DETAILED ACTION
This examination is in response to the communication filed on 03/31/2024. Claims 1-20 are currently pending, where claims 1 and 11 are independent.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 03/31/2024 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 9 is rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 9 recites the limitation "the classification algorithm" in line 1. There is insufficient antecedent basis for this limitation in the claim. For purposes of Examination, this term is interpreted as being “the classification model”.
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 1-7, 9-11 and 17-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea and/or mathematical algorithm without significantly more. Independent claims 1 and 11 recite “receiving the input from a user”, “applying a token classification model to the input to generate a replacement dictionary”, “applying a classification model to the input to classify at least one or a nature or a structure of the input”, “updating the input based on the replacement dictionary”, “identifying, based on the classified nature or the structure of the input, at least one large language model”, “converting the input in view of the at least one large model by a trained machine learning model” and “transmitting the converted input and the replacement dictionary to the at least one large language model.”
The limitations of “receiving…”, “applying…”, “applying…”, “updating…”, “identifying…”, “converting…”and “transmitting…” as drafted, are a process that, under a broadest reasonable interpretation, covers the abstract idea of “mental processes” because they cover concepts performed in the human mind, including observation, evaluation, judgement and opinion. See MPEP 2106.04(a)(2). That is, other than reciting “a trained machine learning model”, a “processor” (claim 1), nothing in the claimed elements preclude the steps from practically being performed by a person
receiving the input from a user (e.g., by the person receiving a written or oral request),
applying a token classification model to the input to generate a replacement dictionary (e.g., by the person using simple rules or algorithms to identify elements to be annotated or replaced such as an acronyms, personal information, or domain specific data, i.e., annotating the request; under a broadest reasonable interpretation a “token classification model” corresponds to rules for normalizing input data),
applying a classification model to the input to classify at least one or a nature or a structure of the input (e.g., by the person analyzing the request to determine the type or structure, i.e., a request for information, a command to be processed etc. ; under a broadest reasonable interpretation a “classification model” corresponds to rules for categorizing the request type),
updating the input based on the replacement dictionary (e.g., by the person annotating the request to replace the acronyms, reformat numbers, etc.),
identifying, based on the classified nature or the structure of the input, at least one large language model (e.g., by the person selecting from a plurality of available models, a specific model to handle the request based on the identified request type),
converting the input in view of the at least one large model
(transmitting of data is considered post solution activity ).
This judicial exception is not integrated into a practical application because the additional elements of “a trained machine learning model”, a “processor” (claim 1), are all recited at a high-level of generality, and paragraph [0044] of the Specification describes the use of a general-purpose processor, e.g., CPU. Accordingly, these additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea. In addition, the added limitation of using “a trained model…” is not recited with sufficient specificity as to provide any details about how the trained model operates or how the converting of the input alters the converting of the input beyond a person revising the input using pen and paper. Thus, the claims as a whole are directed to an abstract idea (Step 2A, prong two).
Claims 1 and 11 do not include any additional elements that are sufficient to amount to significantly more than the judicial exception because, as discussed above with respect to integration of the abstract idea into a practical application, the additional elements of a “processor” (claim 1), or a “trained model…” amount to no more than mere instructions to apply the exception using generic computer components. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept (Step 2B).
With respect to dependent claims 2-7, 9-10 and 17-20, these claims are directed the criteria used update or convert the input or select/identify a large language model. These limitations also relate to the abstract idea of “mental processes.” That is nothing in the claimed elements preclude the steps from practically being performed by a person using the recited criteria when performing the respective steps. No additional elements are present.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
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 non-obviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1-7, 9-11, 17-18 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Cuomo et al. (US 2025/0299053 A1; herein “Cuomo”) further in view of Gharibi et al. (US 2025/0103746 A1; herein “Gharibi”) .
