DETAILED ACTION
This Office Action is responsive to amendments and arguments filed on July 1st, 2026.
Claims 1, 7 and 15 are amended. Claims 1-20 are pending and have been examined; hence, this action is made FINAL.
Any previous objections/rejections not mentioned in this Office Action have been withdrawn by the Examiner.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on October 23 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Response to Amendments and Arguments
Regarding rejections made under 35 U.S.C. 103, Applicant argues, "Krabach discloses in [0022] that the rubric 35 may measure model characteristics such as 'distinct features, style, persona, disposition, voice, identity, temperament, attitude, character traits, or other characteristics that the model 46 displays when generating responses or content' while [0024] discloses measuring model output characteristics such as 'cultural sensitivity, fairness, inclusivity, wittiness, assertiveness, patience, friendliness, formality, empathy, profanity, verbosity, tone, topical relevance, factuality, and creativity.' Accordingly, Applicant respectfully submits that the claims, as amended, are distinguished from the teachings of Krabach.
In addition, while Kamkar may disclose that 'S250 includes determining whether the initial model satisfies one or more constraints.' ([0073]), Applicant respectfully submits that such constraints are later specified to be 'fairness constraints' for 'sensitive attributes.' (Id.) Applicant respectfully submits that determining whether a model satisfies fairness constraints of sensitive attributes is distinct from a classification model to ['classify.. .a task into one or more task categories representing model functional capabilities'] as is claimed. Accordingly, Applicant respectfully submits that Kamkar fails to disclose all elements of the independent claims, as amended," (page 15 of Remarks).
Applicant’s arguments are moot, as new grounds of rejection are raised in view of previously included reference U.S. 2024/0135113 to Zorn et al. Further detail is provided below.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 5, 7-8, 13, 15 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication 2025/0077795 to Krabach et al. (hereinafter, "Krabach") in view of U.S. Patent Application Publication 2024/0135113 to Zorn et al. (hereinafter, "Zorn").
Regarding claims 1, 7 and 15, Krabach teaches a computer-implemented method, a computer program product, and a computer system comprising: training, using a database of tasks, a classifier model to classify a task into one or more task categories representing model functional capabilities, the training resulting in a trained classifier model (paragraph [0024], "Rubric classifier training module 32 is configured to train the rubric classifier 34 using a rubric database 30, which in the illustrated example comprises outputs 31a-d of a plurality of generative language models with one or more output characteristics evaluated by human or machine evaluators observing the outputs 31a-d or responses of each model to a variety of prompts, and then scoring or rating the performance of each generative language model for each output characteristic that is evaluated.");
generating a plurality of prompts (paragraph [0015], "At a high level, the generative model program 22 implements an interaction interface 26 by which a text input 42 is received, and passes the text input 42 to a prompt generator 38, which generates a prompt 40 based on the text input 42. The prompt 40 is input to a generative model 46, which in turn generates output 48 which can be passed to interaction interface 26.");
applying a first prompt in the plurality of prompts to a trained model, the trained model configured to perform tasks in a first task category of the one or more task categories, the trained model producing a first model output in response to the first prompt (paragraph [0050], "At step 202, the method 200 interfaces with a trained generative model that receives input of a prompt including natural language text input and, in response, generates an output that includes natural language text output.");
classifying, using the trained classifier model, the first model output into tasks in a second task category of the one or more task categories (paragraph [0051], "At step 204, the method 200 includes monitoring compliance of the generative language model with the rubric, by feeding the output of the generative language model to a rubric classifier configured to generate a predicted classification for an output characteristic in the rubric. The predicted classification may be a numerical classification or a qualitative classification."); and
adjusting, responsive to determining the second task category is the undesired task category, the trained model, the adjusting altering a functional capability of the trained model to limit the trained model to perform the tasks in the first task category (paragraph [0051], "The generative language model may be intermittently updated over a time period, and compliance may be monitored by feeding a plurality of outputs of the generative language model to the rubric classifier at a series of points in time during the time period, to thereby generate a time-series of predicted classifications for the output characteristic in the rubric.").
While Krabach teaches a method for evaluating and adjusting a machine learning model, the functional capability of the subject model is not taught as the feature under test, and thus, Zorn is introduced.
Zorn teaches a method for discovering machine learning model capabilities that includes determining that the second task category is different than the first task category and thereby an undesired task category (paragraph [0027], "The method 300 may be performed by the capability extraction system 120 of FIG. 1, in order to output the capability indication 121 of FIG. 1. The method 300 includes determining that the capability extraction system is to estimate or determine whether the model has the one or more capabilities (act 301). Then, the capability extraction system performs one or more capability extraction stages (act 302) to determine whether the model has the capability or capabilities (act 303)," and paragraph [0033], "On the other hand, in an 'indirect' capability extraction stage, the output generated by the language model is either not natural language at all, or else is natural language that is not semantically responsive to the natural language input 401 in the language model input 400. As an example, the natural language input might be a request to generate code, and the response could be the generation of the code. Code is not natural language. Thus, this is an indirect capability extraction stage. As another example, the natural language input might be a request to perform a task, and the output is data that is not structured in natural language form.").
