Items

EXAMROOM.AI CORP. filed USPTO patent application US20260094017A1 for an AI and machine learning system that dynamically generates knowledge assessment items for item banks. The patent covers methods, systems, and computer program products for automated test item creation. The application was filed September 1, 2025 and published April 2, 2026.

Federated Learning Acceleration Patent for Intelligent Personalized Services

USPTO published patent application US20260094005A1 by Korea Electronics Technology Institute for a client training acceleration method using federated learning for intelligent personalized services. The invention includes searching similar models in a repository, aggregating them into a global model, distributing to clients for training, and reflecting local data on the model. The application (No. 18941753) was filed on 2024-11-08 and published on 2026-04-02.

Automated Data Visualization and Infographics Using LLMs and Diffusion Models

Microsoft Technology Licensing filed USPTO Patent Application US20260094325A1 for automated generation of data visualizations and infographics using large language models and diffusion models. The system generates candidate analytics from raw data, creates visualization code scaffolds, and produces programmatic outputs including infographics via diffusion models. Application No. 19412362 was filed December 8, 2025.

AI Method for Determining Object Physical State Using PDEs

USPTO published patent application US20260093980A1 describing a computer-implemented method for generating training data using partial differential equations (PDEs) to determine the physical state of objects. The application, filed on September 28, 2025, covers methods for configuring PDE boundary conditions on basic shapes to derive solutions representing physical states. Inventors include Kaixuan Zhang, Jianing Huang, Youjia Wu, and Ze Cheng. This is a routine publication of a patent application under CPC classification G06N 3/08.

Methods for Training Foundation Models for Processing Optical Physiological Signals

Nokia Solutions and Networks Oy filed USPTO Patent Application US20260093979A1 for methods of training artificial intelligence models using Photoplethysmography (PPG) physiological signals. The patent covers AI models incorporating multiple machine learning models trained on PPG data obtained from multiple individuals to determine outputs based on the physiological signals.

Model Training Method, Image Editing Method, Apparatus, Device, Medium, and Product

The USPTO published patent application US20260093978A1 by inventors Gu, Zhu, Chen, and Wen, disclosing methods for training image editing models using original images, editing descriptions, edited images, and multi-item evaluation information. The application covers processing these inputs to generate image editing results and updating the model based on differences between results and edited images.

Machine Learning Systems and Methods for Real Time Anomaly Detection and Prescriptive Feedback

USPTO published patent application US20260093977A1 by Zebra Technologies Corporation disclosing a machine learning method for real-time anomaly detection with prescriptive feedback. The method involves receiving query parameters, retrieving datasets, applying machine learning models trained in real-time to detect anomalies, filtering outputs, generating identification instructions, and transmitting anomalous item information. Application No. 18943495 was filed November 11, 2024.

Fault-Aware Training to Salvage AI Accelerators

The USPTO published patent application US20260093976A1 by inventor Ihab Amer describing a method for fault-aware training of AI accelerators. The method generates multiple neural network model approximations of a compute engine with faults, matching a fault map to an appropriate approximation, and loading the matched model when the engine transitions to approximate mode. The compute engine is a multiply-accumulate unit within an AI accelerator.

Personalized Intent Automation Using Dynamic Prompt Generation for Machine Learning

USPTO published patent application US20260093975A1 titled "Personalized Intent Automation Using Dynamic Prompt Generation for Machine Learning." The invention covers techniques for detecting trigger events in software applications, aggregating user data from multiple sources, generating prompts for ML models based on recommendable actions and associated scores, and providing recommended actions to users with feedback mechanisms for updating scores. Inventors: Rachita Ramesh, Gaurav Budjade, Neha Jayaprakash, Nava Teja Tummalapalli, and Sujay Sundaram.

Neuromorphic photonics with coherent linear optical neurons

The USPTO published patent application US20260093974A1 by Pleros et al. covering neuromorphic photonic computing systems using coherent linear optical neurons implemented with multipath optical interferometers. The application includes optical amplitude modulators, phase shifters, and photodetectors for non-linear activation functions in optical neural networks. The patent was filed on July 7, 2025, under application number 19261861.

Analog Neuromorphic Circuits for Dot-Product Operations Implementing Resistive Memories

The USPTO published patent application US20260093973A1 for analog neuromorphic circuits implementing resistive memories for dot-product operations. Inventors Chris Yakopcic, Tarek M. Taha, and Md Raqibul Hasan disclose a circuit architecture where input voltages representing vector values are processed through paired resistive memories, with resistance values converted to weighted matrix values for neural network computations. The application was filed December 9, 2025, and published April 2, 2026.

Layer-wise Precision Optimization in Analog Compute-in-Memory Accelerators

USPTO published Intel Corporation's patent application US20260093972A1 for a layer-wise precision optimization method in analog compute-in-memory (ACiM) accelerators for neural networks. The invention selectively allocates neural network layers to either digital compute-in-memory (DCiM) or analog compute-in-memory (ACiM) circuits based on signal sensitivity and statistical weight distribution criteria.

