DETAILED ACTION
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 .
All objections/rejections not mentioned in this Office Action have been withdrawn by the Examiner.
Examiner’s Note
In the Applicant’s Summary of interview filed 27 February 2026 and the Response filed 02 March 2026, applicant’s representative indicates that an interview was held on 26 February 2026 between Examiner Serraguard and Mr. Breska, attorney for the applicant. Examiner appreciates the courtesy extended in the remarks. However, upon review, no contemporaneous records of such an interview could be found. Respectfully, there were no available contemporaneous records indicating correspondence, including formal/informal requests for interview, emails presenting the agenda/proposed amendments, or other submission of discussion items, such that the examiner could be aware of the contents and/or could have provided reasoned comments regarding such extensive amendments. Though representative indicates that the amendments were discussed, and that the submitted amendments conform to those “discussed during the examiner interview,” a record of such amendments could not be found. Further, there is no corresponding examiner interview summary filed, which indicates that examiner was not aware, as of 26 February 2026, that an interview was held.
Applicant’s position regarding the amendments being in light of the interview is duly noted. Though examiner cannot rule out the possibility that a phone conversation about the application occurred, contemporaneous records, as maintained for an interview in the normal course of examination, do not support the existence of a substantive review or discussion regarding proposed amendments to claims 1-20. As such, claims 1-20 are reviewed here as newly presented in the Response. Applicant is advised to provide formal interview requests in the future, including formal submissions of agendas and proposed amendments, such that the contents of such interviews can be made clearly of record.
Status of the Claims
Prior to entry of the amendment(s) and/or consideration of the argument(s), the status of the claims is as follows.
Claim(s) 1-20 is/are pending.
Claim(s) 1-20 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Upadhyay (U.S. Pat. App. Pub. No. 2025/0053876, hereinafter Upadhyay).
Response to Amendments
Applicant’s amendment filed on 02 March 2026 has been entered.
In view of the amendment to the claim(s), the amendment of claim(s) 1-9, and 11-19; the cancellation of claim(s) 20; and the addition of claim(s) 21 have been acknowledged and entered.
In view of the amendment to claim(s) 1, 11, and 17 and the cancellation of claim(s) 20, the rejection of claims 1-20 under 35 U.S.C. §102 is withdrawn.
In light of the amended/newly added claims, new grounds for rejection under 35 U.S.C. §103 and 35 U.S.C. §112 are provided in the action below.
Response to Arguments
Applicant’s arguments regarding the prior art rejections under 35 U.S.C. §102/103, see pages 11-14 of the Response to Non-Final Office Action dated 28 November 2026, which was received on 02 March 2026 (hereinafter Response and Office Action, respectively), have been fully considered.
With respect to the rejection(s) of claim(s) 1, 11, and 17 under 35 U.S.C. §102 as being anticipated by Upadhyay, applicant asserts that Upadhyay fails to teach or suggest all limitations of claims 1, 11, and 17. Applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 1, 11, and 17 under 35 U.S.C. §102 are withdrawn.
Applicant further argues that the rejection(s) of dependent claims 2-10, 12-16, and 18-19 should be withdrawn for at least the same reasons as independent claims 1, 11, and 17. Applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 2-10, 12-16, and 18-19 under 35 U.S.C. §102 are withdrawn.
However, upon further consideration, new ground(s) of rejection under 35 U.S.C. §103 are made in light of combinations of Upadhyay and Non-patent literature to Chen (Chen, L., Zaharia, M. and Zou, J., 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, hereinafter Chen).
The Applicant has not provided any further statement and therefore, the Examiner directs the Applicant to the below rationale.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 1-20 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Regarding claim 1, and mutatis mutandis claims 11 and 17, the limitation “sequencing the two or more models…” lacks specification support. Claim 1 recites “sequencing the two or more models in the selected one of the suggested combinations” at lines 16-17. First, we determine the broadest reasonable interpretation (BRI) of the phrase, as used in claim 1. As the specification fails to expressly define “sequencing”, we look to the ordinary meaning of the word in light of the specification to determine the meaning. However, upon further review, the specification also does not use the word “sequencing.” As such, the BRI of sequencing in light of the ordinary meaning of the word is determining an order for something (i.e., a sequence). In light of the BRI, it is acknowledged that the specification does describe “identify[ing] which AI based models should be combined (and in what specific order).” (Instant Application, [0045]). As one skilled in the art would understand the “sequencing” of selected models, in light of the BRI, as a determination of “which AI based models should be combined (and in what specific order),” the word “sequencing” appears initially to have descriptive specification support.
However, in claim 3, which depends from claim 1, applicant provides the further limitation “wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order.” In light of the principles of claim differentiation, the phrase “sequencing the two or more models…” in claim 1 must necessarily be broader than the further limitation of “combining the two or more models in the selected one of the suggested combinations in a specific order” in claim 3. As there is no known support for sequencing meaning something broader than the further limitation in claim 3, “sequencing” as used in claim 1 lacks specification support.