Krabach and Zorn are considered analogous because they are each concerned with assessing and training machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have combined Zorn’s capability extraction system with Krabach’s rubric classifier for the purpose of effectively guiding model training. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Regarding claims 5, 13 and 19 Krabach further teaches generating the plurality of prompts comprises adjusting an initial prompt producing an initial model output, the adjusting generating an adjusted prompt producing an adjusted model output, the initial model output having an initial correctness lower than a correctness threshold, the adjusted model output having an adjusted correctness higher than the initial correctness (paragraph [0047], "As shown in FIG. 3, the output 48 of the generative language model 46 includes exchanges in which John Smith anxiously mentioned his history of ACL tear as he asked if running was a safe workout activity for him. This output 48 is inputted into a rubric classifier 34, which outputs a predicted classification 36 for the output 48 indicating a friendliness score of 4 out of 5. The prompt generator 38 receives the predicted classification 36 as input, determines that the friendliness score of 4 is less than the target friendliness score of 5, and generates a prompt context 44 which suggests, “use more personalized and warm responses, use the user's name in the conversation, express more empathy, add a bit of positive emotional tone” so that the friendliness score of subsequent responses can be raised to a 5.").
Regarding claim 8, Krabach further teaches the stored program instructions are stored in a computer readable storage device in a data processing system, and wherein the stored program instructions are transferred over a network from a remote data processing system (paragraph [0045], "The client computing device 62 may be responsible for communicating over a computer network between the user operating the client computing device 62 and the server computing device 60 which executes the generative model program 22 and contains the rubric classifier 34 and the generative language model 46, via an application programming interface (API) 66 of the generative model program 22. The client computing device 62 may take the form of a personal computer, laptop, tablet, smartphone, smart speaker, etc. The same processes described above with reference to FIG. 1A may be performed, except in this case the natural language text input 42 and output 48 may be communicated between the server computing device 60 and the client computing device via a computer network such as the Internet.").
Claims 2, 10 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Krabach and Zorn as applied to claims 1, 7 and 15 above, and further in view of U.S. Patent Application Publication 2024/0354503 to Baruch et al. (hereinafter, "Baruch").
Regarding claims 2, 10 and 16, the combination of Krabach and Zorn does not teach “generating the plurality of prompts comprises training a reinforcement learning agent to reward a prompt invoking generation of a novel task higher than a prompt invoking generation of a non-novel task, the training resulting in a trained reinforcement learning agent,” and thus, Baruch is introduced.
Baruch teaches generating the plurality of prompts comprises training a reinforcement learning agent to reward a prompt invoking generation of a novel task higher than a prompt invoking generation of a non-novel task, the training resulting in a trained reinforcement learning agent (paragraph [0292], "In some implementations, feedback processor 1210 includes a reinforcement learning component such as a reinforcement learning model that machine-learns a reward function based on feedback associated with prompt-output pairs. For example, given a prompt-output pair 1208, feedback processor 1210 receives or identifies feedback that pertains to the prompt-output pair 1208. The feedback can include pre-distribution feedback and/or post-distribution feedback received from one or more other components of the thought starter generation system. The feedback processor 1210 applies the reward function to the received or identified feedback to generate a reward score for the corresponding prompt-output pair based on the feedback associated with the prompt-output pair. The reward scores are incorporated into the prompt-feedback pairs 1212 and/or output-feedback pairs 1214, which are then used to train or fine tune the generative model 1206 using, for example, supervised or semi-supervised machine learning.").
Krabach, Zorn and Baruch are considered analogous because they are each concerned with training machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Krabach and Zorn with the teachings of Baruch for the purpose of effectively guiding model training. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Claims 3-4, 6, 11-12, 14, 17-18 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Krabach and Zorn as applied to claims 1, 5, 7, 13, 15 and 19 above, and further in view of U.S. Patent Application Publication 2025/0111147 to Pryzant et al. (hereinafter, "Pryzant").
Regarding claims 3, 11 and 17, the combination of Krabach and Zorn does not teach “generating the plurality of prompts comprises using the trained reinforcement learning agent to reward a derived prompt, the derived prompt derived from an existing prompt,” and thus, Pryzant is introduced.
Pryzant teaches generating the plurality of prompts comprises using the trained reinforcement learning agent to reward a derived prompt, the derived prompt derived from an existing prompt (paragraph [0036], "Second, the textual gradients g.sub.1-g.sub.x 230a-230m are provided to another LLM prompt, in this case, editing prompt δ 235, which instructs the LLM to edit the current prompt P 205 in order to fix the problems described by the textual gradients g1-gx 230a-230m. In this way, the LLMs are engaged in a recursive feedback loop.").
