USPTO Patent Applications - AI & Computing (G06N)

Capital One Patent: Trifurcated Prompts for AI Communication Processing

The USPTO published Capital One Services LLC's patent application US20260099519A1 on April 9, 2026, covering systems and methods for processing communications using trifurcated prompts with perturbed contexts for large language model interaction. The system receives inbound communications, determines context, applies perturbation models to generate alternative tokens, and constructs prompts for AI processing.

Topological Sparse Training Process for Machine Learning Models

USPTO published patent application US20260099725A1 for a topological sparse training process for machine learning models. The application covers methods for executing ML models with attention heads, using loss functions configured for preferential attachment of neurons, and modifying parameters including setting values to zero to generate sparse ML models. The invention aims to optimize neural network architecture through preferential attachment mechanisms during training.

Information Classification Using Local, Global Models

USPTO published patent application US20260099726A1 by inventors Song Bai, Yujun Shi, Wenqing Zhang, and Bin Lu on April 9, 2026. The application covers a method and apparatus for information classification using local and global classification models with decorrelation to address dimensional collapse of feature representations. The invention was filed on July 7, 2023, under application number 19099724.

Language Models Having Reduced Size While Maintaining Performance and Reducing Hallucinations

USPTO published patent application US20260099727A1 titled 'Language Models Having a Reduced Size While Maintaining Performance and Reducing Hallucinations' filed January 30, 2025 by inventors Jeffrey Daniel Esposito, Henry Svendsgaard, Aishwarya Dharani Arul, and Tabor Scott. The application discloses a computer program product that iteratively trains a language model by adjusting hyperparameters such as number of layers, hidden units, and parameters, selecting the smallest model configuration that meets a predetermined performance threshold to reduce hallucinations.

Task Agnostic Embedding Based Labeling Escalation On Fly

The USPTO published patent application US20260099728A1 for a machine learning method involving task-agnostic embedding-based labeling escalation. The invention uses a two-pass system: a first-pass model makes an initial decision, and if embedding analysis indicates a need for escalation, a higher-complexity second-pass model generates a final decision. The application was filed on October 7, 2024, by inventors MohammadReza GHAEINI and Muhaimenul ADNAN and classified under CPC G06N 5/01.

Compound Docking Calculation Processing Device Method and Program

Institute of Science Tokyo filed USPTO patent application US20260099729A1 for a compound docking calculation processing device, method, and program. The invention enables combinatorial optimization in fragment-based compound docking calculations for drug discovery by using fragment decomposition, fragment docking calculation, interaction evaluation, and optimization processing units.

Rules Engine AI Guides End Users Through Transactions

The USPTO has published patent application US20260099731A1 filed by inventors Stephen Barrett Cichy and Markus Daniel Bockle. The application covers a rules engine configured to express complex logic, handle time and event synchronization, provide insights into rule execution, and model uncertainties. The system enables AI-guided assistance for end users completing transactions.

SYSTEM FOR AUTOMATED DATA ANALYSIS AND DECISION-MAKING FOR COMPLEX PRODUCT CONFIGURATION

The USPTO published patent application US20260099734A1 filed by Vertiv Corporation on October 7, 2025. The application covers an AI-driven product configurator that pre-processes data from multiple sources, extracts features, and infers results through ML models deployed on edge devices with dynamic resource allocation. Analytical reports with visualizations and alerts are generated based on detected anomalies.

System for Providing Software Related Answer Based on Trained Model

USPTO published patent application US20260099736A1 for a system providing software-related answers using a trained AI model. The invention involves natural language understanding for code base queries, using a custom enhancement model to determine user intent and a trained model (with generic and specific inputs) to generate natural language responses. Inventors: Joel Hart and Douglas Lee of SAFERITE.

Prediction Model Training Using Detected Anomalies

USPTO published patent application US20260099735A1 for a system and method of training prediction models using detected anomalies. The system trains multiple models using historical data, selects the best-performing model based on test data, generates forecasts, identifies anomalies between forecasts and model outputs, and incorporates user feedback to retrain and improve prediction accuracy.

Trustable AI Chat, Privacy, Single Interface

The USPTO published patent application US20260099733A1 on April 9, 2026. The application covers systems and methods for enabling clients to interact with multiple AI-based applications and large language model services through a single user interface with assured privacy and security. The invention involves a network device that stores service level information for multiple external AI services and converts natural language requests into service call sequences.

Variable-bit Adaptive Sensing Circuit for Analog Neuromorphic Systems

USPTO published patent application US20260099703A1 for a variable-bit adaptive sensing circuit system designed for analog neuromorphic systems. The system comprises a sensing circuit, error detection circuit, and analog-to-digital conversion circuit for processing synapse array outputs. The application was filed on April 22, 2025, and names Hyung Min Lee, Min Seong Um, and Min Il Kang as inventors.