Accenture patent enhances LLM performance via data curation

Accenture patent enhances LLM performance via data curation

Efficient Decoding of Output Sequences Using Parameter Sharing

USPTO published patent application US20260093982A1 for methods and systems enabling efficient decoding of output sequences using parameter sharing in machine learning tasks. The application, filed by six inventors including Adam Joshua Fisch and Tal Schuster, covers techniques for generating output sequences by processing embeddings through layer blocks until termination criteria are met.

Embedded Neural Network System and Information Processing Method

USPTO published patent application US20260093983A1 for an embedded neural network system. The application, filed November 12, 2025 (No. 19386261), covers an information processing system with an embedded information acquisition unit and a generation unit that embeds information into neural network elements. Inventors: Yuko ISHIWAKA, Kazuto SUDA, Minoru OWADA.

ML Model Characterizes Solid-State Detectors and Reduces Defect Levels

USPTO published patent application US20260093984A1 by Srutarshi Banerjee et al. describing a physics-based neural network model that learns trapping, detrapping, and charge transport properties in solid-state detectors on a voxel-by-voxel basis. The invention reduces experimental data requirements by using electrode signals or free charge data alone to train the model, with regularization techniques to handle reduced training data.

Training Transformers Using Sliceout Dropout

The USPTO published patent application US20260093985A1 filed by Cohere Inc., covering a system and method for training transformer neural networks using 'Sliceout' dropout operations. The method slices contiguous memory segments of weight matrices instead of randomly dropping weights, preserving regularization while reducing computational overhead and memory requirements. Application No. 19412214 was filed December 8, 2025.

Multiple Instance Learning for Content Feedback Localization Without Annotation

USPTO published patent application US20260093986A1 for a method predicting annotation spans without labeled annotation data, treating Automated Essay Scoring (AES) as a Multiple Instance Learning (MIL) task. The invention enables models to predict content scores and localize content using sentence-level predictions without annotation training data. Inventors include Scott HELLMAN, Peter W. FOLTZ, Lee BECKER, and William R. MURRAY.

Computer System and Method for Quantizing Artificial Neural Network Model

USPTO published patent application US20260093987A1 by NOTA, INC. for a neural network quantization method that identifies outliers from activation elements and regularizes weights based on outlier relevance before quantization. The application was filed November 12, 2024.

Attention Mechanism Adjustment Method Based on Attention Score and Computing Device

The USPTO published patent application US20260093988A1 by Industrial Technology Research Institute for an attention mechanism adjustment method in Transformer models. The invention performs cross-head column-wise aggregation on attention score matrices to determine token importance, then prunes less important tokens before softmax operations. This reduces computational complexity while maintaining model performance.

Electronic Device, Terminal, and Operating Method with Neural Network Lightweighting

Samsung Electronics Co., Ltd. filed patent application US20260093989A1 for an electronic device and operating method that generates candidate neural networks by excluding nonlinear layers, selects candidates based on importance and latency values, and merges successive convolution layers. The patent (Application No. 19329970) was published April 2, 2026, with inventors Jinuk Kim and Hyun Oh Song. The invention relates to CPC classifications G06N 3/082 and G06N 3/0464, covering neural network optimization and compression techniques.

Neural Network Alignment Using Filter Layers

The USPTO published patent application US20260093990A1 describing methods for aligning pre-trained generative neural networks by introducing trainable filter layers. The filter layers process outputs from pre-trained network stacks while keeping the original pre-trained parameters fixed. Inventors: Xiangyu Qi, Xiao Ma, Ahmad Beirami.

Using Compressed Representations to Adapt Generative Models to New Context Data

USPTO published patent application US20260093991A1 for methods and systems using compressed representations to adapt generative models to new context data. The application was filed on October 2, 2025, by inventors Yoel Yehuda Zeldes, Amir Zait, Efrat Farkash, Ilia Labzovsky, and Danny Karmon. The invention involves processing compressed representations with a trained compression model to generate responses to queries through a generative neural network.

Consistency model mimics diffusion model denoising trajectory

Consistency model mimics diffusion model denoising trajectory

Amplifying Non-Linearity in Feedforward Network Module

USPTO published patent application US20260093963A1 by eight inventors (Yixing Xu, Chao Li, Dong Li, Xiao Sheng, Fan Jiang, Lu Tian, Ashish Sirasao, Emad Barsoum) covering modifications to the FFN module framework in machine learning models. The application describes an improved nonlinear function designed to decrease hidden dimensions of the FFN module, thereby reducing computational costs. Application No. 18957488 was filed on November 22, 2024.

Block-Based Compression of Neural Network Data

The USPTO published patent application US20260093964A1 by inventor Navin Patel describing a method for compressing neural network data blocks. The system converts trained neural network weights into 2D blocks and treats them as color data (monochrome or RGB) for block-level compression, reducing memory and bandwidth requirements while maintaining data fidelity during inference operations.