Krabach, Zorn and Pryzant are considered analogous because they are each concerned with training machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Krabach and Zorn with the teachings of Pryzant for the purpose of improving training results. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Regarding claims 4, 12 and 18, the combination of Krabach and Zorn do not teach “generating the plurality of prompts comprises prompting the trained model to generate a generated prompt, the generated prompt resulting in a desired output of the trained model,” however, Pryzant teaches generating the plurality of prompts comprises prompting the trained model to generate a generated prompt, the generated prompt resulting in a desired output of the trained model (paragraph [0036], "Third, additional candidate prompts are generated by running the existing candidate prompts or optimized prompts P′11-P′mq 240 through a paraphrasing prompt mc 245 or an LLM referred to as LLMmc, to explore the local Monte Carlo search space around the new prompt candidates. This prompt 245 asks the LLM to generate new candidate prompts or paraphrased prompts P”111-P”mqs 250, which are worded differently but semantically similar to their inputs (i.e., optimized prompts P′11-P′mq 240).").
Krabach, Zorn and Pryzant are considered analogous because they are each concerned with training machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Krabach and Zorn with the teachings of Pryzant for the purpose of improving training results. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Regarding claims 6, 14 and 20, the combination of Krabach and Zorn does not teach “the adjusted prompt has a semantic meaning above a semantic meaning threshold,” however, Pryzant teaches the adjusted prompt has a semantic meaning above a semantic meaning threshold (paragraph [0073], "An example paraphrasing prompt, which was used for each of the examples shown in FIGS. 4A-4D, may include prompt language such as: “Generate a variation of the following instruction while keeping the semantic meaning. Input: {prompt_instruction}. Output: ______.”").
Krabach, Zorn and Pryzant are considered analogous because they are each concerned with training machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Krabach and Zorn with the teachings of Pryzant for the purpose of improving training results. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Krabach and Zorn as applied to claim 7 above, and further in view of U.S. Patent 11,775,867 to Jamei (hereinafter, "Jamei").
Regarding claim 9, Krabach further teaches the stored program instructions are stored in a computer readable storage device in a server data processing system, and wherein the stored program instructions are downloaded in response to a request over a network to a remote data processing system for use in a computer readable storage device associated with the remote data processing system (paragraph [0045], "The client computing device 62 may be responsible for communicating over a computer network between the user operating the client computing device 62 and the server computing device 60 which executes the generative model program 22 and contains the rubric classifier 34 and the generative language model 46, via an application programming interface (API) 66 of the generative model program 22. The client computing device 62 may take the form of a personal computer, laptop, tablet, smartphone, smart speaker, etc. The same processes described above with reference to FIG. 1A may be performed, except in this case the natural language text input 42 and output 48 may be communicated between the server computing device 60 and the client computing device via a computer network such as the Internet.").
The combination of Krabach and Zorn does not teach “program instructions to meter use of the program instructions associated with the request; and program instructions to generate an invoice based on the metered use,” and thus, Jamei is introduced.
Jamei teaches program instructions to meter use of the program instructions associated with the request; and program instructions to generate an invoice based on the metered use (column 7, lines 12-19, "Each testing container (or “worker”) logs the resources that it uses to evaluate a submitted model. This includes required RAM, CPU memory, processing time, data storage, network traffic to transfer the data, etc. The log data is aggregated by user ID and used for billing and enforcement of limits or quotas set on resource use. This allows for a system in which each user pays for the amount of services that they use as part of the model evaluation processes.").
Krabach, Zorn and Jamei are considered analogous because they are each concerned with assessing machine learning models. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Krabach and Zorn with the teachings of Jamei for the purpose of expanding monetization opportunities for model training. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
U.S. Patent 8,762,299 to Breckenridge et al.
U.S. Patent 11,494,689 to Efstathiou et al.
U.S. Patent 11,983,238 to Nagalapatti et al.
U.S. Patent 12,361,215 to Wei et al.
U.S. Patent Application Publication 2024/0428937 to Natarajan et al.
U.S. Patent Application Publication 2023/0267307 to Wang et al.
U.S. Patent Application Publication 2024/0119361 to Yin et al.
U.S. Patent Application Publication 2025/0124300 to Maia et al.
International Publication WO 2018/153806 to Gendron-Bellemare et al.
International Publication WO 2020/191057 to Kamkar et al.
U.S. Patent Application Publication 2019/0095557 to Sehgal et al.
"Coauthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities” by Lee et al.
“FairRover: Explorative Model Building for Fair and Responsible Machine Learning” by Zhang et al.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SEAN T SMITH whose telephone number is (571)272-6643. The examiner can normally be reached Monday - Friday 8:00am - 5:00pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, PIERRE-LOUIS DESIR can be reached at (571) 272-7799. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/SEAN THOMAS SMITH/Examiner, Art Unit 2659
/BHAVESH M MEHTA/Supervisory Patent Examiner, Art Unit 2656
Prosecution Insights