Semantic Communication Method for AI Model Transmission

USPTO published patent application US20260099699A1 disclosing a semantic communication method for AI model transmission. The method encodes a target AI model using a preset semantic encoder to output semantic information, which is then transmitted through a wireless channel to a semantic decoder that reconstructs a corresponding model. The semantic encoder and decoder are trained using a semantic contrastive loss function to minimize semantic distance between target and enhanced samples while maximizing distance with remaining samples.

Pyramid Key-Value Cache Compression for Transformer Models

USPTO published patent application US20260099695A1 on April 9, 2026, for a method of operating transformer models with algorithmic key-value cache memory allocation across decoding layers. The invention allocates a fixed memory budget progressively across layers, with higher layers receiving smaller cache allocations. Each layer independently determines maximum key-value vector pairs based on its allocated cache.

RAG Content Quality Evaluation Method Using Large Language Models

USPTO published patent application US20260099693A1 titled 'Content Quality Evaluation for Retrieval Augmented Generation Systems.' The patent covers a method for objectively evaluating content output by RAG systems using large language models to generate evaluation metrics and present comparative quality data across multiple RAG system configurations.

Patent for 3D Colon Model From CT Scans

Patent for 3D Colon Model From CT Scans

Convolution Network for Relevant Motion Detection in Surveillance Video

The USPTO published patent application US20260099927A1 describing AI methods for detecting relevant motion of persons and vehicles in surveillance videos. The application covers a convolution network with spatial-wise and temporal-wise max pooling elements that generates prediction results for relevant motion detection. The application was filed on May 20, 2025, by inventors Ruichi Yu and Hongcheng Wang.

Fujitsu Patent - Discrete Optimization Using Continuous Relaxation and Machine Learning

USPTO published patent application US20260099565A1 assigned to Fujitsu Limited. The application covers a non-transitory computer-readable medium and calculation method for discrete optimization using continuous relaxation with machine learning. The invention involves applying perturbations to discrete optimization problems and training a machine learning model to output solutions.

AI Classification Detects Software Intrusive Action Provisions

USPTO published patent application US20260099582A1 for an AI system that classifies software information to detect provisions indicating intrusive actions. The system uses machine learning to identify when a software program will perform intrusive actions and provides classification to destination devices. Inventors: Shannon Sabens, Marian Radu, Jeffrey Kaplan.

Creating and Extracting Training Data from Storage Systems to Train Machine Learning Models for Ransomware Detection

The USPTO published patent application US20260099597A1 on April 9, 2026, describing methods for generating machine learning training data from storage systems to train ransomware detection models. The invention creates snapshots of storage volumes, generates ransomware traces using hidden volumes and benign traces from regular volumes, extracts features into an advanced features table, and trains ML models using the generated training data. The application was filed on October 4, 2024, under Application No. 18907467.

Lane Violation Detection Using Convolutional Neural Networks

The USPTO published patent application US20260100058A1 filed by Hayden AI Technologies, Inc. on August 12, 2025. The application discloses systems and methods for detecting traffic lane violations using convolutional neural networks, with bounding boxes for vehicles and polygons for lane detection. Inventors include Vaibhav Ghadiok, Christopher Carson, and Bo Shen.

Modeling Agents Using Local and Global Models

USPTO granted Patent US20260099646A1 for a method of modeling agents within an environment using local continuous models, local discrete models, and global continuous models to predict agent behaviors. The patent covers determining local and global continuous aspects alongside local discrete aspects, then using these to determine agent behaviors via modeling systems. Inventors include Leslie Ann Canavera, Lauren Brooke Decker, and Christopher Rex Curry.

Metadata Centric AI Class Reassignment Patent

USPTO published patent application US20260099765A1 for a metadata-centric AI system that reassigns data classifications based on performance metrics and confusion matrix analysis. The invention derives group-specific thresholds from prior classification instances to evaluate and update predicted classifications. Inventors include Madhusoodhana Chari Sesha, Pradeep Kumar Surenran, Ankush Anshuman, Akshay Jain, and Surya Thankamony Somanathan.

Integrated Customer Intelligence Platform and Method

The USPTO has granted Patent Application US20260099856A1 to Kyocera Document Solutions Inc. for an integrated customer intelligence platform and method in document management systems. The system collects customer market data including characteristics, engagement status, and usage patterns to generate actionable recommendations using predictive heuristics models with cyclicality and seasonality probability scoring. Inventors include Selim ZAMAN.

Computing User-Specific Item Prices Using AI

USPTO published patent application US20260099858A1 by inventors Veijo Heinonen and Mikko Saikko disclosing a method and apparatus for calculating user-specific item prices using artificial intelligence. The system determines personalized pricing based on user attribute profile vectors, item attribute vectors, price elasticity metrics, and appeal scores. The application (No. 19416874) was filed on December 11, 2025.