Though the above argument is described in the context of claims 1 and 3, the same argument is applicable to the combinations of claims 11 and 12, and claims 17 and 18, respectively. Therefore, claims 1, 11, and 17 contain limitations which lack specification support and are rejected.
Regarding claims 2-10, 12-19, and 21, claims 2-10, 12-19, and 21 depend from claims 1, 11, and 17 and incorporate all limitations therefrom. Therefore, claims 2-10, 12-19, and 21 are rejected under 35 U.S.C. 112(a) for at least the same reasons as claims 1, 11, and 17.
Appropriate correction is required.
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.
Claim(s) 1-19 and 21 is/are rejected under 35 U.S.C. 103 as being unpatentable over Upadhyay in view of Chen.
Regarding claim 1, Upadhyay discloses A computer-implemented method (CIM) (Systems and methods for machine learning prompt routing; Upadhyay, ¶ [0025] (Further evidence at Prov. App. No. 63/588,591, [0016])), comprising: receiving, at a context-aware multi-model aggregator, a user query from an endpoint device (the system using “router 100...receive[s] a runtime prompt from a prompter 10 (e.g., a service, a user, a user device, etc.).”; Upadhyay, ¶ [0134] (Further evidence at Prov. App. No. 63/588,591, [0052])); evaluating the user query (The system can further “determine a performance score for each candidate model in the set of candidate models 200 given a runtime prompt” which is based on the user query; Upadhyay, ¶ [0135], [0138] (Further evidence at Prov. App. No. 63/588,591, [0052])); selecting a combination of models to evaluate the user query based at least in part on the evaluation of the user query (A “candidate model can be selected based on: the determined performance score or set of performance scores (e.g., from S330), the determined metadata (e.g., from S320), selection parameters (e.g., user preferences), and/or any other suitable information” and “Each candidate model… can differ in:... architecture (e.g., transformers, diffusion models, generative models, etc.)” where a transformer model is, by definition a combination of models (i.e., an encoder and a decoder); Upadhyay, ¶ [0042], [0149] (Further evidence at Prov. App. No. 63/588,591, [0073])) by submitting contextual information extracted from the user query by a context analyzer, to a service mapper of the context-aware multi- model aggregator (Discloses the Score Prediction Model 112 (service mapper) which “can receive a prompt and optionally additional information (e.g., prompter identifier, context, etc.) as input and generate an output”; Upadhyay, ¶ [0072] (Further evidence at Prov. App. No. 63/588,591, [0047])), receiving suggested combinations of two or more models that may be used to evaluate the user query (“The SPM preferably includes a first set of encoding layers configured to determine an encoding for a prompt and a second set of parallel sets of decoder layers (e.g., one set of decoder layers for each candidate model) configured to determine a score” and each scoring head “outputs a predicted performance score for the respective candidate model (e.g., indicative of the quality of the response that the candidate model is expected to produce responsive to the prompt)”; Upadhyay, ¶ [0071], [0074] (Further evidence at Prov. App. No. 63/588,591, [0055]-[0056])), and selecting one of the suggested combinations of two or more models to evaluate the user query (“Selecting a candidate model S340 functions to identify the best candidate model to use to generate a response (e.g., wherein the best candidate model is the selected model 210). The candidate model can be selected from the set of candidate models corresponding to the set of similar stored prompt representations 410”; Upadhyay, ¶ [0149] (Further evidence at Prov. App. No. 63/588,591, [0056]-[0057]))... the two or more models in the selected one of the suggested combinations…[being in a sequence] (The encoder and decoder models of a transformer model, which may be the combination of two or more models selected as the candidate model, are sequenced, understood as positioned in a specified order, in the selected one of the suggested combinations.; Upadhyay, ¶ [0042]-[0045]); using the sequenced two or more models to process the user query by passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models (The encoder and decoder models of a transformer model processes the user query at the encoder model to produce an encoded output. The encoded output is then provided as the input to the decoder model, the decoder model being the successive model of the two models.; Upadhyay, ¶ [0042]-[0045]); transmitting a final output, produced by the sequenced two or more models, to the endpoint device (“A result can optionally be returned to the prompter 10 (e.g., directly, via the system, etc.). A result can include a response from the selected candidate model, an identifier for a selected candidate model, a set of ranked candidate models 200, a message (e.g., “no suitable model could be found”), predicted performance scores, performance scores determined from the reward model 300, metadata, and/or other suitable information or combination of information.”; Upadhyay, ¶ [0160] (Further evidence at Prov. App. No. 63/588,591, [0081])); and performing reinforcement learning based at least in part on feedback received from the endpoint device (“The models can be trained or learned using... reinforcement learning... on: labeled data (e.g., data labeled with the target label), unlabeled data, positive training sets (e.g., a set of data with true positive labels, negative training sets (e.g., a set of data with true negative labels)” and can be “reinforced... based on newly received, up-to-date measurements; past measurements recorded during the operating session; historic measurements recorded during past operating sessions” and training data can include “generating a new set of training data and/or amending an existing set of training data (e.g., based on user metrics)”; Upadhyay, ¶ [0095]-[0096], [0100] (Further evidence at Prov. App. No. 63/588,591, [0074], [0082]-[0083])). However, Upadhyay fails to expressly recite the active step of “sequencing” said models.
