Back to results
US 12,754,959 B1Granted
On-device agentic validation via personalized model ensemble
AssigneeU.S. Bank National Association
InventorRahul Sharma
Life of this patent
Estimating the term…
Abstract
An agentic recommendation is generated in response to a query from an agentic model executing in a user device. Further in response to the query, a set of personalized recommendations is generated from a set of personalized models forming an ensemble of models executing in the user device. In an output corresponding to the query, the agentic recommendation is replaced with at least one personalized recommendation from the set of personalized recommendations.
Claims18
A method, comprising:
- generating, responsive to a query, from an agentic model executing in a user device, a multistep plan corresponding to the query;
- executing, in the agentic model, a first step from the multistep plan using a first agentic prompt corresponding to the first step, the executing outputting a first agentic recommendation;
- generating, from a first set of personalized models in an ensemble of models executing in the user device, and responsive to the first step, a first set of personalized recommendations;
- replacing, in an output corresponding to the first step, the first agentic recommendation with a first personalized recommendation from the first set of personalized recommendations;
- executing, in the agentic model, a second step from the multistep plan using a second agentic prompt corresponding to the second step, the executing outputting a second agentic recommendation;
- computing a cumulative confidence value as a product of confidence values of outputs chosen at a completed subset of steps from the set of steps in the multistep plan, wherein the completed subset of steps comprises those steps of the multistep plan for which at least one output has been generated;
- using a combination of the agentic model and the ensemble of models for processing an incomplete subset of steps from the set of steps in the multistep plan, wherein the incomplete subset of steps comprises those steps of the multistep plan for which no output has been generated, the using being responsive to the cumulative confidence value being more than a critical threshold for cumulative confidence but less than a moderate threshold for cumulative confidence;
- and
- outputting, responsive to the query, a final recommendation using the first personalized recommendation and the second agentic recommendation.
6 more elements
A computer program product comprising a set of one or more non-transitory computer readable storage media, and program instructions collectively stored on the one or more non-transitory computer readable storage media, the program instructions executable by a set of one or more processors to cause the set of one or more processors to perform operations comprising:
- generating, responsive to a query, from an agentic model executing in a user device, a multistep plan corresponding to the query;
- executing, in the agentic model, a first step from the multistep plan using a first agentic prompt corresponding to the first step, the executing outputting a first agentic recommendation;
- generating, from a first set of personalized models in an ensemble of models executing in the user device, and responsive to the first step, a first set of personalized recommendations;
- replacing, in an output corresponding to the first step, the first agentic recommendation with a first personalized recommendation from the first set of personalized recommendations;
- executing, in the agentic model, a second step from the multistep plan using a second agentic prompt corresponding to the second step, the executing outputting a second agentic recommendation;
- computing a cumulative confidence value as a product of confidence values of outputs chosen at a completed subset of steps from the set of steps in the multistep plan, wherein the completed subset of steps comprises those steps of the multistep plan for which at least one output has been generated;
- using a combination of the agentic model and the ensemble of models for processing an incomplete subset of steps from the set of steps in the multistep plan, wherein the incomplete subset of steps comprises those steps of the multistep plan for which no output has been generated, the using being responsive to the cumulative confidence value being more than a critical threshold for cumulative confidence but less than a moderate threshold for cumulative confidence;
- and
- outputting, responsive to the query, a final recommendation using the first personalized recommendation and the second agentic recommendation.
6 more elements
A system comprising a set of one or more processors and one or more non-transitory computer readable storage media, and program instructions collectively stored on the one or more non-transitory computer readable storage media, the program instructions executable by the set of one or more processors to cause the processor to perform operations comprising:
- generating, responsive to a query, from an agentic model executing in a user device, a multistep plan corresponding to the query;
- executing, in the agentic model, a first step from the multistep plan using a first agentic prompt corresponding to the first step, the executing outputting a first agentic recommendation;
- generating, from a first set of personalized models in an ensemble of models executing in the user device, and responsive to the first step, a first set of personalized recommendations;
- replacing, in an output corresponding to the first step, the first agentic recommendation with a first personalized recommendation from the first set of personalized recommendations;
- executing, in the agentic model, a second step from the multistep plan using a second agentic prompt corresponding to the second step, the executing outputting a second agentic recommendation;
- computing a cumulative confidence value as a product of confidence values of outputs chosen at a completed subset of steps from the set of steps in the multistep plan, wherein the completed subset of steps comprises those steps of the multistep plan for which at least one output has been generated;
- using a combination of the agentic model and the ensemble of models for processing an incomplete subset of steps from the set of steps in the multistep plan, wherein the incomplete subset of steps comprises those steps of the multistep plan for which no output has been generated, the using being responsive to the cumulative confidence value being more than a critical threshold for cumulative confidence but less than a moderate threshold for cumulative confidence;
- and
- outputting, responsive to the query, a final recommendation using the first personalized recommendation and the second agentic recommendation.
6 more elements
Dependencies are inferred from the claim text.