Regarding claims 1 and 11, Cuomo teaches a computer-implemented method (Abstract teaches “…provide computer-implemented methods…”) and non-transitory computer readable medium including instructions that, when executed by at least one processor, cause the at least one processor to perform operations for updating an input for at least one large language model (¶[0021] teaches “…there is provided a computer program product that include one or more computer readable storage media and program instructions…”), the operations comprising:
receiving the input from a user (Fig. 3, step 302 “Receive user Prompt” );
applying a classification model to the input to classify at least one of a nature or a structure of the input (Fig. 3, step 308 “Categorize prompt” and ¶[0071] teaches “…virtualization manager 110 categorizes the enriched user prompt into specific domains and styles.” );
updating the input based on the replacement dictionary (¶¶[0058]-[0059] teaches “…prompt tailoring engine 120 can include one or more tailoring algorithms …responsible for adapting the user’s query to better match the specific characteristics of the selected LLM…adaptions can include…rephrasing the query entirely…prompt tailoring engine 120 can include a pre-prompt repository (not shown). The pre-prompt repository is a database containing pre-engineered prompts or segments of prompts…Virtualization manager 110 can utilize this repository to speed up the tailoring process by providing ready-to-use templates or building blocks” Under a broadest reasonable interpretation “replacement dictionary” includes prompt templates or building blocks generated for respective LLMs);
identifying, based on the classified nature or the structure of the input, at least one large language model (Fig. 3, step 310 “Tailor Prompt for Specified LLM” and ¶[0072] teaches “In step 310, virtualization manager 110 tailors prompt for specified LLM.” );
converting the input in view of the at least one large language model by a trained machine learning model (Fig. 3, step 310 “Tailor Prompt for Specified LLM” and ¶[0072] teaches “…virtualization manager 110 tailors prompts based on specific characteristics and API requirements for a specified LLM…” and ¶[0075] teaches “In step 312, virtualization manager 110 optimizes prompt format…by formatting the prompt to align with the specific API requirements of the chosen LLM” ); and
transmitting the converted input and the replacement dictionary to the at least one large language model (Fig. 3, step 314 “xxx” and ¶[0076] teaches “In step 314… virtualization manager 110 evaluates the response generated by the LLM, processes the optimized prompt, and evaluates the quality and relevance of the generated response”. Evaluation of the LLM response inherently requires transmitting the enriched prompt to the chosen LLM ).
Although Cuomo teaches enriching the user prompt and optimizing the prompt format (See Fig. 3 and ¶¶[0072]-[0075]) utilizing, for example, a dictionary of stored templates, Cuomo fails to explicitly disclose applying a token classification model to the input to generate a replacement dictionary.
Gharibi teaches systems and methods for providing privacy-aware LLM prompting, wherein the system includes a pre-processing model for redacting and/or augmenting a prompt (Abstract). More specifically, Gharibi teaches applying a token classification model to the input to generate a replacement dictionary (¶[0055] [provisional Appendix A] teaches “The pre-processing step utilizes a combination of machine learning models and regex expressions to identify sensitive information in the prompt…can perform this task using a transformer-encoder architecture with the head of the neural network performing classification per token”; ¶[0056] [provisional Appendix A] teaches “A model and/or an ensemble of models…are capable of classifying named entities into over thirty categories of sensitive information…accept a user prompt or any other text and produce probability distribution for each token, indicating its likelihood of belonging to a specific category of sensitive information”; and ¶[0059] [provisional Appendix A] teaches “Step 2 involves crating mapping dictionaries. The system includes a dictionary that maps every term (or token) classified as sensitive information to a placeholder or token” ).
Cuomo differs from the claimed invention, as defined in claims 1 and 11, in that Cuomo fails to explicitly disclose applying a token classification model to the input to generate a replacement dictionary. Applying token classification models to prompt input to generate a replacement dictionary is known in the art as evidenced by Gharibi. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have modified the system of Cuomo to include applying a token classification model to the input prompts to generate a replacement dictionary as taught by Gharibi in order “to manage the privacy risks associated with prompting any machine learning model” (Gharibi, ¶[0006]).
Regarding claims 2 and 18, the combination of Cuomo and Gharibi teaches all of the elements of claims 1 and 11 (see detailed element mapping above). In addition, Cuomo further teaches the operations further comprise:
identifying, based on the classified nature or the structure of the input, a large language model from the at least one large language model (¶[0082] teaches “…virtualization manage 110 automatically selects model using one or more algorithms to select an LLM that best aligns with the prompt’s requirements, informed by the…prompt’s categorization”); and
transmitting the updated input to the identified large language model (Fig. 4, steps 408 “Integrate Feedback Loop” and ¶[0083] teaches “In step 408, virtualization manager 110 integrates the selection into the feedback loop…user feedback directly from the user regarding their satisfaction with the response…” Rating satisfaction of the response necessarily requires the prompt to be provided to the identified LLM for processing).
Regarding claim 3, the combination of Cuomo and Gharibi teaches all of the elements of claim 1(see detailed element mapping above). In addition, Cuomo further teaches the operations further comprise converting the input into a text format (¶[0058] teaches “…prompt tailoring engine 120 can include one or more tailing algorithms…adaptation can include…rephrasing the query entirely” ).
Regarding claim 4, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). In addition, Gharibi further teaches the replacement dictionary comprises one or more classified entities associated with the input (¶[0056] teaches “A model and/or ensemble of models …that are capable of classifying named entities in to over thirty categories of sensitive information (e.g., name, age, IP address, gender, etc.)…” ).