Chen teaches “FrugalGPT, a simple yet flexible instantiation of LLM cascade which learns which combinations of LLMs to use for different queries” thus selecting both specified models and combinations thereof for sequential presentation of queries. (Chen, ¶ Abstract). Regarding claim 1, Chen teaches sequencing the two or more models in the selected one of the suggested combinations (Discloses the use of a “Model fine-tuning” which “uses expensive LLMs’ responses to fine-tune cheap LLMs.” As shown in FIG. 2(d), the prompt is received by GPT-4, where the output of GPT-4 is fed to GPT-J, to fine-tune GPT-J to provide the final answer to the query. It is noted that this is both a selection of models from a set of available models and a sequencing of that selection.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include sequencing the two or more models in the selected one of the suggested combinations. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 2, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the user query is evaluated using a context analyzer of the context-aware multi-model aggregator (The system discloses a context analyzer in the form of the encoder 20/similarity model 111 (e.g., the “encoder 20 “ can be “integrated within the similarity model 111 and/or separate from the similarity model 111”) that analyzes (“the router 100 includes a similarity model 111 configured to determine a similarity of the prompt to another prompt”) and encodes (the encoder is “configured to encode the prompt into a prompt encoding”) the prompt.; Upadhyay, ¶ [0035], [0062], [0114] (Further evidence at Prov. App. No. 63/588,591, [0018], [0022])).
Regarding claim 3, the rejection of claim 2 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. However, Upadhyay fail(s) to expressly recite wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order.
The relevance of Chen is described above with relation to claim 1. Regarding claim 3, Chen teaches wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, (The “Model fine-tuning” includes combining two selected models, being GPT-4 and GPT-J, in the specific order shown in FIG. 2(d). This specific order and combination allows GPT-4 to produce an output that, when used as an input at GPT-J, fine-tunes GPT-J to provide the final answer to the query.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)) wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order (The specific order and combination disclosed at FIG. 2(d) allows GPT-4 to produce an output which is received by GPT-J and used for fine tuning, where GPT-J is the successive one of the two or more models.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 4, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the... models includes one or more publicly available models and one or more private models (Discloses candidate models that are third party (publicly available), such as “OpenAI” and “Anthropic” models, and locally executed/user specific (private). “Each candidate model can be executed by the user (e.g., on a user device)” which is understood as a local version of the available models, where a local version of a model is a private model.; Upadhyay, ¶ [0041]-[0042] (Further evidence at Prov. App. No. 63/588,591, [0029]; App. No. 63/532,199, pg. 1, para. 5)). However, Upadhyay fail(s) to expressly recite wherein the selected one of the suggested combinations of models includes one or more publicly available models and one or more private models.
The relevance of Chen is described above with relation to claim 1. Regarding claim 4, Chen teaches wherein the selected one of the suggested combinations of models includes one or more publicly available models and one or more private models (The specific order and combination disclosed at FIG. 2(d), regarding using GPT-4 to produce an output which is received by GPT-J, is part of “optimizing over the selection of different LLM APIs (e.g., GPT-J, ChatGPT, and GPT-4) as well as prompting strategies” to maintain quality while reducing costs. It would be well within ordinary skill in the art to combine public and private models, as disclosed by Upadhyay, as part of the fine tuning of Chen, as the overarching goal of FrugalGPT is maintaining or improving quality while reducing inference costs. The system uses all available models and, so long as the goals are met, the source of the model is irrelevant. Further, in the context of this example, each of these models is private from the perspective of the owner (in that the models are presented in a black box format to the end user and the models are wholly owned by the respective corporation). Thus, a combination created by OpenGPT which includes GPT-4 and GPT-J is a combination of public and private models from the perspective of OpenGPT.; Chen, ¶ Pg, 2, lines 6-13; Pg. 4, FIG. 2(a)-(e))..
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the selected one of the suggested combinations of models includes one or more publicly available models and one or more private models. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 5, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the … models includes two or more different models that are only private models (“The set of candidate models 200” can “include a set of models provided by the same provider” and the “candidate models” can be run “on the same system as running the router” where a local version of a model is a private model. As it indicates that each candidate model “can be executed by the user (e.g., on a user device)”, it discloses the use of only private models.; Upadhyay, ¶ [0041]-[0042] (Further evidence at Prov. App. No. 63/588,591, [0029]; App. No. 63/532,199, pg. 1, para. 5)). However, Upadhyay fail(s) to expressly recite wherein the selected one of the suggested combinations of models includes two or more different models that are only private models.
