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 .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-5, 7-12, and 14-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Taking independent claim 1 as representative, the claims recite, in part:
A method for selecting a prompt to input to a large language model (LLM), the method comprising:
determining, by a predictive model and for a first set of data to be the subject of a prompt of a pre-defined set of prompts, a score for each prompt of the pre-defined set of prompts,
wherein the score for each prompt is based on a predicted probability of receiving a defined feedback value on an output of the LLM generated based on that prompt and the first set of data;
inputting to the LLM the prompt with the highest score and the first set of data; and
outputting a result generated by the LLM based on the input prompt and the first set of data.
Regarding the claim as presented above, the Examiner has bolded the elements believe to read on an abstract idea and has underlined the elements believed to read on secondary elements in the claim.
This judicial exception is not integrated into a practical application because selecting a prompt, e.g. from many available prompts, is essentially a decision under the broadest reasonable interpretation of the claim, and hence a mental step for purposes of this analysis. More granularly, and in relation to prompt selection, determining a score for essentially each eligible/available prompt in a set of prompts could be understood as either a mental step or as a mathematical step. In either case, score determination when recited at this level generality is essentially an abstract idea. Further, the clarification that a score as just mentioned is based on a predicted probability or receiving a defined feedback value on an output of a LLM based on the prompt and subject data is still just a mathematical step and/or evaluation or judgment. For example, a person could compute the score based on their prediction as mentally determined or as mathematically computed.
The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception because (i) the providing of the prompt as selected on some mental step or mathematical basis as an input to a model, and then (ii) outputting the model’s result having used the prompt as an input, are essentially insignificant extra-solution activity that (a) do not effectively integrate the abstract idea into a practical application or (b) provide significantly more than the abstract idea.
Independent claims 8 and 15 are similar to independent claim 1, and essentially feature some or all of the same limitations. Hence, they too are rejected as being directed to an abstract idea without significantly more, under a similar reasoning as provided above per claim 1.
Dependent claims 2-5, 7, 9-12, 14, and 16-20 each depend from one of the independent claims addressed above, and are likewise rejected. The Examiner does not believe their respective subject matter as claimed to provide any further rationale to integrate the abstract idea, per the independent claims 1, into a practical application or otherwise provide significantly more. With respect to these additional claims, the Examiner provides the following further comment:
Claims 2 and 9 are directed to a clarification that the model used to score the prompts is pre-trained in a particular way. While this is a useful clarification for implementing the invention / understanding it, the Examiner notes that it does not provide an active training step or active feedback step to the scope of claims as pending, and rather, for purposes of using the model within the active scope of the claim, the model is akin to a static model that can be equated with a static algorithm or an off the shelf model that does not actually improve. The Examiner would decide differently, see e.g., the Examiner’s non-rejection under section 101 for claims 6 and 13, if the improvement was clearly within the active scope of the claim’s steps, but rather the Examiner interprets these clarifications to merely define how the model was arrived at prior to the execution of the present claim’s active scope. For these reasons, the Examiner does not find the additional subject matter claimed here to integrate the claimed invention into a practical application and/or provide significantly more.
Claims 3 and 10 are directed to a clarification as to what type of data is subject to the understood mental steps/math/abstract idea, which does not add for any further meaningful integration into a practical application or provide significantly more. Said another way, merely defining or changing the type of data that is subject to a mental step and/or math, e.g. here tabular data, does not change the reality of the subject matter still being directed to an abstract idea.
Claims 4 and 11 are directed to a clarification of a generic feedback receipt step, which the Examiner characterizes as insignificant extra-solution activity, particularly since this present claim does not do anything meaningful with the feedback as received, e.g. in contrast to claims 6 and 13 which the Examiner has declined to reject under section 101.
Claims 5, 12, and 17 are directed to clarification of what the aforementioned feedback (see just above) is in terms of a type or a format. The Examiner’s rationale provided above per claim 4 still applies here. Merely defining the feedback’s type/format but not actively doing anything with it does not persuasively integrate the abstract idea of the independent claims into a practical application or provide something significantly more.
Claims 7 and 14
Claim 16 is directed to a clarification of what constitutes the recited data, e.g. tabular data essentially, upon which the abstract idea is applied, and further what the prompt is substantively defined to do. While these are useful clarifications for implementing the invention / understanding it, the Examiner does not believe these details to make the claim any more integrated into a practical application or otherwise provide something significantly more; rather, even with these details, the claim is still largely steps/details relating to abstract ideas in the form of mental steps and/or mathematical computations with some additional insignificant extra-solution activity.
