IP Library › Granted Patent US 12682180
Granted Patent B1
US 12682180 · App. 19/195,387 · Granted Jul 14, 2026

Large language model performance evaluation and modification

Inventors: Xiang Gao (Mountain View, CA); Avinash Baidya (Mountain View, CA); Kamalika Das (Mountain View, CA)
Assignee: INTUIT INC.
G06F40/35G06F40/30G06N20/00H04L51/02H04L51/216
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12682180
App. No.
19/195,387
Granted
Jul 14, 2026
Kind
B1
Abstract

At least one processor may receive conversation data comprising a user input and a response to the user input generated by a large language model (LLM). The at least one processor may detect a behavior gap between the conversation data and at least one test conversation data comprising human-to-human communication samples. Based on the behavior gap, the at least one processor may retrain the LLM and/or provide an alternative response workflow for responding to the user input.

Claims (39)

1 . A method comprising:

receiving, by at least one processor, conversation data comprising a user input included in a prompt to a large language model (LLM) and a response to the user input generated by the LLM;

detecting, by the at least one processor, a behavior gap between the conversation data and at least one test conversation data comprising human-to-human communication samples,

the detecting comprising detecting:

at least one overused dialog act indicated by at least one of a misalignment of LLM dialog acts with dialog acts found in the at least one test conversation data and a use of more dialog acts per turn than found in the at least one test conversation data,

at least one overused tool use indicated by at least one of a different LLM tool selection than selections found in the at least one test conversation data and a more frequent LLM tool use than uses found in the at least one test conversation data, and

at least one overused external knowledge use indicated by a more frequent LLM direct copying than copying found in the at least one test conversation data and a lower LLM compression ratio than a compression ratio found in the at least one test conversation data;

determining, by the at least one processor, an alternative response workflow configured to generate a second response to the user input having a second behavior gap that is less than the behavior gap, the determining comprising configuring the LLM to produce the second response without incorporating the at least one overused element; and

causing, by the at least one processor, generation of the second response to the user input by the LLM and presentation of the second response in a user interface in response to the prompt to the LLM.

2 . The method of claim 1 , wherein the detecting the at least one dialog act comprises prompting a test LLM with a dialog act system prompt and the conversation data and receiving the at least one dialog act in a response generated by the test LLM.

3 . The method of claim 1 , wherein the detecting at least one tool use comprises prompting a test LLM with a tool use system prompt and the conversation data and receiving an indication of the at least one tool use in a response generated by the test LLM.

4 . The method of claim 1 , wherein the detecting at least one external knowledge use comprises detecting copying of external knowledge into the response to the user input.

5 . The method of claim 1 , wherein the detecting at least one external knowledge use comprises determining a compression ratio of at least a portion of the response to the user input over referenced external knowledge.

6 . The method of claim 1 , wherein the determining comprises:

receiving, by the at least one processor, second conversation data comprising the user input and the second response to the user input generated by a second LLM;

detecting, by the at least one processor, the second behavior gap between the second conversation data and at the least one test conversation data comprising human-to-human communication samples;

determining, by the at least one processor, that the second behavior gap is less than the behavior gap; and

causing, by the at least one processor, output of the second response to the user input by the second LLM.

7 . The method of claim 1 , wherein the alternative response workflow comprises injecting at least one known human dialog act into the second response.

8 . A system comprising:

at least one processor; and

at least one non-transitory computer-readable medium storing instructions that, when executed by the at least one processor, cause the at least one processor to perform processing comprising:

receiving conversation data comprising a user input included in a prompt to a large language model (LLM) and a response to the user input generated by the LLM;

detecting a behavior gap between the conversation data and at least one test conversation data comprising human-to-human communication samples, the detecting comprising detecting:

at least one overused dialog act indicated by at least one of a misalignment of LLM dialog acts with dialog acts found in the at least one test conversation data and a use of more dialog acts per turn than found in the at least one test conversation data,

at least one overused tool use indicated by at least one of a different LLM tool selection than selections found in the at least one test conversation data and a more frequent LLM tool use than uses found in the at least one test conversation data, and

at least one overused external knowledge use indicated by a more frequent LLM direct copying than copying found in the at least one test conversation data and a lower LLM compression ratio than a compression ratio found in the at least one test conversation data;

determining an alternative response workflow configured to generate a second response to the user input having a second behavior gap that is less than the behavior gap, the determining comprising configuring the LLM to produce the second response without incorporating the at least one overused element; and

causing generation of the second response to the user input by the LLM and presentation of the second response in a user interface in response to the prompt to the LLM.

9 . The system of claim 8 , wherein the detecting the at least one dialog act comprises prompting a test LLM with a dialog act system prompt and the conversation data and receiving the at least one dialog act in a response generated by the test LLM.

10 . The system of claim 8 , wherein the detecting at least one tool use comprises prompting a test LLM with a tool use system prompt and the conversation data and receiving an indication of the at least one tool use in a response generated by the test LLM.

11 . The system of claim 8 , wherein the detecting at least one external knowledge use comprises detecting copying of external knowledge into the response to the user input.

12 . The system of claim 8 , wherein the detecting at least one external knowledge use comprises determining a compression ratio of at least a portion of the response to the user input over referenced external knowledge.

13 . The system of claim 8 , wherein the determining comprises:

receiving second conversation data comprising the user input and the second response to the user input generated by a second LLM;

detecting the second behavior gap between the second conversation data and at the least one test conversation data comprising human-to-human communication samples;

determining that the second behavior gap is less than the behavior gap; and

causing output of the second response to the user input by the second LLM.

14 . The system of claim 8 , wherein the alternative response workflow comprises injecting at least one known human dialog act into the second response.