AI Measures Net Gain Loss for Uncertain Events Using Monte Carlo Simulations

The USPTO published patent application US20260099795A1 titled 'System and a Method for Measuring Net Gain and Loss of Alternatives for Uncertain Events.' The application, filed by inventor Ernest Forman on August 31, 2024, discloses an AI system that receives uncertain events, alternatives, objectives, and certain events, then uses Monte Carlo simulations on evaluated likelihoods and consequences to determine expected loss or gain for each alternative. This is a routine USPTO publication of a patent application in the decision-support software field.

Universal Machine Learning Pipeline Execution System and Method

USPTO published patent application US20260099305A1 titled 'Systems and Methods for Universal Machine Learning Pipeline Execution,' filed December 10, 2025. The application discloses methods for automated machine learning model development, including parsing configuration files, generating model code, creating ML pipelines, monitoring execution, and producing trained models with performance data. Inventors: Rameshchandra Bhaskar Ketharaju, Anjeet Kumar, and Shuvam Sengupta. CPC classifications include G06F 8/35, G06F 8/31, G06F 11/3476, and G06N 20/00.

Hierarchical Reinforcement Learning Controls Industrial Facility

The USPTO published patent application US20260099128A1 on April 9, 2026, filed by inventors William Wong, Praneet Dutta, and Jerry Jiayu Luo. The application covers methods and systems for controlling industrial facilities using hierarchical reinforcement learning with high-level and low-level neural network controllers. CPC classifications include F28F 27/003, G05B 13/027, and G06N 3/092.

Asynchronous Quantum Information Processing System Reduces QIPU Dead Time

USPTO published patent application US20260099354A1 by William Joseph Zeng describing an asynchronous approach to implementing quantum algorithms to reduce dead time of quantum information processing units (QIPUs). The system uses a controller to manage multiple parameter sets for quantum programs, allowing the QIPU to continue executing while the controller processes results and determines updated parameters. This approach enables the QIPU to operate with minimal or no idle time.

Dendritic Computation Neural Network Patent, Apr 9

Dendritic Computation Neural Network Patent, Apr 9

Enhanced Artificial Intelligence Virtual Assistants Patent Application

The USPTO published patent application US20260099676A1 for Zoom Video Communications, Inc. covering enhanced AI virtual assistant methods. The application describes receiving user requests, determining intent, identifying services, and generating responses. The application was filed October 7, 2024, and published April 9, 2026.

SAFETY ALIGNMENT FOR LANGUAGE MODELS USING MODEL-GENERATED SAFETY CATEGORIES

USPTO published patent application US20260099707A1 for safety alignment techniques in language models. The application describes using an ensemble of generative AI models to generate machine-defined safety labels for interactions, applying majority voting with predefined safety labels to revise training data labels, and training language models to implement guardrails restricting unsafe content generation. The application covers ensemble-based safety labeling and alignment training methodologies for AI systems.

METHOD AND SYSTEM FOR DEPLOYMENT OF LARGE LANGUAGE MODELS (LLM) IN CLOUD INSTANCES

Tata Consultancy Services Limited filed USPTO patent application US20260099706A1 for a method and system to deploy LLMs in cloud instances. The system evaluates cloud instance feasibility based on LLM model size and available storage, determines latency values for batch sizes across LLM-accelerator pairs, and generates deployment recommendations based on latency, cost, workload, application type, and performance metrics.

Flexible Prompt Guardrails System for Generative AI

USPTO published patent application US20260099719A1 for a flexible and extensible prompt guardrails system for generative AI. The system intercepts prompts intended for a generative AI system, extracts feature vectors using specialized models, and evaluates them against rules to determine whether to block or allow the prompt. The system supports adding or removing features and updating evaluation rules based on testing. The application was filed on October 3, 2024.

NEC Corporation Multi-Sensor Encoding and Adversarial Estimation Machine Learning Device Patent Application

USPTO published patent application US20260099724A1 for NEC Corporation's machine learning device that trains encoding models for sensor data and adversarial estimation models. The invention encodes first and second sensor data into codes, trains an adversarial estimation model to estimate cross-modal codes, and trains the encoding model to resist adversarial estimation. This publication affects technology companies and manufacturers developing multi-sensor machine learning systems.

Quantum Circuit Optimization via Coordinate-Descent Method

Quantum Circuit Optimization via Coordinate-Descent Method

Quantum Error Mitigation for Probability Distributions

The USPTO published patent application US20260099753A1 for a quantum computing system that performs error mitigation on probability distributions obtained from quantum circuit observables. The system executes multiple shots of a quantum circuit to obtain noise probabilities, determines expectation values, performs error mitigation, and transforms results into error mitigated probability distributions. The application was filed on October 9, 2024.