Claims 18-19 feature essentially the same limitations already discussed above in relation to claim 1, so no further comment is provided at this time.
Claim 20 is directed to a step for using the feedback that goes beyond the mere receipt of it, as the Examiner has addressed per claims 4-5 just above, but short of using it in a manner that actively improves the functioning of a machine, field of endeavor and so forth – mostly because of its high level of generality. The Examiner contrasts the detail here, and the Examiner’s conclusion, with that found in claim 6 for example and the Examiner’s corresponding decision not to issue a rejection under section 101 for that claim.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-2, 4-6, 8-9, 11-13, 15, and 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2025/0209372 (“Muthu”) in view of Non-Patent Literature “A Gentle Introduction to Probability Scoring Methods in Python” (“Brownlee”).
Regarding claim 1, MUTHU teaches A method for selecting a prompt to input to a large language model (LLM) ([0001] discussing the automated optimization of a prompt, based on a scoring of the prompt to satisfy criteria, such that an optimized version of the prompt is essentially determined/selected as being sufficient (e.g., by the parent ML model, essentially a prompt selector/optimizer) to a child ML model (e.g., [0035] teaches that the child model may be a generative/GPT model, which the Examiner understands to be an example of a LLM as recited), and further per [0029] many iterations of the prompt can be considered, per the taught optimization process, before one is selected as sufficient (i.e., the prompt selected as sufficient is one of many prompts evaluated), and further still per [0031] a selected prompt may be stored and subject to reuse many times to accomplish different tasks, in which the case that prompt would be newly selected as sufficient for some other task, implying from this that storing of prompts involves a store of prompts from which prompts may be considered and selected), the method comprising:
determining, by a predictive model and for a first set of data to be the subject of a prompt of a pre-defined set of prompts, a score for each prompt of the pre-defined set of prompts ([0045]-[0046] discussing the scoring of prompts by a parent model to determine whether a prompt is sufficient, as part of the prompt optimization process as taught, and further where the prompt is understood to result in the execution of the child model to generate an output based on some input ([0003) or “detail” ([0042]) (i.e., corresponding to the recited “first set of data”), and that the prompt as selected may be one of many prompt iterations considered or one of many prompts that have been stored ([0029] and [0031] as previously discussed just above in relation to the preamble)),
wherein the score for each prompt is based on … predicting … receiving a defined feedback value on an output of the LLM generated based on that prompt and the first set of data (the child model’s output is based on the prompt and information/detail ([0003] and [0042] as discussed just above), and per [0046] and [0056] the scoring is understood to possibly be “a binary indication of whether or not each of the scoring criteria was satisfied” and predictively relates to (per [0052]) whether such a prompt would garner user feedback in the way of accepting or rejecting it’s use);
inputting to the LLM the prompt with the highest score and the first set of data and outputting a result generated by the LLM based on the input prompt and the first set of data ([0031] discussing that the optimized prompt is ultimately used with the child model to perform tasks, and the Examiner understands that prompts generally may include information/detail ([0003] and [0042] as discussed just above) as inputs needed for the substantive task at hand as facilitated by the child model, and where the child model’s performance of the task is understood to result in an “output” per [0036]-[0037] which corresponds to the recited “result generated by the LLM”).
Regarding the limitation addressed above for scoring each prompt on a particular basis, Muthu is silent as to any mention of probabilities relating to its scoring of prompts, e.g. per the entirety of the limitation clarifying “wherein the score for each prompt is based on a predicted probability of receiving a defined feedback value on an output of the LLM generated based on that prompt and the first set of data.” As mentioned above, Muthu’s parent model generates/computes a score for each prompt, and selects the prompt on that basis. The model is clearly a machine-learning model ([0040]) and hence would be understood by one of ordinary skill in the art to inherently feature probabilities, the evaluation of which drive the ultimate result of its model output, e.g. the score as mentioned here. However, to the extent that Muthu’s teaching isn’t explicit enough or unsound for this further proposition, the Examiner then relies upon BROWNLEE to teach what Muthu otherwise lacks, see e.g., Brownlee’s page 2 teaching “Log loss, also called “logistic loss,” “logarithmic loss,” or “cross entropy” can be used as a measure for evaluating predicted probabilities” (just under the heading Log Loss Score), such that a classifier/model’s performance is evaluated in terms of a loss which in turn is an examination of “predicted probabilities.” Hence, the probabilities that are arrived at as part of executing the model serve as a basis for the model’s output. See also Brownlee’s pages 8-9, under the heading ROC AUC Score, which details the role of predicted probabilities used in implementing a binary classification problem via such modelling. Hence, again, probabilities that are arrived at as part of executing the model serve as a basis for the model’s output.
