Granted Patent B2

US 12,572,749 · App. 18/232,485 · Granted Mar 10, 2026

Inventors: Mohsen Fayyaz (Berlin, DE); Eric Chris Wolfgang Sommerlade (Oxford, GB); Justin James Wagle (Pacifica, CA); Vivek Pradeep (Bellevue, WA)

Assignee: Microsoft Technology Licensing, LLC

G06F40/35G06F40/284G06N20/00

Abstract

Claims (60)

1 . A method for interacting with a machine-trained model over plural dialogue turns of a dialogue using resources of one or more computing systems, comprising:

receiving input information for a current dialogue turn of a dialogue;

generating a first prompt that includes the input information and first abstract token information that has been generated by the machine-trained model in one or more prior dialogue turns;

using the machine-trained model to map the first prompt to a response;

generating a second prompt that includes at least the input information and the response; and

using the machine-trained model to map the second prompt to second abstract token information for use in a subsequent dialogue turn,

the receiving, generating a first prompt, using the machine-trained model to map the first prompt, generating a second prompt, and using the machine-trained model to map the second prompt being repeated at least one time, wherein the method interleaves generating of responses with generating of abstract token information,

each instance of abstract token information expressing a current state of a dialogue history of the dialogue with fewer tokens than a full dialogue history, which reduces a size of a prompt that incorporates said each instance of abstract token information,

wherein latency of the machine-trained model and an amount of the resources that are consumed by the machine-trained model depends on sizes of prompts processed by the machine-trained model.

2 . The method of claim 1 , wherein the machine-trained model has two parts, including a response-generating model part and an abstract-token-generating model part.

3 . The method of claim 2 , wherein the abstract-token-generating model part has a smaller size, and has fewer weights, than the response-generating model part.

4 . The method of claim 3 , wherein the abstract-token-generating model part is implemented by a local computing system and the response-generating model part is implemented by a network-accessible computing system.

5 . The method of claim 1 , wherein the machine-trained model is a language model that includes attention logic for processing input embedding information by determining relevance of a first part of the input embedding information when interpreting a second part of the input embedding information.

6 . The method of claim 1 , wherein the first abstract token information includes one more abstract tokens generated by the machine-trained model.

7 . The method of claim 1 , wherein the second prompt also includes the first abstract token information.

8 . The method of claim 1 ,

wherein the second prompt also includes machine-trained abstract-token-prompting information, the machine-trained model interpreting the abstract-token-prompting information as a signal that the machine-trained model is to generate the second abstract token information,

wherein the abstract-token-prompting information includes one more abstraction tokens, and

wherein the first prompt does not include the machine-trained abstract-token-prompting information.

9 . The method of claim 1 , wherein the input information expresses an input query and text extracted from a document.

10 . The method of claim 1 , wherein weights of the machine-trained model have been trained by successively improving predictive accuracy of the machine-trained model, while rewarding the machine-trained model based on an extent to which the machine-trained model compresses instances of abstract token information.

11 . The method of claim 1 , wherein weights of the machine-trained model have been trained by:

using a teacher version of the machine-trained model to generate a ground-truth response for an input example, the teacher version of the machine-trained model having fixed weights during the training;

using a student version of the machine-trained model to generate a student-model response for the input example;

using the student version of the machine-trained model to generate student-model abstract token information, the student-model abstract token information expressing a current state of a dialogue history of a dialogue with fewer tokens than the dialogue history;

using the ground-truth response and the student-model response to generate prediction loss; and

using the student-model abstract token information to generate sparsity loss.

12 . The method of claim 11 ,

wherein the teacher version of the machine-trained model generates the ground-truth response independent of any instance of abstract token information, and

wherein the student version of the machine-trained model generates the student-model response based on prior-generated abstract token information.

13 . The method of claim 11 , wherein the student version of the machine-trained model generates the student-model abstract token information based on the input example, the student-model response, and prior-generated abstract token information.

14 . The method of claim 13 , wherein the student version of the machine-trained model also generates the student-model abstract token information based on machine-trained abstract-token-prompting information, the student version of the machine-trained model interpreting the machine-trained abstract-token-prompting information as a signal that the student version of the machine-trained model is to generate the student-model abstract token information.

15 . The method of claim 1 , wherein weights of the machine-trained model have been trained by:

using a teacher version of the machine-trained model to generate a ground-truth response for an input example, the teacher version of the machine-trained model having fixed weights during training, wherein the teacher version of the machine-trained model generates the ground-truth response independent of any instance of abstract token information,

using a student version of the machine-trained model to generate a student-model response for the input example, wherein the student version of the machine-trained model generates the student-model response based on prior-generated abstract token information,

using the student version of the machine-trained model to generate student-model abstract token information; and

using the ground-truth response, the student-model response, and the student-model abstract token information to generate loss.

16 . A computing system for interacting with a machine-trained model, comprising:

an instruction data store for storing computer-readable instructions;

a processing system for executing the computer-readable instructions in the data store, to perform operations including, in an inference stage:

using the machine-trained model to generate a response to a first prompt, the first prompt expressing input information for a current dialogue turn of a dialogue and first abstract token information generated in one or more prior dialogue turns,

each instance of abstract token information expressing a current state of a dialogue history of the dialogue with fewer tokens than the dialogue history,

wherein the machine-trained model includes weights that have been trained in a training stage by successively improving predictive accuracy of the machine-trained model, while rewarding the machine-trained model based on an extent to which the machine-trained model compresses instances of abstract token information,

wherein the weights have been trained in the training stage by:

using a teacher version of the machine-trained model to generate a ground-truth response for an input example, the teacher version of the machine-trained model having fixed weights during training, wherein the teacher version of the machine-trained model generates the ground-truth response independent of any instance of abstract token information,

using a student version of the machine-trained model to generate a student-model response for the input example, wherein the student version of the machine-trained model generates the student-model response based on prior-generated abstract token information,

using the student version of the machine-trained model to generate student-model abstract token information; and

using the ground-truth response, the student-model response, and the student-model abstract token information to generate loss.

17 . The computing system of claim 16 , wherein the operations further include, in the inference stage, mapping a second prompt to second abstract token information using the machine-trained model, for input to the machine-trained model in a subsequent dialogue turn.

18 . The computing system of claim 17 , wherein the second prompt also includes machine-trained abstract-token-prompting information, the machine-trained model interpreting the machine-trained abstract-token-prompting information as a signal that the machine-trained model is to generate the second abstract token information.

19 . A computer-readable storage medium for storing computer-readable instructions, a processing system executing the computer-readable instructions to perform operations, the operations comprising:

using a teacher version of a machine-trained model to generate a ground-truth response for an input example, the teacher version of the machine-trained model having fixed weights during training;

using a student version of the machine-trained model to generate a student-model response for the input example;

using the student version of the machine-trained model to generate student-model abstract token information;

using the ground-truth response and the student-model response to generate prediction loss;

using the student-model abstract token information to generate sparsity loss; and

using plural instances of the prediction loss and plural instances of the sparsity loss to update the student version of the machine-trained model,

wherein the teacher version of the machine-trained model generates the ground-truth response independent of any instance of abstract token information,

wherein the student version of the machine-trained model generates the student-model response based on prior-generated abstract token information, and

wherein the student version of the machine-trained model generates the student-model abstract token information based on the input example, the student-model response, the prior-generated abstract token information, and machine-trained abstract-token-prompting information.

Assignments (1)

Continuity (1)

References Cited (44)