Technical Disclosure Common
Not a member yet
9157 research outputs found
Sort by
METHOD AND APPROACH IN LARGE LANGUAGE MODELS FOR COMPLEX QUESTION ANSWERING
No full textThe present disclosure relates to a method and system for enhancing the accuracy and flexibility of Large Language Models (LLMs) during complex question answering. The method involves identifying a plurality of decision nodes within the LLM\u27s internal reasoning structure, each representing a point of ambiguity or multiple potential reasoning paths. The reasoning process is paused at each identified decision node to present contextually relevant options to the user. User input is received at these decision nodes, specifying preferences or additional data to guide subsequent reasoning. This input is integrated into the ongoing reasoning process of the LLM, dynamically generating a response that reflects improved contextual relevance and aligns with user specifications. Furthermore, the system comprises a decision node identification module, a user input interface, a reasoning adaptation module, and a dynamic response generation module to facilitate enhanced collaborative problem-solving during complex inquiries
FINE TUNING CONTEXT WINDOW FOR OPTIMAL UTILIZATION IN GENERATIVE MODELS
No full textThe present disclosure relates to a method and a system for fine-tuning a context window for optimal utilization in generative models. The present disclosure suggests receiving an input prompt comprising a plurality of tokens and generating a baseline output based on the input prompt. Upon generating the baseline output, the present disclosure suggests evaluating at least one token of the plurality of tokens for contribution to the generated output. Subsequently, the present disclosure suggests computing rarity scores and possibility scores associated with the evaluated tokens by using a rarity matrix model and a possibility matrix model. Further, the present disclosure suggests generating a reward signal based on a combination of the baseline output, the rarity scores, and the possibility scores. Finally, the present disclosure suggests identifying and removing at least one non-contributory token based on the reward signal. As a result, the present disclosure provides enhanced resource efficiency for context utilization in generative models
System for Context-Aware Payment Recommendations Using Retrieval-Augmented Generation
No full textA system can generate context-aware payment recommendations to assist users in selecting a suitable payment method, which may be beneficial given the complexity of financial product benefits. The system may employ a multi-stage processing pipeline where, for example, an intent classifier can first infer a user\u27s need from a transactional context. A request may then be processed by a handler that can query a retrieval-augmented generation system to obtain relevant data from a knowledge base. A large language model may synthesize this data to generate a structured output, such as a JSON payload containing both human-readable text and machine-readable user interface directives. The structured JSON output is designed to bridge the gap between unstructured natural language and platform-specific interactive UI components. This process can enable a client application, such as one on a smartphone or wearable device, to dynamically present payment guidance to a user, for example, within a checkout experience
PROVIDING UNIFIED INTERFACE FOR RENDERING USER INTERFACE (UI) COMPONENTS AND APPLICATION PROGRAMMING INTERFACE (API) DOCUMENTATION
No full textThe present disclosure relates to a system and a method for dynamically fetching, filtering, and rendering one or more user interface (UI) component examples and application programming interface (API) documentation for multiple libraries and versions. The system enables users to select a library and version. The system fetches a corresponding metadata and code examples from a structured JavaScript Object Notation (JSON) file. The system provides device emulations, code/preview toggling, and API reference display in a unified interface
A SECURITY ASSESSMENT SYSTEM FOR BLOCKCHAIN WALLET RECOVERY MECHANISMS
No full textThe present disclosure relates to a system for analyzing security vulnerabilities in blockchain wallet recovery mechanisms. The system includes a blockchain wallet configured with a plurality of distinct recovery agents, a model-checking framework operable to simulate various potential adversarial scenarios, and an interface to specify parameters such as agent reputation, transaction spending power, and user response probability. Additionally, the security assessment output visually indicates secure and insecure configurations of the blockchain wallet based on interactions among the recovery agents. Furthermore, the model-checking framework simulates diverse adversarial conditions, generating an assessment matrix that highlights potential risks associated with different agent configurations. This effectively provides a comprehensive security analysis for blockchain wallets
Distributed Security Orchestration with Adaptive Filtering and Data Minimization
No full textTraditional security analytics, which relies on centralized data collection and processing, can raise concerns relating to privacy, data-transmission costs, threat-response latency, and jurisdictional data-handling requirements. This disclosure describes techniques to perform security analytics across a distributed network of edge computing nodes. By leveraging edge computing to collect and analyze data from endpoint devices, data collection/transmission is optimized and privacy is enhanced. Centralized processing via federated/distributed processing is limited. Adaptive filtering is used to dynamically adjust data processing and transmission based on detected events and risk levels. A tiered zoning approach processes high-risk events centrally, quarantines moderate-risk events for further analysis regionally, and handles low-risk events locally. The techniques optimize resource utilization and ensure efficient, scalable, and privacy-conscious security analytics. Tiered granularity improves the privacy and utility trade-off in privacy-preserving machine learning. Security learnings are globally learned and aggregated in a privacy-preserving manner
P2P SERVICE WITH DYNAMIC ACQUIRING
No full textOne embodiment includes a processing network computer implemented method for facilitating a transaction between a sender and recipient. The method uses a token or token reference to dynamically select an authorizing entity as an acquirer, thus enabling transactions among a variety of authorizing entities
Pet-Aware Navigation Using Aggregated and Verified Point-of-Interest Data
No full textDigital mapping applications may lack integrated data on pet policies for points of interest (POIs), which can lead users to consult fragmented and potentially inaccurate external sources when planning travel with pets. A system and method are described for integrating pet-compatibility data into a navigation application. The system can utilize a multi-source data engine to aggregate and harmonize information from sources such as crowdsourced user reports, third-party data providers, and artificial intelligence analysis of public web data. A harmonization framework may use confidence scoring and a rules engine to resolve data conflicts and verify the pet policies of POIs, such as restaurants, hotels, or parks. This aggregated data can be used to provide users with features such as advanced search filters, custom map visualizations that identify pet-friendly locations, and a routing service that can generate travel paths to, for example, avoid pet-restricted areas
OPEX-Aware Relaying
No full textThis technical disclosure proposes an OPEX proxy that routes client requests to cloud application instances based on the operating expenses (OPEX) of potential host data centers. Unlike traditional routing methods, which are focused on load or network health, this proxy selects data centers with the lowest current or expected OPEX, reducing energy costs without requiring ownership of the power source. The approach supports integration with existing routing mechanisms and can be extended with environmental impact metrics
Rubric-Based Evaluation for On-Brand Generative Media
No full textThis publication describes a novel system and method for evaluating artificial intelligence (AI)-generated media, focusing on prompt adherence, technical quality, and brand alignment. The framework addresses challenges in current evaluation methods by employing dynamic, multi-component rubric generation and a multimodal, rubric-based evaluation and scoring process. This approach provides a scalable, reliable, and interpretable solution for comprehensively assessing generative media