Intuit Inc.

Certain aspects of the disclosure provide techniques for categorization by a device. An example method includes receiving input information regarding a plurality of classification targets and a plurality of categories for classification of the plurality of classification targets; generating a plurality of embeddings for the input information using a first model, the plurality of embeddings including: a set of first embeddings associated with the plurality of classification targets, and a set of second embeddings associated with the plurality of categories; determining that a similarity score for the set of first embeddings and the set of second embeddings fails to satisfy a threshold; generating, based on the similarity score failing to satisfy the threshold and using a graph neural network (GNN), a classification of the plurality of classification targets in accordance with the plurality of categories; and outputting information regarding the classification.

Aspects of the disclosure provide techniques for defining communication parameters. A method includes obtaining input information relating to a communication source and a set of communication targets; generating, in accordance with a target function that indicates a set of constraints to be satisfied by a set of communication parameters, the set of communication parameters for communications from the communication source to the set of communication targets, the generating including: performing a linear programming operation to generate one or more first communication parameters of the set of communication parameters based on the set of constraints, and performing a mixed-integer programming operation to generate one or more second communication parameters of the set of communication parameters based on the set of constraints, wherein the set of communication parameters is based on a predictive model configured to predict a behavior of the set of communication targets; and providing the set of communication parameters.

A computer-implemented method involves receiving input with information about treatments for targets. A large language model (LLM) agent identifies tools for configuring treatments, selecting targets, and allocating treatments. The LLM agent generates and provides application programming interface (API) inputs to the tools. The LLM agent receives results from the tools, determines if the results satisfy a condition, and outputs an indication based on the results. The method includes generating computer code for treatments, applying treatments, and performing statistical analysis. The system iteratively adjusts treatments, targets, or allocations until conditions are met. The system identifies ambiguities and obtains clarifications, training the LLM agent accordingly. The processing system includes memory and processors to execute instructions, receive input, identify tools, generate API inputs, provide inputs to tools, receive results, determine conditions, and output indications.

Certain aspects of the disclosure provide a method for fraud detection. The method may include obtaining a plurality of fraud detection rule combinations, each fraud detection rule combination of the plurality of fraud detection rule combinations having an assigned risk level, wherein each fraud detection rule combination comprises a combination of two or more individual fraud detection rules; receiving transaction data associated with an account; identifying, from the plurality of fraud detection rule combinations, one or more fraud detection rule combinations triggered by the transaction data, and generating a consolidated risk score for the account by aggregating the assigned risk levels of each triggered fraud detection rule combination; and generating a fraud alert when the consolidated risk score exceeds a threshold, wherein the fraud alert includes an explanation identifying the one or more fraud detection rule combinations that contributed to the consolidated risk score.

Aspects of the present disclosure relate to structuring prompt frameworks in machine learning models. Embodiments include instructing a machine learning model via a prompt to generate an output according to a series of steps that reference one or more sections of the prompt. Embodiments include providing the machine learning model, via the prompt, with the one or more sections delineated with corresponding tags, each section of the one or more sections being referenced in the prompt via a corresponding tag. Embodiments include providing the machine learning model, via the prompt, with an output template indicating a target structure for the output and instructing the machine learning model to score the generated output according to a set of scoring criteria. Embodiments include instructing the machine learning model via the prompt to provide the output only when a calculated score, based on the scoring of the generated output, exceeds a threshold value.

Security against identity theft in generative artificial intelligence applications includes receiving, from a user computing device, an online user prompt to a large language model (LLM), the online user prompt associated with a user identifier, tokenizing the online user prompt to generate a target set of tokens, and tagging each token in the target set based on parts of speech to obtain a tagged target set. Security further includes processing, by a vector embedding model, the tagged target set to generate multiple vector embeddings, processing, by a sequential model, the vector embeddings to generate a target vector, processing, by an anomaly detection model using or trained with a signature of the user identifier, the target vector to detect whether user impersonation exists. Security further includes blocking, in real time with receiving the online user prompt, access to the LLM based on detecting that user impersonation exists.

Systems and methods for hardening system prompts are disclosed herein. An example method is performed by one or more processors of a hardening system. The example method may include receiving an initial prompt for a language model (LM), generating an initial accuracy score representative of an extent to which output generated by the LM matches a target output when the initial prompt is used as its system prompt, generating an initial robustness score representative of an extent to which the LM resists adversarial attacks when the initial prompt is used as its system prompt, and iteratively transforming, using an artificial intelligence (AI)-based hardening agent in conjunction with a set of machine learning (ML)-based optimization tools and a reinforcement learning (RL) technique, the initial prompt into a hardened prompt such that the hardened prompt maximizes an increase of the initial robustness score and minimizes a decrease of the initial accuracy score.

A system and method for dynamically optimizing large language model (LLM) inference by selectively deactivating layers based on query complexity. A multi-label classifier is trained on diverse user queries and their optimal layer configurations. During inference, the classifier analyzes incoming queries to predict which LLM layers can be safely deactivated without compromising output quality. The system processes user queries through the LLM with the predicted layer configuration, reducing computational resources while maintaining accuracy. A database stores historical queries, layer configurations, and performance metrics for continuous system improvement.

A method includes obtaining multiple user interface (UI) elements by a UI generator. A set of descriptors is added to the UI elements. The UI elements are further populated with data values. A UI document including the populated UI elements with descriptors is transmitted to a client application and rendered on a user interface. A prompt is generated based on the descriptors, and the data of the UI elements. The prompt is processed by the LLM to generate an insight based on the prompt. The insight is transmitted to the client application to be rendered on the user interface.

A method of improving a retrieval augmented generation (RAG) language model ensemble. The language models are executed on first prompts to generate test outputs. Each of the first prompts includes a natural language question corresponding to correct answers only obtainable from the context. The prompts are assigned to each of the language models. The prompts avoid referencing the context. A removed question is removed to generate filtered questions. The removed question includes a corresponding one of the natural language questions for which a corresponding test output of the test outputs represents a correct answer relative to the correct answers. The language models are executed on filtered prompts to generate generated answers. Each of the filtered prompts includes the filtered questions and a corresponding command to return the generated answers from the context. A semantic model is executed on the generated answers to generate scores representing accuracies of the language models.