Muthu and Brownlee both relate to machine-learning models used in classification or similar problems. Hence, they are similarly directed and therefore analogous. It would have been obvious to implement the parent model per Muthu to evaluate the prompt in a probabilistic manner, as Brownlee teaches, with a reasonable expectation of success, to arrive at the same scoring indication result that Muthu contemplates for that same model’s output. The Examiner believes essentially Brownlee teaches the nuts and bolts of model operation with a depth that is granular enough to detail the role of probabilities, and that such teachings would be foundational to how the models taught by Muthu work, but that Muthu is more concerned with higher level aspects of using the models once understood/implemented.
Regarding claim 2, Muthu in view of Brownlee teach the method of claim 1, as discussed above. The references further teach the additional limitations wherein the predictive model is pre-trained based on feedback values for a training set of outputs of the LLM generated based on each prompt of the pre-defined set of prompts and one or more training sets of data (Muthu: model training, inclusive of the parent model, is understood to involve conventional neural network training aspects as detailed in [0040], including that of training inputs/outputs, and in the context of the invention it would be understood to leverage the outputs from the child model as a function of the optimized prompt provided via the parent model, e.g. as shown in the flow diagrams per FIGs. 1-2). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 4, Muthu in view of Brownlee teach the method of claim 1, as discussed above. The references further teach the additional limitation comprising receiving a feedback value on the output result generated by the LLM (Muthu: [0039] discussing the user’s provision of a feedback aspect relating to the prompt as optimized and the corresponding output of the child model). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 5, Muthu in view of Brownlee teach the method of claim 4, as discussed above. The references further teach the additional limitation wherein the feedback value comprises one of: binary feedback; a rating on a scale; and a rating on a scale converted from textual feedback (Muthu: [0039] includes the user’s accepting or rejecting the optimized prompt, i.e. a binary feedback as recited). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 6, Muthu in view of Brownlee teach the method of claim 4, as discussed above. The references further teach the additional limitation wherein the feedback value on the output result generated by the LLM is used to further train the predictive model (Muthu: [0039] teaches “… Feedback may also be received in the form of a user 108 accepting or rejecting optimized prompts 115, or a user 108 modifying an optimized prompt 115 or otherwise providing alternative prompts. The information gathered by the user feedback engine 130 may then be used to re-train one or more of the models of the prompt optimization engine 110 and/or child machine learning model 100 …”). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 8, the claim includes the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale. The claim additionally recites at least one computer processor and a computer readable storage medium comprising instructions essentially corresponding to the subject matter of claim 1, and the Examiner believes Muthu teaches these further features: see e.g., Muthu’s [0060]-[0064] discussing CPU and memory elements that read on the additional features being addressed here.
Regarding claim 9, the claim includes the same or similar limitations as claim 2 discussed above, and is therefore rejected under the same rationale.
Regarding claim 11, the claim includes the same or similar limitations as claim 4 discussed above, and is therefore rejected under the same rationale.
Regarding claim 12, the claim includes the same or similar limitations as claim 5 discussed above, and is therefore rejected under the same rationale.
Regarding claim 13, the claim includes the same or similar limitations as claim 6 discussed above, and is therefore rejected under the same rationale.
Regarding claim 15, the claim includes some of the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale. The Examiner notes that present claim uses the term “generative artificial intelligence” instead of claim 1’s “large language model (LLM)”, and also the term “positive feedback” instead of claim 1’s “defined feedback value”; however, the Examiner believes the mappings as provided per claim 1 and its recited terms still read on these similar terminologies of the present claim.
Regarding claim 17, the claim includes the same or similar limitations as claim 5 discussed above, and is therefore rejected under the same rationale.
Regarding claim 18, the claim includes some of the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale.
Regarding claim 19, the claim includes some of the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale.
Regarding claim 20, the claim includes the same or similar limitations as claim 6 discussed above, and is therefore rejected under the same rationale.
Claims 3, 10, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Muthu in view of Brownlee and further in view of Non-Patent Literature “Table-GPT: Empower LLMs to Understand Tables” (“AI Papers”).
