IP Library Granted Patent US 12,743,259
Granted Patent B2
US 12,743,259 · App. 19/532,268 · Granted Sep 22, 2026

Method for group relative policy optimization reinforcement learning on structured data for code generation

Inventors: Joseph Lo (Newark, NJ); German Soto Sanchez (Newark, NJ); Vijay Mayadas (Newark, NJ); Richard Chern (Newark, NJ)
Assignee: Broadridge Financial Solutions, Inc.
G06F8/35G06N20/00
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 12,743,259
App. No.
19/532,268
Granted
Sep 22, 2026
Kind
B2
Abstract

In some embodiments, the present disclosure provides an exemplary method that may include steps of integrating received structured data and received unstructured data from at least one enterprise system for a plurality of data outputs; generating one or more smart prompts; retrieving documentation relevant to a particular workflow and appending the particular workflow to the one or more smart prompts; utilizing a large language model to produce a plurality of candidate outputs; scoring each candidate output of the plurality of candidate outputs; calculating an average score across the plurality of candidate outputs; selecting at least one candidate output; and updating a plurality of parameters associated with a large language model based on the at least one selected candidate output.

Claims (48)

1 . A computer-implemented method comprising:

integrating, by a processor, received structured data and received unstructured data from at least one enterprise system for a plurality of data outputs;

generating, by the processor, one or more smart prompts based on an alignment of one or more data outputs of the plurality of data outputs;

retrieving, by the processor, documentation relevant to a particular workflow and appending the particular workflow to the one or more smart prompts;

utilizing, by the processor, a large language model to produce a plurality of candidate outputs based on the one or more smart prompts;

scoring, by the processor, each candidate output of the plurality of candidate outputs using a language model scoring function and at least one domain-specific scoring function,

wherein the domain-specific scoring function provides additional weighted values to one or more candidate outputs;

calculating, by the processor, an average score across the plurality of candidate outputs;

selecting, by the processor, at least one candidate output that exceeds a predetermined threshold, the predetermined threshold associated with the average score across the plurality of candidate outputs; and

updating, by the processor, a plurality of parameters associated with the large language model based on the at least one selected candidate output.

2 . The computer-implemented method of claim 1 , wherein the structured data comprises transaction data and the unstructured data comprises associated data logs from at least one mainframe system.

3 . The computer-implemented method of claim 1 , wherein generating the one or more smart prompts comprises aligning transaction data outputs associated with a production of the data outputs.

4 . The computer-implemented method of claim 1 , wherein retrieving documentation relevant to the particular workflow comprises extracting predetermined passages from system manuals, regulatory guidelines, and architectural documentation.

5 . The computer-implemented method of claim 1 , wherein the large language model is trained using group relative policy optimization.

6 . The computer-implemented method of claim 1 , wherein scoring each candidate output comprises applying a domain-specific validation algorithm to generate numeric scores quantifying compliance with predetermined domain criteria.

7 . The computer-implemented method of claim 1 , wherein the domain-specific scoring function compares candidate outputs against known regulatory constraints and historical transaction records.

8 . The computer-implemented method of claim 1 , wherein updating the plurality of parameters associated with the large language model comprises selective reinforcement based on candidate outputs that exceed the predetermined threshold.

9 . The computer-implemented method of claim 1 , further comprising automatically generating code associated with the at least one selected candidate output.

10 . The computer-implemented method of claim 1 , wherein the method is performed in real-time using streaming input data to dynamically construct smart prompts.

11 . The computer-implemented method of claim 1 , further comprising parsing identified processing steps to generate structured data pairs representing temporal or logical relationships between the identified processing steps.

12 . The computer-implemented method of claim 1 , further comprising storing final outputs with associated log references establishing traceability to source data.

13 . The computer-implemented method of claim 1 , wherein the method further comprises utilizing auxiliary domain-specific scoring algorithms to fine-tune or train the large language model in a continual manner.

14 . The computer-implemented method of claim 1 , wherein the method further comprises communicating with a plurality of models and modules to automatically generate code associated with the plurality of candidate outputs.

15 . The computer-implemented method of claim 1 , wherein the method further comprises receiving real-time streaming data associated with financial events, wherein newly received data triggers repeated execution of one or more workflows.