The relevance of Chen is described above with relation to claim 1. Regarding claim 5, Chen teaches wherein the selected one of the suggested combinations of models includes two or more different models that are only private models (The specific order and combination disclosed at FIG. 2(d), regarding using GPT-4 to produce an output which is received by GPT-J, is part of “optimizing over the selection of different LLM APIs (e.g., GPT-J, ChatGPT, and GPT-4) as well as prompting strategies” to maintain quality while reducing costs. It would be well within ordinary skill in the art to combine two or more private models, as part of the fine tuning of Chen, as the overarching goal of FrugalGPT is maintaining or improving quality while reducing inference costs. The system uses all available models and, so long as the goals are met, the source of the model is irrelevant. Thus, the use of any or all public and/or private is both an incidental fact of the operation of FrugalGPT and part of routine optimization of the same.; Chen, ¶ Pg, 2, lines 6-13; Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the selected one of the suggested combinations of models includes two or more different models that are only private models. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 6, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. However, Upadhyay fail(s) to expressly recite further comprising: selecting a second of the suggested combinations of two or more models to evaluate the user query.
The relevance of Chen is described above with relation to claim 1. Regarding claim 6, Chen teaches further comprising: selecting a second of the suggested combinations of two or more models to evaluate the user query (Chen further discloses “LLM cascade” which “sends a query to a list of LLM APIs sequentially” where “If one LLM API’s response is reliable, then its response is returned” and the “remaining LLM APIs are queried only if the previous APIs’ generations are deemed insufficiently reliable.” It would be well within ordinary skill in the art to combine this technique, depicted in FIG. 2(e), with the LLM fine-tuning described in FIG. 2(d), such that two separate and cascaded LLM fine-tuning combinations are selected. The first LLM fine tuning combination produces an insufficient response, the query would be forwarded to the second LLM-fine-tuning combination.; Chen, ¶ Pg. 4, FIG. 2(a)-(e); pg. 5, “Strategy 3: LLM Cascade”).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include further comprising: selecting a second of the suggested combinations of two or more models to evaluate the user query. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 7, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the selected one of the suggested combinations of models includes two or more artificial intelligence (AI) based models selected from the group consisting of: machine learning models, generative AI models, deep learning models, natural language processing models, large language models, and foundation models (The “transformers” cited above includes the encoder and decoder, which are, at least, machine learning models.; Upadhyay, ¶ [0044] (Further evidence at Prov. App. No. 63/588,591, [0029], [0082])).
Regarding claim 8, the rejection of claim 7 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. However, Upadhyay fail(s) to expressly recite wherein the selected one of the suggested combinations of models includes two or more different types of generative AI models and large language models.
The relevance of Chen is described above with relation to claim 1. Regarding claim 8, Chen teaches wherein the selected one of the suggested combinations of models includes two or more different types of generative AI models and large language models (The “Model fine-tuning” includes combining two selected models, being GPT-4 and GPT-J, in the specific order shown in FIG. 2(d), where both GPT-4 and GPT-J are both different types of generative AI models and different types of large language models.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the selected one of the suggested combinations of models includes two or more different types of generative AI models and large language models. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 9, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the performing of the reinforcement learning includes: in response to receiving positive feedback from the endpoint device, increasing a score assigned to the one of the suggested combinations of models that produced the transmitted final output (“target performance scores can be generated from user metrics received from the user and/or generated from indicators relating to the user” including “user metrics can be used to evaluate a result determined in S300 (e.g., a response from the selected model 210, etc.)” which includes both positive and negative evaluations, and “the performance score” of the “selected prompt representation-candidate model pair” is increased or decreased, based on the user evaluation being positive or negative, respectively.; Upadhyay, ¶ [0109] (Further evidence at Prov. App. No. 63/588,591, [0016], [0078], [0080]-[0081])); and in response to receiving negative feedback from the endpoint device, decreasing a score assigned to the one of the suggested combinations of models that produced the transmitted final output (In an example, “when a user metric indicates that a result was poor quality, the reward model can generate a performance score from the prompt-response pair and penalize it based on the user metric. In another example, in the similarity variant of the router, the stored prompt representation associated with the selected prompt representation-candidate model pair can be determined, and the performance score associated with the stored prompt representation can be decreased.”; Upadhyay, ¶ [0109] (Further evidence at Prov. App. No. 63/588,591, [0016], [0078], [0080]-[0081])).
Regarding claim 10, the rejection of claim 1 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the context-aware multi-model aggregator is located at a central server, (“The router 100, reward models 300, and/or other system models are preferably hosted by the same entity (e.g., a “router provider”), but can alternatively be hosted by different entities” and “the entities can be in communication with one another via APIs”; Upadhyay, ¶ [0094] (Further evidence at Prov. App. No. 63/588,591, [0080])) wherein the endpoint device and the central server are both connected to a network (“The router 100, reward models 300, and/or other system models... can be in communication with one another via APIs” and “a runtime prompt” is received “from a prompter 10 (e.g., a service, a user, a user device, etc.)” where “runtime prompt can be received from a user” such as via “an API call, [or] a third party system”; Upadhyay, ¶ [0094], [0134] (Further evidence at Prov. App. No. 63/588,591, [0080])).