Regarding claim 3, Muthu in view of Brownlee teach the method of claim 1, as discussed above, but not the further limitation wherein the first set of data is a set of tabular data. At best, Muthu teaches the use of a GPT type model for its child model but is rather agnostic as to what type/format of data it takes in. Hence, Muthu seems generalized for that type of model at least, if not even broader and extensible to other types of models/AI that can receive prompts for example. Rather, the Examiner relies upon AI PAPERS to teach what Muthu etc. otherwise lack, see e.g., AI Papers’s focus is a modification to GPT-based model pipelines that permit the providing of input data in tabular form.
Like Muthu, AI Papers is directed to prompt-based generative modeling. Hence, they are similarly directed and therefore analogous. It would have been obvious to incorporate AI Papers’s feature, as relied upon, into a framework such as Muthu’s as modified, with a reasonable expectation of success, to permit a broader and more explicit of different types of data inputs for the prompting and prompting tuning/engineering, and specifically one inclusive of tabular data since tables are a common and widely-used data construct in software and numerical-based information systems that lend themselves to leverage/interrogation by computing processes, such as GPT models for example.
Regarding claim 10, the claim includes the same or similar limitations as claim 3 discussed above, and is therefore rejected under the same rationale.
Regarding claim 16, the claim includes the same or similar limitations as claim 3 discussed above, and is therefore rejected under the same rationale.
Claims 7 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Muthu in view of Brownlee and further in view of Non-Patent Literature “Evaluating Prompts: A Developer’s Guide” (“Dhinakaran”).
Regarding claim 7, Muthu in view of Brownlee teach the method of claim 4, as discussed above, but not the further limitation wherein the feedback value on the output result generated by the LLM is used to create at least one new prompt to be added to the pre-defined set of prompts. As discussed above in relation to claim 1, the Examiner believes Muthu can be understood to contemplate prompt reuse and hence prompt storage and the consideration of existing prompts for other tasks. See, e.g., Muthu’s [0031] teaching that a selected prompt may be stored and subject to reuse many times to accomplish different tasks, in which the case that prompt would be newly selected as sufficient for some other task, implying from this that storing of prompts involves a store of prompts from which prompts may be considered and selected Hence, the Examiner reasons that it would be obvious to not just keep/store prompts as Muthu appears to do so but to do so especially for prompts that have been given favorable feedback by the user, as Muthu also provides for and as discussed previously in relation to claim 1. There would be a natural motivation to reuse prompts, since storage thereof is taught, that have been given positive/favorable feedback, within Muthu’s framework, since the aim of Muthu is to improve in part by improving the prompts themselves via feedback, and hence there is value in keeping prompts that have been valued per that user feedback.
However, to the extent that Muthu is not concrete enough in this regard, the Examiner would then further rely upon DHINAKARAN to teach what Muthu etc. otherwise lack, see e.g., Dhinakaran’s page 13 explicitly detailing the use and advantages of a prompt registry, which facilitates prompt reuse and study/evaluation in a manner that improves prompt engineering/tuning and thereby improves the larger framework within which such a feature is implemented.
Like Muthu, Dhinakaran is directed to prompt-based generative modeling and more specifically prompt tuning/engineering. Hence, they are similarly directed and therefore analogous. It would have been obvious to incorporate Dhinakaran’s registry feature, as relied upon, into a framework such as Muthu’s as modified, with a reasonable expectation of success, to concretely provide a storage with the stated purpose of improving prompt engineering/tuning and promoting their reuse and sharing, e.g. across tasks as Muthu itself contemplates.
Regarding claim 14, the claim includes the same or similar limitations as claim 7 discussed above, and is therefore rejected under the same rationale.
Conclusion
The prior art made of record and not relied upon is considered pertinent to Applicants’ disclosure:
CN-116484879-A
CN-117194078-A
Non-Patent Literature “Applying Large Language Models to Tabular Data: A New Approach” (Lopatecki)
Non-Patent Literature “Softmax Activation Function: Everything You Need to Know” (Bala)
Non-Patent Literature “Table-GPT: Empower LLMs to Understand Tables” (AI Papers Academy)
Non-Patent Literature “Deep Learning: Getting Started” (Ponnambalan)
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SHOURJO DASGUPTA whose telephone number is (571)272-7207. The examiner can normally be reached M-F 8am-5pm CST.
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, Tamara Kyle can be reached at 571 272 4241. 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.
/SHOURJO DASGUPTA/Primary Examiner, Art Unit 2144
Prosecution Insights