Instruction-guided visual embeddings and feedback-based learning in large vision-language models

Patent metadata

Field	Value
Publication number	US-12411879-B2
Application number	US-202418924763-A
Country	US
Kind code	B2
Filing date	Oct 23, 2024
Priority date	Oct 24, 2023
Publication date	Sep 9, 2025
Grant date	Sep 9, 2025

Abstract

In an example, a method for fine-tuning a Large Visual Language Model (LVLM) includes providing visual queries, each of the visual queries comprises at least an image and a textual query related to the image; processing, by the LVLM, the visual queries to extract visual embeddings from the visual queries, wherein the LVLM comprises a Visual Language Model (VLM), a first Large Language Model (LLM), and a linear projection layer interconnecting the VLM and the LLM; for visual queries: i) generating, by the LVLM, a response to the corresponding visual query based on the corresponding visual embedding; ii) evaluating, by a second LLM, the generated response to verify that the generated response satisfies predefined criteria; and iii) providing, by the second LLM, a feedback to the LVLM, in response to the evaluating the generated response; and fine-tuning the LVLM using aggregated feedback provided by the second LLM for the visual queries.

First claim

What is claimed is:

1. A method for fine-tuning a Large Visual Language Model (LVLM), the method comprising:
providing a plurality of visual queries, wherein each of the plurality of visual queries comprises at least an image and a textual query related to the image;
processing, by the LVLM, the plurality of visual queries to extract one or more visual embeddings from each of the plurality of visual queries, wherein the LVLM comprises a Visual Language Model (VLM), a first Large Language Model (LLM), and a linear projection layer interconnecting the VLM and the LLM;
for each of the plurality of visual queries:
i) generating, by the LVLM, a response to the corresponding visual query based on the corresponding one or more visual embeddings;
ii) evaluating, by a second LLM, the generated response to verify that the generated response satisfies one or more predefined criteria; and
iii) providing, by the second LLM, a feedback to the LVLM, in response to the evaluating the generated response; and

fine-tuning the LVLM using aggregated feedback provided by the second LLM for the plurality of visual queries.

2. The method of claim 1 , wherein the one or more predefined criteria comprise at least one of helpfulness, honesty, and harmlessness.

3. The method of claim 1 , wherein the feedback comprises a Natural Language Feedback (NLF).

4. The method of claim 3 , wherein the feedback comprises at least a numerical score, critique feedback, and refinement feedback.

5. The method of claim 4 , wherein the refinement feedback suggests improvements or modifications to the generated response.

6. The method of claim 3 , further comprising training the LVLM using the NLF.

7. The method of claim 6 , wherein training the LVLM further comprises:
training the LVLM using the NLF incorporated into a conditional Reinforcement Learning (RL) algorithm.

8. A computing system for fine-tuning a Large Visual Language Model (LVLM), the computing system comprising:
processing circuitry in communication with storage media, the processing circuitry configured to execute a machine learning system comprising the LVLM, the processing circuitry configured to:
provide a plurality of visual queries, wherein each of the plurality of visual queries comprises at least an image and a textual query related to the image;
process, by the LVLM, the plurality of visual queries to extract one or more visual embeddings from each of the plurality of visual queries, wherein the LVLM comprises a Visual Language Model (VLM), a first Large Language Model (LLM), and a linear projection layer interconnecting the VLM and the LLM;

for each of the plurality of visual queries:
i) generate, by the LVLM, a response to the corresponding visual query based on the corresponding one or more visual embeddings;
ii) evaluate, by a second LLM, the generated response to verify that the generated response satisfies one or more predefined criteria; and
iii) provide, by the second LLM, a feedback to the LVLM, in response to the evaluating the generated response; and
fine-tune the LVLM using aggregated feedback provided by the second LLM for the plurality of visual queries.

9. The system of claim 8 , wherein the one or more predefined criteria comprise at least one of helpfulness, honesty, and harmlessness.

10. The system of claim 8 , wherein the feedback comprises a Natural Language Feedback (NLF).

11. The system of claim 10 , wherein the feedback comprises at least a numerical score, critique feedback, and refinement feedback.

12. The system of claim 11 , wherein the refinement feedback suggests improvements or modifications to the generated response.

13. The system of claim 10 , the processing circuitry further configured to:
train the LVLM using the NLF.

14. The system of claim 13 , wherein the processing circuitry configured to train the LVLM is further configured to:
train the LVLM using the NLF incorporated into a conditional Reinforcement Learning (RL) algorithm.

15. Non-transitory computer-readable storage media having instructions encoded thereon, the instructions configured to cause processing circuitry to:
provide a plurality of visual queries, wherein each of the plurality of visual queries comprises at least an image and a textual query related to the image;
process, by a Large Visual Language Model (LVLM), the plurality of visual queries to extract one or more visual embeddings from each of the plurality of visual queries, wherein the LVLM comprises a Visual Language Model (VLM), a first Large Language Model (LLM), and a linear projection layer interconnecting the VLM and the LLM;
for each of the plurality of visual queries:
i) generate, by the LVLM, a response to the corresponding visual query based on the corresponding one or more visual embeddings;
ii) evaluate, by a second LLM, the generated response to verify that the generated response satisfies one or more predefined criteria; and
iii) provide, by the second LLM, a feedback to the LVLM, in response to the evaluating the generated response; and

fine-tune the LVLM using aggregated feedback provided by the second LLM for the plurality of visual queries.

16. The storage media of claim 15 , wherein the one or more predefined criteria comprise at least one of helpfulness, honesty, and harmlessness.

17. The storage media of claim 15 , wherein the feedback comprises a Natural Language Feedback (NLF).

18. The storage media of claim 17 , wherein the feedback comprises at least a numerical score, critique feedback, and refinement feedback.

19. The storage media of claim 18 , wherein the refinement feedback suggests improvements or modifications to the generated response.

20. The storage media of claim 17 , the instructions further configured to cause processing circuitry to:
train the LVLM using the NLF.

Assignees

• Stanford Res Inst Int

Inventors

• Chen Yangyi
• Sikka Karan
• Cogswell Michael A
• Divakaran Ajay

Classifications

• G06F16/3344

  using natural language analysis · CPC title

• G06F16/532 Primary

  Query formulation, e.g. graphical querying · CPC title

• G06F16/338 Primary

  Presentation of query results · CPC title