Regarding claim 11, Upadhyay discloses A computer program product (CPP) (Systems and methods for machine learning prompt routing; Upadhyay, ¶ [0025] (Further evidence at Prov. App. No. 63/588,591, [0016])), comprising: a set of one or more computer-readable storage media; and program instructions, collectively stored in the set of one or more storage media, for causing a processor set to perform the following computer operations (Discloses “implement[ing] the above methods and/or processing modules in non-transitory computer-readable media, storing computer-readable instructions that, when executed by a processing system, cause the processing system to perform the method(s) discussed herein. The instructions can be executed by computer-executable components integrated with the computer-readable medium and/or processing system. The computer-readable medium may include any suitable computer readable media”; Upadhyay, ¶ [0165] (Further evidence at Prov. App. No. 63/588,591, [0085])): receive, at a context-aware multi-model aggregator, a user query from an endpoint device (the system using “router 100...receive[s] a runtime prompt from a prompter 10 (e.g., a service, a user, a user device, etc.).”; Upadhyay, ¶ [0134] (Further evidence at Prov. App. No. 63/588,591, [0052])); evaluate the user query (The system can further “determine a performance score for each candidate model in the set of candidate models 200 given a runtime prompt” which is based on the user query; Upadhyay, ¶ [0135], [0138] (Further evidence at Prov. App. No. 63/588,591, [0052])); select a combination of models to evaluate the user query based at least in part on the evaluation of the user query (A “candidate model can be selected based on: the determined performance score or set of performance scores (e.g., from S330), the determined metadata (e.g., from S320), selection parameters (e.g., user preferences), and/or any other suitable information” and “Each candidate model… can differ in:... architecture (e.g., transformers, diffusion models, generative models, etc.)” where a transformer model is, by definition a combination of models (i.e., an encoder and a decoder); Upadhyay, ¶ [0042], [0149] (Further evidence at Prov. App. No. 63/588,591, [0073])) by submitting contextual information extracted from the user query by a context analyzer, to a service mapper of the context- aware multi-model aggregator (Discloses the Score Prediction Model 112 (service mapper) which “can receive a prompt and optionally additional information (e.g., prompter identifier, context, etc.) as input and generate an output”; Upadhyay, ¶ [0072] (Further evidence at Prov. App. No. 63/588,591, [0047])), receiving suggested combinations of two or more models that may be used to evaluate the user query (“The SPM preferably includes a first set of encoding layers configured to determine an encoding for a prompt and a second set of parallel sets of decoder layers (e.g., one set of decoder layers for each candidate model) configured to determine a score” and each scoring head “outputs a predicted performance score for the respective candidate model (e.g., indicative of the quality of the response that the candidate model is expected to produce responsive to the prompt)”; Upadhyay, ¶ [0071], [0074] (Further evidence at Prov. App. No. 63/588,591, [0055]-[0056])), and selecting one of the suggested combinations of two or more models to evaluate the user query (“Selecting a candidate model S340 functions to identify the best candidate model to use to generate a response (e.g., wherein the best candidate model is the selected model 210). The candidate model can be selected from the set of candidate models corresponding to the set of similar stored prompt representations 410”; Upadhyay, ¶ [0149] (Further evidence at Prov. App. No. 63/588,591, [0056]-[0057]))... the two or more models in the selected one of the suggested combinations…[being in a sequence] (The encoder and decoder models of a transformer model, which may be the combination of two or more models selected as the candidate model, are sequenced, understood as positioned in a specified order, in the selected one of the suggested combinations.; Upadhyay, ¶ [0042]-[0045]); use the sequenced two or more models to process the user query by passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models (The encoder and decoder models of a transformer model processes the user query at the encoder model to produce an encoded output. The encoded output is then provided as the input to the decoder model, the decoder model being the successive model of the two models.; Upadhyay, ¶ [0042]-[0045]); transmit a final output, produced by the sequenced two or more models, to the endpoint device (“A result can optionally be returned to the prompter 10 (e.g., directly, via the system, etc.). A result can include a response from the selected candidate model, an identifier for a selected candidate model, a set of ranked candidate models 200, a message (e.g., “no suitable model could be found”), predicted performance scores, performance scores determined from the reward model 300, metadata, and/or other suitable information or combination of information.”; Upadhyay, ¶ [0160] (Further evidence at Prov. App. No. 63/588,591, [0081])); and perform reinforcement learning based at least in part on feedback received from the endpoint device (“The models can be trained or learned using... reinforcement learning... on: labeled data (e.g., data labeled with the target label), unlabeled data, positive training sets (e.g., a set of data with true positive labels, negative training sets (e.g., a set of data with true negative labels)” and can be “reinforced... based on newly received, up-to-date measurements; past measurements recorded during the operating session; historic measurements recorded during past operating sessions” and training data can include “generating a new set of training data and/or amending an existing set of training data (e.g., based on user metrics)”; Upadhyay, ¶ [0095]-[0096], [0100] (Further evidence at Prov. App. No. 63/588,591, [0074], [0082]-[0083])). However, Upadhyay fails to expressly recite the active step of “sequencing” said models.