Sparse activation-aware weight loading for ML inference

The USPTO published patent application US20260093965A1 by inventors Mahdi Heydari et al. describing techniques for sparse activation-aware weight loading and inference for machine learning models. The application covers methods for identifying non-zero values in activation tensors and loading only corresponding weight channels to optimize memory usage and computational efficiency during ML inference.

Pre-compute Storage Memory Neural Network System

The USPTO published patent application US20260093966A1 filed by Swarup Bhunia et al. describing a pre-compute storage memory system for neural networks. The system includes a memory array storing weight activation sets and an address decoder that retrieves pre-computed results matching input operands for use in determining neural network layer activations during inference or training.

Neural Network Sparsity Methods for Computational Efficiency

USPTO published patent application US20260093967A1 for methods improving neural network efficiency through increased sparsity. The application, filed October 2, 2025, covers systems that use predictor values to determine item importance and selectively limit computation to a proper subset. Named inventors include researchers from Google and academic institutions.

Memory Circuits and Methods for Encoder/Decoder Dual Mode for Compute-in-Memory

The USPTO published patent application US20260093968A1 by Taiwan Semiconductor Manufacturing Company (TSMC) covering integrated circuit memory circuits for compute-in-memory (CIM) operations. The patent describes dual-mode encoder/decoder circuits using multiply-accumulate (MAC) operations with data multiplexer paths for efficient processing of neural network and AI workloads.

Digital Injection-Locked Oscillator

The USPTO published patent application US20260093969A1 for a digital injection-locked oscillator invented by Franck Badets. The invention describes a circuit with an adder adding digital words, a register updating based on clock signals, and a first circuit receiving reference signals at natural frequencies. The application (No. 19341044) was filed September 26, 2025, and published April 2, 2026.

Neural Network Processor, System-on-Chip, Data Processing Method, and Storage Medium

USPTO published patent application US20260093970A1 on April 2, 2026, disclosing a neural network processor and system-on-chip design. Inventors Yongkang XU and Yibo HE filed Application No. 19388033 on November 13, 2025. The invention (CPC: G06N 3/063) relates to systems-on-a-chip technology for neural network operations.

Neural Network Weight Fetching Based on Calibration Operations

USPTO published patent application US20260093971A1 by inventor Weiliang Liu, disclosing techniques for fetching neural network weights based on calibration operations. The application (G06N 3/065) covers apparatuses and systems for executing neural networks with optimized weight retrieval. No compliance obligations or deadlines are associated with this publication.

Method and Apparatus for Determining Physical State of Object via Neural Network

USPTO published patent application US20260093961A1 for a computer-implemented method using neural networks to determine physical states of objects. The method divides object shapes into sub-shapes, computes local solutions via a neural network model, and derives global solutions representing the object's physical state. Invented by Jianing Huang, Youjia Wu, Kaixuan Zhang, and Ze Cheng. This is an informational publication with no regulatory obligations.

System and Method for Identification of Archaeological Features Using Remotely Sensed Data

The USPTO published patent application US20260092777A1 for an AI-driven system designed to identify gravesites and archaeological features through remote sensing and machine learning. Invented by Nicholas A. Kuncewicz, the technology integrates multimodal data including spectral imagery and LiDAR to detect and classify features. No compliance obligations arise from this patent application publication.

Artificial Intelligence Aiding for Gyrocompassing

Honeywell International Inc. filed patent application US20260092782A1 for an AI-assisted gyrocompassing system. The invention uses a trained machine learning model to predict vehicle attitude, heading, or latitude from inertial measurement unit data, reducing initial alignment time during gyrocompass unit start-up. The application was filed October 1, 2024, and published April 2, 2026.

Methods and Systems of Predicting Total Loss Events

USPTO published patent application US20260091748A1 assigned to Cambridge Mobile Telematics Inc., covering machine learning methods for predicting total loss events. The system uses mobile device sensors to detect crash events and generates confidence scores for total loss predictions using multiple ML models. Application was filed December 9, 2025 and published April 2, 2026.

Interactive Driving Environment Stories via Learning Model

USPTO published patent application US20260091321A1 disclosing a system and method for generating interactive stories about driving environments using a learning model. The invention acquires sensor data from vehicles, contextual cues about occupants, and their preferences to generate and dynamically adjust interactive virtual driving stories displayed within the vehicle. Application No. 18902256 was filed on September 30, 2024.

ML-based Furniture Layout Generation in Virtual Spaces

USPTO published patent application US20260091318A1 on April 2, 2026, disclosing machine learning techniques for generating recommended furniture layouts in virtual spaces within online games. The system processes space configurations, user criteria such as furniture style and budget, and characteristics of proximate spaces to output AI-generated furnished layouts. Inventors include Anushka Nair, Nitish Victor, Caedmon Somers, Ashwin Nathan, Han Liu, Jesse Hans Stokes Harder, and Marjan Moodi.