16 . A system comprising: one or more processors; and a memory in communication with the one or more processors and storing instructions that, when executed by the one or more processors, cause the one or more processors to:

integrate received structured data and received unstructured data from at least one enterprise system for a plurality of data outputs;

generate one or more smart prompts based on an alignment of one or more data outputs of the plurality of data outputs;

retrieve documentation relevant to a particular workflow and appending the particular workflow to the one or more smart prompts;

utilize a large language model to produce a plurality of candidate outputs based on the one or more smart prompts;

score each candidate output of the plurality of candidate outputs using a language model scoring function and at least one domain-specific scoring function,

wherein the domain-specific scoring function provides additional weighted values to one or more candidate outputs;

calculate an average score across the plurality of candidate outputs;

select at least one candidate output that exceeds a predetermined threshold, the predetermined threshold associated with the average score across the plurality of candidate outputs; and

update a plurality of parameters associated with the large language model based on the at least one selected candidate output.

17 . The system of claim 16 , wherein the structured data comprises transaction data and the unstructured data comprises associated data logs from at least one mainframe system.

18 . The system of claim 16 , wherein the large language model is trained using group relative policy optimization.

19 . The system of claim 16 , wherein the domain-specific scoring function compares candidate outputs against known regulatory constraints and historical transaction records.

20 . A computer-implemented method comprising:

integrating, by a processor, received structured data and received unstructured data from at least one enterprise system for a plurality of data outputs;

generating, by the processor, one or more smart prompts based on an alignment of one or more data outputs of the plurality of data outputs;

retrieving, by the processor, documentation relevant to a particular workflow and appending the particular workflow to the one or more smart prompts;

utilizing, by the processor, a large language model to produce a plurality of candidate outputs based on the one or more smart prompts;

scoring, by the processor, each candidate output of the plurality of candidate outputs using a language model scoring function and at least one domain-specific scoring function,

wherein the domain-specific scoring function provides additional weighted values to one or more candidate outputs;

calculating, by the processor, an average score across the plurality of candidate outputs;

selecting, by the processor, at least one candidate output that exceeds a predetermined threshold, the predetermined threshold associated with the average score across the plurality of candidate outputs;

automatically generating, by the processor, code associated with the at least one selected candidate output; and

updating, by the processor, a plurality of parameters associated with the large language model based on the at least one selected candidate output.

Continuity (2)
Provisional Application 63754962 · Feb 6, 2025
Related Publication 20260227970A1 · Aug 6, 2026
References Cited (29)
US 10331659B2 · Ahuja · 2019 [cited by examiner]
US 11726750B1 · Arcadinho et al. · 2023 [cited by applicant]
US 12222898B1 · Madan · 2025 [cited by examiner]
US 12614026B1 · Leonard et al. · 2026 [cited by applicant]
US 20200293562A1 · Pasternack · 2020 [cited by examiner]
US 20230092969A1 · Chen et al. · 2023 [cited by applicant]
US 20240062120A1 · He · 2024 [cited by examiner]
US 20240086767A1 · Xu · 2024 [cited by examiner]
US 20240095584A1 · Xu · 2024 [cited by examiner]
US 20240202221A1 · Juban et al. · 2024 [cited by applicant]
US 20240345551A1 · Selvaraj et al. · 2024 [cited by applicant]
US 20240394545A1 · Eisenschlos · 2024 [cited by examiner]
US 20250021650A1 · Corlatescu · 2025 [cited by examiner]
US 20250094473A1 · Wilczynski · 2025 [cited by examiner]
US 20250117417A1 · Aghajanyan · 2025 [cited by examiner]
US 20250131028A1 · Siebel · 2025 [cited by examiner]
US 20250131044A1 · Ken-Kwofie · 2025 [cited by examiner]
US 20250181619A1 · Jain · 2025 [cited by examiner]
US 20250188823A1 · Rathore · 2025 [cited by examiner]
US 20250245665A1 · Tholar · 2025 [cited by examiner]
US 20250252445A1 · Gaddam · 2025 [cited by examiner]
US 20250315683A1 · Feehan · 2025 [cited by examiner]
US 20250348728A1 · Bang · 2025 [cited by examiner]
US 20250356187A1 · Papancea · 2025 [cited by examiner]
US 20250372219A1 · Boussina · 2025 [cited by examiner]
US 20260064731A1 · Rezaeian · 2026 [cited by examiner]
US 20260087010A1 · Agarwal · 2026 [cited by examiner]
KR 102895547B1 · 2025 [cited by examiner]
International Search Report and Written Opinion from International Application No. PCT/US2026/014288 dated May 15, 2026. [cited by applicant]