The relevance of Chen is described above with relation to claim 1. Regarding claim 11, Chen teaches sequence the two or more models in the selected one of the suggested combinations (Discloses the use of a “Model fine-tuning” which “uses expensive LLMs’ responses to fine-tune cheap LLMs.” As shown in FIG. 2(d), the prompt is received by GPT-4, where the output of GPT-4 is fed to GPT-J, to fine-tune GPT-J to provide the final answer to the query. It is noted that this is both a selection of models from a set of available models and a sequencing of that selection.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include sequencing the two or more models in the selected one of the suggested combinations. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 12, the rejection of claim 11 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the user query is evaluated using a context analyzer of the context-aware multi-model aggregator, (The system discloses a context analyzer in the form of the encoder 20/similarity model 111 (e.g., the “encoder 20 “ can be “integrated within the similarity model 111 and/or separate from the similarity model 111”) that analyzes (“the router 100 includes a similarity model 111 configured to determine a similarity of the prompt to another prompt”) and encodes (the encoder is “configured to encode the prompt into a prompt encoding”) the prompt.; Upadhyay, ¶ [0035], [0062], [0114] (Further evidence at Prov. App. No. 63/588,591, [0018], [0022])). However, Upadhyay fail(s) to expressly recite wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order.
The relevance of Chen is described above with relation to claim 1. Regarding claim 12, Chen teaches wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, (The “Model fine-tuning” includes combining two selected models, being GPT-4 and GPT-J, in the specific order shown in FIG. 2(d). This specific order and combination allows GPT-4 to produce an output that, when used as an input at GPT-J, fine-tunes GPT-J to provide the final answer to the query.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)) wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order (The specific order and combination disclosed at FIG. 2(d) allows GPT-4 to produce an output which is received by GPT-J and used for fine tuning, where GPT-J is the successive one of the two or more models.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 13, the rejection of claim 11 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the selected one of the suggested combinations of models include publicly available models and private models (Discloses candidate models that are third party (publicly available), such as “OpenAI” and “Anthropic” models, and locally executed/user specific (private). “Each candidate model can be executed by the user (e.g., on a user device)” which is understood as a local version of the available models, where a local version of a model is a private model.; Upadhyay, ¶ [0041]-[0042] (Further evidence at Prov. App. No. 63/588,591, [0029]; App. No. 63/532,199, pg. 1, para. 5)).
Regarding claim 14, the rejection of claim 11 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the selected one of the suggested combinations of models include only private models (“The set of candidate models 200” can “include a set of models provided by the same provider” and the “candidate models” can be run “on the same system as running the router” where a local version of a model is a private model. As it indicates that each candidate model “can be executed by the user (e.g., on a user device)”, it discloses the use of only private models.; Upadhyay, ¶ [0041]-[0042] (Further evidence at Prov. App. No. 63/588,591, [0029]; App. No. 63/532,199, pg. 1, para. 5)).
Regarding claim 15, the rejection of claim 11 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. However, Upadhyay fail(s) to expressly recite wherein the program instructions further cause the processor set to perform the following computer operation: selecting a second of the suggested combinations of two or more models to evaluate the user query.
The relevance of Chen is described above with relation to claim 1. Regarding claim 15, Chen teaches wherein the program instructions further cause the processor set to perform the following computer operation: selecting a second of the suggested combinations of two or more models to evaluate the user query (Chen further discloses “LLM cascade” which “sends a query to a list of LLM APIs sequentially” where “If one LLM API’s response is reliable, then its response is returned” and the “remaining LLM APIs are queried only if the previous APIs’ generations are deemed insufficiently reliable.” It would be well within ordinary skill in the art to combine this technique, depicted in FIG. 2(e), with the LLM fine-tuning described in FIG. 2(d), such that two separate and cascaded LLM fine-tuning combinations are selected. The first LLM fine tuning combination produces an insufficient response, the query would be forwarded to the second LLM-fine-tuning combination.; Chen, ¶ Pg. 4, FIG. 2(a)-(e); pg. 5, “Strategy 3: LLM Cascade”).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the program instructions further cause the processor set to perform the following computer operation: selecting a second of the suggested combinations of two or more models to evaluate the user query. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 16, the rejection of claim 11 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the selected one of the suggested combinations of models includes artificial intelligence (AI) based models selected from the group consisting of: machine learning models, generative AI models, deep learning models, natural language processing models, large language models, and foundation models (The “transformers” cited above includes the encoder and decoder, which are, at least, machine learning models.; Upadhyay, ¶ [0044] (Further evidence at Prov. App. No. 63/588,591, [0029], [0082])).
