Microsoft Technology Licensing, LLC

Carina Suzana Negreanu

Andrew Donald Gordon

Advait Sarkar

John Herbert Martin Williams

Rasika Mudumbai Chakravarthy

Daniel Fabian

Brian Paul Slininger

Sumit Gulwani

Benjamin Zorn

Vu Minh Le

Neil Toronto

José Pablo Cambronero Sánchez

Sruti Srinivasa Ragavan

Gust Ben Anneloes Verbruggen

Christian Leopold Bejamin Poelitz

Mukul Singh

Technology is disclosed herein for the integration of spreadsheet environments and LLM services. In an implementation, an application receives a natural language input from a user associated with a spreadsheet. The application generates a prompt based on the user input and at least a portion of the spreadsheet. The prompt includes a statement of the problem, a request for an LLM service to identify preparatory steps before generating a solution to the problem, and a request to include the preparatory steps in output that includes the solution to problem. The application service receives a reply to the prompt from the LLM service that includes the output. The application service implements the preparatory steps with respect to data in the spreadsheet and implements the solution to the problem with respect to the data in the spreadsheet.

Predicate-guided preparation for LLM integrations in spreadsheet environments

The indirect querying of models to determine capabilities possessed by the model. Such indirect queries take the form of model input that potentially includes a natural language input user data. Such model input is structured such that the output of the model is either not natural language at all, or else is natural language that is not semantically responsive to the natural language input. Nevertheless, the output is evaluated to estimate or determine the capability possessed by the model. Thus, models may be more fully utilized to their better potential.

Model capability extraction

Technology is disclosed herein for the integration of spreadsheet environments and LLM services. In an implementation, an application service inputs a first prompt to a LLM service to provide a formula suggestion for a spreadsheet. The application service receives a first output from the LLM service that includes a first formula in a programming language having a syntax that does not support comments. The application service generates a second prompt instructing the LLM service to provide a chain-of-thought breakdown of the first formula and receives a second output. The second output includes a second formula in the same programming language as the first formula and multiple comments corresponding to multiple portions of the second formula.

Chain of thought reasoning for LLM integrations in spreadsheet environments

The generation of a response to a task prompt that represents a task to perform on declarative code. The response is generated with the aid of a language model that was trained on imperative code. The declarative code includes declarations about data. A task prompt represents a task to perform on the declarative code. At least a portion of the declarative code and at least a portion of the task prompt are converted into input imperative code. The input imperative code is then caused to be provided as input to the language model, resulting in the language model generating output imperative code. At least a portion of the output imperative code is then converted into a response to the task prompt.

Responding to task prompt on declarative code using language model

Some embodiments perform fully unsupervised string error detection and repair, for strings containing both syntactic errors and semantic errors. In some embodiments, string error mitigation includes learning a regular-expression-based pattern that characterizes at least a predefined fraction of a set of string data values, identifying and masking a semantic substring, deriving edits to repair a string which contains the semantic substring, and replacing a mask value with a replacement substring that conforms to the pattern. Some embodiments utilize a language model. Some embodiments include or utilize a pattern learner, one or more learned significant patterns, and a repair engine. Some embodiments exploit program execution information to generate patterns when significant patterns are not initially found.

String data error detection and repair

Growing community of inventors

Cambridge, United Kingdom

Carina Suzana Negreanu