Microsoft Technology Licensing, LLC

Albert Ziegler

Max Schaefer

Johan Sebastian Heesemann Rosenkilde

Oegerikus De Moor

Edward Emil Aftandilian

Balabhadra Graveley

Rahul Pandita

Arjuna Ganesh Sittampalam

Maksym Taran

Krzysztof Stanisław Cieślak

A large language model, trained on source code and natural language text generates a unit test for a change to a file in a pull request of a code repository. An ordered sequence of prompts is created and each is applied serially to the large language model to perform an individual task that leads to the generation of the unit test. The unit test may be added to an existing file or generated as a newly-created file. Each prompt includes the data from a previously-issued prompt of the ordered sequence in order for the model to retain contextual knowledge learned previously. The model generates the unit test as update commands when the unit test is added to an existing file.

Conversational unit test generation using large language model

Using indentation to trim output of a language synthesis model. A computer system determines a first indentation level associated with a cursor within an editor buffer. The computer system identifies a block of line(s) from an output from a language synthesis model. Each line in the block meets at least one criterion from a set of criteria, including a first criterion that the line is an initial line in the output from the language synthesis model, a second criterion that the line has a second indentation level that is greater than the first indentation level, and a third criterion that the line contains a continuation token. The computer system identifies a particular line, immediately following the block, that does not meet any criterion from the set of criteria. The computer system trims the output from the language synthesis model from the particular line onward.

Using indentation to trim output of a language synthesis model

An elidable text is constructed that prioritizes the content included in a prompt to a large language model having a fixed-size context window. The elidable text is generated from developer-generated instructions or automatically for source code within a source code editor. A source code editor may include a feature that selects certain lines of code as important or focused which are assigned a high-priority value. A changed line, a line of source code at a current cursor position, lines of source code at the beginning of a file and those that output data are considered focused lines. Non-focused lines are assigned a priority based on a distance from a focused line. The elidable text constrains the data included in a prompt to the context window size by replacing the lowest-valued lines of text and source code with a replacement string.

Elidable text for prompt crafting

Classifying packages based on generic feature extraction. A computer system identifies a set of training packages, including a first subset known to be malicious, and a second subset known to be benign. The computer system extracts a set of training feature vectors from the set of training packages by inputting each training package to a feature extraction model, which generates a training feature vector for each training package. The computer system trains a classification model using the set of training feature vectors. After training the classification model using the set of training feature vectors, a subject package is classified as malicious or benign based on extracting a feature vector for the subject package by inputting the subject package to the feature extraction model, and inputting the feature vector to the classification model.

Generic feature extraction for identifying malicious packages

A computer system identifies a code synthesis model and a set of software repositories, each comprising corresponding source code and corresponding tests for testing the corresponding source code. The computer system tests a plurality of generated source code predictions using the set of software repositories. The testing includes, for each software repository in the set of software repositories, identifying a portion of the corresponding source code of the software repository that is covered by the corresponding tests of the software repository, using the code synthesis model to generate a source code prediction for the portion of the corresponding source code, and using the corresponding tests of the software repository to test the source code prediction that was generated using the code synthesis model. Based on testing the plurality of generated source code predictions using the set of software repositories, the computer system generates an evaluation of the code synthesis model.

Code synthesis model evaluation harnessing real-world code repositories

Inferring source code structure based on indentation. A portion of source code comprising a plurality of lines is identified, and a corresponding indentation level for each line is determined. A plurality of nodes, each representing a different line in the plurality of lines are generated. The nodes are arranged into an ordered tree based on the corresponding indentation level for each line. Based on an identified line in the plurality of lines, a subset of the plurality of lines are identified from the ordered tree, based on identifying a subtree that comprises a node representing the identified line. A source code action is initiated based on the subset of the plurality of lines.

Growing community of inventors

Uppsala, Sweden

Albert Ziegler