Regarding claim 17, Upadhyay discloses A computer system (CS) (Systems and methods for machine learning prompt routing; Upadhyay, ¶ [0025] (Further evidence at Prov. App. No. 63/588,591, [0016])), comprising: a processor set; a set of one or more computer-readable storage media; program instructions, collectively stored in the set of one or more storage media, for causing the processor set to perform the following computer operations (Discloses “implement[ing] the above methods and/or processing modules in non-transitory computer-readable media, storing computer-readable instructions that, when executed by a processing system, cause the processing system to perform the method(s) discussed herein. The instructions can be executed by computer-executable components integrated with the computer-readable medium and/or processing system. The computer-readable medium may include any suitable computer readable media”; Upadhyay, ¶ [0165] (Further evidence at Prov. App. No. 63/588,591, [0085])): receive, at a context-aware multi-model aggregator, a user query from an endpoint device (the system using “router 100...receive[s] a runtime prompt from a prompter 10 (e.g., a service, a user, a user device, etc.).”; Upadhyay, ¶ [0134] (Further evidence at Prov. App. No. 63/588,591, [0052])); evaluate the user query (The system can further “determine a performance score for each candidate model in the set of candidate models 200 given a runtime prompt” which is based on the user query; Upadhyay, ¶ [0135], [0138] (Further evidence at Prov. App. No. 63/588,591, [0052])); select a combination of models to evaluate the user query based at least in part on the evaluation of the user query (A “candidate model can be selected based on: the determined performance score or set of performance scores (e.g., from S330), the determined metadata (e.g., from S320), selection parameters (e.g., user preferences), and/or any other suitable information” and “Each candidate model… can differ in:... architecture (e.g., transformers, diffusion models, generative models, etc.)” where a transformer model is, by definition a combination of models (i.e., an encoder and a decoder); Upadhyay, ¶ [0042], [0149] (Further evidence at Prov. App. No. 63/588,591, [0073])) by submitting contextual information extracted from the user query by a context analyzer, to a service mapper of the context- aware multi-model aggregator (Discloses the Score Prediction Model 112 (service mapper) which “can receive a prompt and optionally additional information (e.g., prompter identifier, context, etc.) as input and generate an output”; Upadhyay, ¶ [0072] (Further evidence at Prov. App. No. 63/588,591, [0047])), receiving suggested combinations of two or more models that may be used to evaluate the user query (“The SPM preferably includes a first set of encoding layers configured to determine an encoding for a prompt and a second set of parallel sets of decoder layers (e.g., one set of decoder layers for each candidate model) configured to determine a score” and each scoring head “outputs a predicted performance score for the respective candidate model (e.g., indicative of the quality of the response that the candidate model is expected to produce responsive to the prompt)”; Upadhyay, ¶ [0071], [0074] (Further evidence at Prov. App. No. 63/588,591, [0055]-[0056])), and selecting one of the suggested combinations of two or more models to evaluate the user query (“Selecting a candidate model S340 functions to identify the best candidate model to use to generate a response (e.g., wherein the best candidate model is the selected model 210). The candidate model can be selected from the set of candidate models corresponding to the set of similar stored prompt representations 410”; Upadhyay, ¶ [0149] (Further evidence at Prov. App. No. 63/588,591, [0056]-[0057]))... the two or more models in the selected one of the suggested combinations…[being in a sequence] (The encoder and decoder models of a transformer model, which may be the combination of two or more models selected as the candidate model, are sequenced, understood as positioned in a specified order, in the selected one of the suggested combinations.; Upadhyay, ¶ [0042]-[0045]); use the sequenced two or more models to process the user query by passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models (The encoder and decoder models of a transformer model processes the user query at the encoder model to produce an encoded output. The encoded output is then provided as the input to the decoder model, the decoder model being the successive model of the two models.; Upadhyay, ¶ [0042]-[0045]); transmit a final output, produced by the sequenced two or more models, to the endpoint device (“A result can optionally be returned to the prompter 10 (e.g., directly, via the system, etc.). A result can include a response from the selected candidate model, an identifier for a selected candidate model, a set of ranked candidate models 200, a message (e.g., “no suitable model could be found”), predicted performance scores, performance scores determined from the reward model 300, metadata, and/or other suitable information or combination of information.”; Upadhyay, ¶ [0160] (Further evidence at Prov. App. No. 63/588,591, [0081])); and perform reinforcement learning based at least in part on feedback received from the endpoint device (“The models can be trained or learned using... reinforcement learning... on: labeled data (e.g., data labeled with the target label), unlabeled data, positive training sets (e.g., a set of data with true positive labels, negative training sets (e.g., a set of data with true negative labels)” and can be “reinforced... based on newly received, up-to-date measurements; past measurements recorded during the operating session; historic measurements recorded during past operating sessions” and training data can include “generating a new set of training data and/or amending an existing set of training data (e.g., based on user metrics)”; Upadhyay, ¶ [0095]-[0096], [0100] (Further evidence at Prov. App. No. 63/588,591, [0074], [0082]-[0083])). However, Upadhyay fails to expressly recite the active step of “sequencing” said models.
