Systems and methods for editing a large language model

Patent metadata

Field	Value
Publication number	US-12530541-B2
Application number	US-202418428530-A
Country	US
Kind code	B2
Filing date	Jan 31, 2024
Priority date	Oct 13, 2023
Publication date	Jan 20, 2026
Grant date	Jan 20, 2026

Abstract

Systems and methods for editing a large language model are provided. The large language model generates a sequence of tokens, a first probability of a pre-edit output based on the sequence of tokens, and a second probability of a target output based on the sequence of tokens. A loss function is provided based on the first probability and the second probability. A plurality of gradients of the large language model with respect to the loss function is computed. An edit location of the large language model is determined based on the plurality of gradients. The large language model is edited by editing weights at the edit location of the large language model, such that the updated large language model generates the target output for an input including the sequence of words.

First claim

What is claimed is:

1 . A method of editing a large language model, the method comprising:
receiving, via a data interface, a sequence of words, a target output, and a pre-edit output; generating, using the large language model, a sequence of tokens of the sequence of words; generating, using the large language model, a first probability of the pre-edit output based on the sequence of tokens; generating, using the large language model, a second probability of the target output based on the sequence of tokens; providing a loss function based on the first probability of the pre-edit output and the second probability of the target output; computing a plurality of gradients of the large language model with respect to the loss function; determining an edit location of the large language model based on the plurality of gradients; and editing the large language model by editing weights at the edit location of the large language model, such that the updated large language model generates the target output for an input including the sequence of words.

2 . The method of claim 1 , wherein the sequence of words is associated with a non-binary proposition.

3 . The method of claim 1 , wherein the sequence of words is associated with a Boolean classification prompt, and
wherein the edit location includes an edit token location associated with a subject of the Boolean classification prompt.

4 . The method of claim 3 , wherein the subject is associated with a plurality of tokens including a last token, and
wherein the edit token location is associated with a token of the plurality of tokens before the last token.

5 . The method of claim 1 , wherein the computing the plurality of gradients of the large language model with respect to the loss function includes:
computing the plurality of gradients of the large language model with respect to the loss function over a group of tokens from the sequence of tokens and a first group of layers of the large language model.

6 . The method of claim 5 , wherein the determining the edit location of the large language model based on the plurality of gradients includes:
determining an edit layer location by selecting from a second group of layers of the large language model based on the plurality of gradients associated with the second group of layers.

7 . The method of claim 1 , wherein the editing the large language includes:
editing the weights at the edit location without using a subject label.

8 . A system for language model editing, the system comprising:
a memory that stores a large language model and a plurality of processor-executable instructions; a communication interface that receives a sequence of words, a target output, and a pre-edit output; one or more hardware processors that read and execute the plurality of processor-executable instructions from the memory to perform operations comprising:
generating, using the large language model, a sequence of tokens of the sequence of words;
generating, using the large language model, a first probability of the pre-edit output based on the sequence of tokens;
generating, using the large language model, a second probability of the target output based on the sequence of tokens;
providing a loss function based on the first probability of the pre-edit output and the second probability of the target output;
computing a plurality of gradients of the large language model with respect to the loss function;
determining an edit location of the large language model based on the plurality of gradients; and
editing the large language model by editing weights at the edit location of the large language model, such that the updated large language model generates the target output for an input including the sequence of words.

9 . The system of claim 8 , wherein the sequence of words is associated with a non-binary proposition.

10 . The system of claim 8 , wherein the sequence of words is associated with a Boolean classification prompt, and
wherein the edit location includes an edit token location associated with a subject of the Boolean classification prompt.

11 . The system of claim 10 , wherein the subject is associated with a plurality of tokens including a last token, and
wherein the edit token location is associated with a token of the plurality of tokens before the last token.

12 . The system of claim 10 , wherein the computing the plurality of gradients of the large language model with respect to the loss function includes:
computing the plurality of gradients of the large language model with respect to the loss function over a group of tokens from the sequence of tokens and a first group of layers of the large language model.

13 . The system of claim 12 , wherein the determining the edit location of the large language model based on the plurality of gradients includes:
determining an edit layer location by selecting from a second group of layers of the large language model based on the plurality of gradients associated with the second group of layers.

14 . The system of claim 8 , wherein the editing the weights at the edit location of the large language includes:
editing the weights at the edit location without using a subject label.

15 . A non-transitory machine-readable medium comprising a plurality of machine-executable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform operations comprising:
receiving, via a data interface, a sequence of words, a target output, and a pre-edit output; generating, using the large language model, a sequence of tokens of the sequence of words;
generating, using the large language model, a first probability of the pre-edit output based on the sequence of tokens;
generating, using the large language model, a second probability of the target output based on the sequence of tokens;
providing a loss function based on the first probability of the pre-edit output and the second probability of the target output;
computing a plurality of gradients of the large language model with respect to the loss function;
determining an edit location of the large language model based on the plurality of gradients; and
editing the large language model by editing weights at the edit location of the large language model, such that the updated large language model generates the target output for an input including the sequence of words.

16 . The non-transitory machine-readable medium of claim 15 , wherein the sequence of words is associated with a non-binary proposition.

17 . The non-transitory machine-readable medium of claim 15 , wherein the sequence of words is associated with a Boolean classification prompt, and
wherein the edit location includes an edit token location associated with a subject of the Boolean classification prompt.

18 . The non-transitory machine-readable medium of claim 17 , wherein the subject is associated with a plurality of tokens including a last token, and
wherein the edit token location is associated with a token of the plurality of tokens before the last token.

19 . The non-transitory machine-readable medium of claim 15 , wherein the computing the plurality of gradients of the large language model with respect to the loss function includes:
computing the plurality of gradients of the large language model with respect to the loss function over a group of tokens from the sequence of tokens and a first group of layers of the large language model.

Assignees

• Salesforce Inc

Inventors

• Feigenbaum Itai Izhak
• Arpit Devansh
• Heinecke Shelby
• Niebles Duque Juan Carlos
• Yao Weiran
• Wang Huan
• Xiong Caiming
• Savarese Silvio

Classifications

• G06N7/01 Primary

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06F40/166

  Editing, e.g. inserting or deleting · CPC title

• G06F40/284

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F40/40 Primary

  Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30) · CPC title