Intuit Inc.

A method including receiving an input for a language model including decoder layers having attention heads. The language model predicts a chunk of tokens based on the input. Predicting includes determining, for each of the attention heads, corresponding attention map weights from the input. The method also includes determining, from the corresponding attention map weights of each of the attention heads, a lookback ratio vector for the chunk of tokens. A hallucination score for the chunk of tokens is predicted using the lookback ratio vector. A determination is made that the hallucination score fails to satisfy a predefined threshold, λ, such that the chunk of tokens represents a hallucination of the language model. An edit direction for the corresponding attention map weights for the chunk of tokens is determined. The attention map weights of at least some attention heads are modified to generate a post-process language model.

Aspects of the present disclosure relate to a two-stage architecture for cross-encoder models that allows for improved text retrieval. The two-stage architecture may involve a baseline cross-encoder that generates baseline joint embedding representations and similarity scores for query-text pairs. Then, a kernel interpolation cross-encoder model may be used to augment the baseline similarity scores based on kernels that are placed according to a location in an embedding space that corresponds to a training example used to build the kernel interpolation cross-encoder model. The level of augmentation of a similarity score for an input query-text pair may thus be based on the location of a joint embedding representation of the query-text pair relative to a location of the training example in the embedding space.

One or more embodiments provide for a method. The method includes receiving an input to a language model having a first computational efficiency. The method also includes determining, from the input, a pruning threshold. The method also includes determining a pruning weight mask for the input from a combination of the pruning threshold and a predetermined weight mask defined for the language model. The method also includes modifying, using the pruning weight mask, the language model to generate a modified language model having a second computational efficiency greater than the first computational efficiency. The method also includes executing the modified language model on the input to generate a token. The method also includes returning the token.

A method includes adding lineage metadata to a dataset of data records. The method further includes obtaining permitted data access operations of the dataset of data records from a registered data access policy. The method further includes evaluating, responsive to a write operation being permitted to the dataset of data records, the dataset of data records against registered data schemas to determine a compliance of the dataset of data records to the at least one registered data schema. The method further includes obtaining, responsive to the dataset of data records being compliant with the at least one registered data schema, at least one registered query pattern of the dataset of data records. The method further includes determining, from the at least one registered query pattern, a storage location for the dataset of data records. The method further includes storing the dataset of data records in the storage location.

Performing web application based screen share annotations includes obtaining, from a source browser executing a web application performing a screen sharing, a location point of the source browser with respect to a shared source screen being shared. Performing web application based screen share annotations further includes calculating, from the location point, a source portion of the shared source screen that shows the web application, translating the source portion to a target portion within a target web browser executing on a target device. Performing web application based screen share annotations further includes demarcating, on the target display, the target portion within the target browser.

Aspects of the present disclosure relate to techniques for deriving cryptographic keys. Embodiments include using a key encapsulation mechanism to encrypt various data items transferred between a first remote server and a client device during a key derivation process. In some embodiments, the encrypted data items include a blinding value and a client key. Embodiments further include using a hybrid post-quantum key exchange technique to securely store and transfer keys used by the remote servers as part of the key derivation process.

At least one processor may receive a user prompt from a user and determine a user role associated with the user. The at least one processor may select a system prompt associated with the user role and identify at least one data source from among a plurality of available system prompts each associated with a different respective user role and indicating one or more data sources accessible according to the respective user role. The at least one processor may build a final prompt configured to retrieve data responsive to the user prompt from the at least one data source. The at least one processor may query a generative artificial intelligence (GenAl) system using the final prompt, the querying causing the GenAl system to access the at least one data source identified by the system prompt and provide a response incorporating data obtained from the at least one data source.

A method includes obtaining an expert availability forecasting model including at least a first time slot, a multitude of available experts corresponding to the at least first time slot, and a predicted customer demand corresponding to the at least first time slot. An expert allocation model is generated, including a first set of available experts allocated to the at least first time slot based on the expert availability forecasting model. A first set of customers is selected, corresponding to the predicted customer demand. The first set of customers is matched to the corresponding first set of available experts in the at least first time slot, based on the expert allocation model, to obtain a customer-expert allocation schedule. The method further includes transmitting the customer-expert allocation schedule to a performance analytics dashboard. The method further includes monitoring at least one data source for at least one expert availability event.

Aspects of the present disclosure relate to automated document integration. Embodiments include receiving one or more documents from a user. Embodiments further include extracting, using a computer vision machine learning model, a first data item from the one or more documents. Embodiments further include generating a prompt based on a second data item not being present in the one or more documents, wherein the prompt comprises a request for the user to provide the second data item. Embodiments further include receiving, based on the prompt, an additional document from the user. Embodiments further extracting, using the computer vision machine learning model, the second data item from the additional document. Embodiments further include generating, via a particular machine learning model based on the first data item and the second data item, an electronic document that corresponds to a particular format.

Certain aspects of the disclosure provide systems and methods for field narration. For example, field narration may be configured to populate and verify a form populated based on audio data. A user interface may send a verification request for the form, including a data field and a value to be verified to a field narrative service. The field narrative service may determine relevant portions of the audio data and the transcript for the value to be verified. The user interface may play the portion of audio data, and display the transcription for verification.