The relevance of Chen is described above with relation to claim 1. Regarding claim 17, Chen teaches sequence the two or more models in the selected one of the suggested combinations (Discloses the use of a “Model fine-tuning” which “uses expensive LLMs’ responses to fine-tune cheap LLMs.” As shown in FIG. 2(d), the prompt is received by GPT-4, where the output of GPT-4 is fed to GPT-J, to fine-tune GPT-J to provide the final answer to the query. It is noted that this is both a selection of models from a set of available models and a sequencing of that selection.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include sequencing the two or more models in the selected one of the suggested combinations. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 18, the rejection of claim 17 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the user query is evaluated using a context analyzer of the context-aware multi-model aggregator, (The system discloses a context analyzer in the form of the encoder 20/similarity model 111 (e.g., the “encoder 20 “ can be “integrated within the similarity model 111 and/or separate from the similarity model 111”) that analyzes (“the router 100 includes a similarity model 111 configured to determine a similarity of the prompt to another prompt”) and encodes (the encoder is “configured to encode the prompt into a prompt encoding”) the prompt.; Upadhyay, ¶ [0035], [0062], [0114] (Further evidence at Prov. App. No. 63/588,591, [0018], [0022])). However, Upadhyay fail(s) to expressly recite wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order.
The relevance of Chen is described above with relation to claim 1. Regarding claim 18, Chen teaches wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, (The “Model fine-tuning” includes combining two selected models, being GPT-4 and GPT-J, in the specific order shown in FIG. 2(d). This specific order and combination allows GPT-4 to produce an output that, when used as an input at GPT-J, fine-tunes GPT-J to provide the final answer to the query.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)) wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order (The specific order and combination disclosed at FIG. 2(d) allows GPT-4 to produce an output which is received by GPT-J and used for fine tuning, where GPT-J is the successive one of the two or more models.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Regarding claim 19, the rejection of claim 17 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. Upadhyay further discloses wherein the selected one of the suggested combinations of models includes publicly available models and private models (Discloses candidate models that are third party (publicly available), such as “OpenAI” and “Anthropic” models, and locally executed/user specific (private). “Each candidate model can be executed by the user (e.g., on a user device)” which is understood as a local version of the available models, where a local version of a model is a private model.; Upadhyay, ¶ [0041]-[0042] (Further evidence at Prov. App. No. 63/588,591, [0029]; App. No. 63/532,199, pg. 1, para. 5)).
Regarding claim 21, the rejection of claim 17 is incorporated. Upadhyay and Chen disclose all of the elements of the current invention as stated above. However, Upadhyay fail(s) to expressly recite wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order.
The relevance of Chen is described above with relation to claim 1. Regarding claim 21, Chen teaches wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, (The “Model fine-tuning” includes combining two selected models, being GPT-4 and GPT-J, in the specific order shown in FIG. 2(d). This specific order and combination allows GPT-4 to produce an output that, when used as an input at GPT-J, fine-tunes GPT-J to provide the final answer to the query.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)) wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order (The specific order and combination disclosed at FIG. 2(d) allows GPT-4 to produce an output which is received by GPT-J and used for fine tuning, where GPT-J is the successive one of the two or more models.; Chen, ¶ Pg. 4, FIG. 2(a)-(e)).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the prompt routing systems of Upadhyay to incorporate the teachings of Chen to include wherein the sequencing the two or more models in the selected one of the suggested combinations to process the user query includes: combining the two or more models in the selected one of the suggested combinations in a specific order, wherein using the sequenced two or more models to process the user query includes passing outputs produced by the respective two or more models as inputs to successive ones of the two or more models according to the specific order. The generative model cascading systems of Chen can learn which combinations of generative models “to use for different queries in order to reduce cost and improve accuracy” which Chen shows, through experimentation “can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.” (Chen, ¶ Abstract). One having ordinary skill in the art would be motivated to combine the teachings of Upadhyay and Chen to achieve a more cost efficient contextual prompt routing without sacrificing quality of results, as recognized by Chen. (Id.).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Non-patent literature to Zhang (Zhang, S., Chen, Z., Chen, S., Shen, Y., Sun, Z. and Gan, C., 2024. Improving reinforcement learning from human feedback with efficient reward model ensemble. arXiv preprint arXiv:2401.16635v1.) discloses a novel approach to enhancing the alignment of large language models through efficient reward model ensemble in RLHF, including using linear-layer ensemble and LoRA-based ensemble.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Sean E. Serraguard whose telephone number is (313)446-6627. The examiner can normally be reached 07:00-17:00 M-F.
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, Daniel C. Washburn can be reached at (571) 272-5551. 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 E Serraguard/Primary Examiner, Art Unit 2657
Prosecution Insights