Hardware and Parameter-Aware ML Model GPU Efficiency Tuning Systems

USPTO published patent application US20260099757A1 for hardware and parameter-aware machine learning model GPU efficiency tuning systems. The application includes claims for methods and systems that receive ML training requests with fixed and dynamic configurations, generate task embeddings, train prediction modules on known configurations, and return optimal training efficiency configurations based on model utilization scores. Inventors include Pin-Lun Hsu, Vignesh KOTHAPALLI, Animesh SINGH, Qingquan SONG, Yun DAI, and Shao TANG. Filing date was October 4, 2024, with application number 18906517.

Machine Learning Model Training Using Randomized Solutions to Find Global Minimum

The USPTO published patent application US20260099758A1, filed October 4, 2024, for a machine learning technique that identifies global minimums across local minimums. Inventors Bikramaditya Padhi and Ramprasadh Kothandaraman disclosed an application server method using randomized solutions and threshold-based evaluation to optimize model training.

Methods and Apparatus to Process Training Data for an AI-Based Model

The USPTO published patent application US20260099759A1 by Niall Fitzgerald, covering methods and apparatus for processing AI training data using feature transformation, hash signature generation, and clustering techniques. The application relates to apparatus comprising interface circuitry and programmable circuits to filter training data clusters and train AI-based models. The application was filed on October 4, 2024, and published on April 9, 2026.

NEC Language Model Efficient Training Method, Decision Support

NEC Language Model Efficient Training Method, Decision Support

Multimodal Retrieval Augmented Generation for Visually Rich Documents

USPTO published patent application US20260099698A1 filed by JPMorgan Chase Bank, N.A. The application covers multimodal retrieval augmented generation (RAG) methods for visually rich documents using a page-wise chunking algorithm. The system embeds text, spatial, and visual features from document pages into vectors, retrieves relevant chunks based on query similarity, and generates responses via a generative model.

Training Content Authenticity Validators via Variable Resolution Game

USPTO published patent application US20260099720A1 assigned to D5AI LLC, covering computer-implemented systems and methods for training content authenticity validators using variable resolution game theory in adversarial machine learning. The application, filed on October 25, 2024, describes training generators and discriminators in a simulated multi-player adversarial game relationship, including two-person zero-sum game configurations with non-simultaneous parameter updates.

Generating Three-Dimensional Molecule Structures Using Generative Artificial Intelligence Models

USPTO published patent application US20260100254A1 titled 'Generating Three-Dimensional Molecule Structures Using Generative Artificial Intelligence Models' filed on April 15, 2025 by inventors Daniel Alexander Reidenbach and Filipp Nikitin. The application discloses methods for using transformer layers and graph neural networks in a generative AI model to generate and refine 3D molecular structures for drug discovery based on initial structure, atom type, and bond type inputs.

Adaptive Data Loader for Bridging Legacy Data Sources and Artificial Intelligence Model Training

USPTO published patent application US20260099766A1 titled 'Adaptive Data Loader for Bridging Legacy Data Sources and Artificial Intelligence Model Training.' The application discloses a process for identifying legacy data sources, adapting data loaders to process data into a training format, and training AI models using the processed batches. Application No. 19064832 was filed on February 27, 2025.

Neural Network Parameter Scaling Method for Federated Learning

The USPTO published patent application US20260099768A1 on April 9, 2026. The application discloses a method for scaling model parameters in federated learning systems where a user equipment applies a scaling factor to parameters based on the number of local training samples before transmitting scaled parameters to a network node. The application was filed on October 11, 2023.

AI Apparatus for Machine Operator Performance Feedback Correlation

The USPTO published patent application US20260099769A1 by inventors Bradford Everman and Brian Bradke, assigned to GMECI LLC, covering an AI apparatus for machine operator feedback correlation. The apparatus includes a processor and memory configured to receive performance data from sensing devices, classify performance data into categories, calculate performance determinations, and generate feedback correlations through a machine learning model. The system employs neural network architectures (CPC: G06N 20/00, G06N 3/0442) for correlating operator feedback with performance metrics.

Patent for Unlearning Text Data in LLMs

Patent for Unlearning Text Data in LLMs

Knowledge base article generation using AI

Knowledge base article generation using AI

Information Processing Device Using Attention Regions for Recognition Tasks

USPTO published patent application US20260099737A1 for Toshiba (Kabushiki Kaisha Toshiba) on April 9, 2026. The application covers an information processing device that executes recognition tasks using reference data, attention region information, and explanatory information about attention regions. No compliance obligations or deadlines are established by this publication.