Cuomo differs from the claimed invention, as defined in claims 1 and 11, in that Cuomo fails to explicitly disclose applying a token classification model to the input to generate a replacement dictionary. Applying token classification models to prompt input to generate a replacement dictionary is known in the art as evidenced by Gharibi. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have modified the system of Cuomo to include applying a token classification model to the input prompts to generate a replacement dictionary as taught by Gharibi in order “to manage the privacy risks associated with prompting any machine learning model” (Gharibi, ¶[0006]).
Regarding claims 5 and 17 , the combination of Cuomo and Gharibi teaches all of the elements of claims 1 and 11 (see detailed element mapping above). In addition, Cuomo further teaches the operations further comprise identifying from a plurality of trained machine learning models the trained machine learning model for updating the input based on the classified nature or the structure of the input (¶[0057] teaches “Prompt tailoring engine 120 includes one or more natural language understanding algorithms to refine and customize user prompts…” and ¶[0058] teaches “…prompt tailoring engine 120 can utilize tailorizing algorithms to apply various adaptations…adaptations can include adding few-shot examples, modifying the tone, or even rephrasing the query entirely” ).
Regarding claim 6, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). In addition, Cuomo further teaches converting the input for the at least one large language model comprises updating the input in view of at least one of a summarization related task (¶[0054] teaches “Model behavior database 118 serves as a repository containing meta information about various Large Language Models (LLMs) available for selection…For example, meta information can include information that identifies models fine-tuned for specific tasks such as summarization, sentiment analysis, or technical coding”; the “or” makes the additionally listed tasks optional).
Regarding claim 7, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). In addition, Cuomo further teaches the input comprises at least one of a prompt (Fig. 3, step 302 “Receive User Prompt” ), a recorded session, an audit log, a policy, a code snippet, or a computer file (The “or” makes the additional elements optional).
Regarding claim 9, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). In addition, Cuomo further teaches the classification algorithm identifies a structure or a nature of the input and a corresponding large language model (¶[0071] teaches “…virtualization manager 110 categories the enriched user prompt into specific domains and styles…based on content, style, and complexity identified from natural language understating”).
Regarding claim 10, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). In addition, Cuomo further teaches the nature of the input comprises a task type of the input (¶[0068] teaches “In this embodiment as user prompt can include a user request to perform a task”).
Regarding claim 20, the combination of Cuomo and Gharibi teaches all of the elements of claim 11 (see detailed element mapping above). In addition, Gharibi further teaches replacing a value of the input with a variable from the replacement dictionary (¶[0059] teaches “Step 2 involves creating mapping dictionaries. The system includes a dictionary that maps every term (or token) classified as sensitive information to a placeholder or token. For example, "John" will be mapped to "Name_l" and "New York" will be mapped to "Location_]". This dictionary enables the separation of different named entities in the prompt to maintain their context and facilitate re-introduction of the redacted terms in the post-processing step via the second privacy monitor 104B.”).
Cuomo fails to explicitly disclose applying a token classification model to the input to generate a replacement dictionary. Applying token classification models to prompt input to generate a replacement dictionary is known in the art as evidenced by Gharibi. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have modified the system of Cuomo to include applying a token classification model to the input prompts to generate a replacement dictionary as taught by Gharibi in order “to manage the privacy risks associated with prompting any machine learning model” (Gharibi, ¶[0006]).
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Cuomo and Gharibi as applied to claim 1 above, and further in view of Bengio et al. (US 2016/0140435 A1; herein “Bengio”).
Regarding claim 8, the combination of Cuomo and Gharibi teaches all of the elements of claim 1 (see detailed element mapping above). Although the combination of Cuomo and Gharibi discusses machine learning models, the combination of Cuomo and Gharibi fails to explicitly disclose utilizing a sequence-to-sequence model with an encoder/decoder structure using LSTM layers.
Bengio teaches a method and system for generating natural language descriptions that utilizes a sequence-to-sequence model with an encoder-decoder neural network architecture using long short-term memory layers (Fig. 2 and ¶[0031]).
The combination of Cuomo and Gharibi differs from the claimed invention, as defined in claim 8, in that the combination fails to explicitly disclose utilizing a sequence-to-sequence modeling including an encoder, decoder and LSTM layers as claimed. Sequence-to-sequence models having an encoder, decoder, and LSTM structure are known in the art as evidenced by Bengio. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo and Gharibi utilizing a trained encoder/decoder structured machine learning model as taught by Bengio as it merely constitutes the substitution of known machine learning structures to achieve the predictable result of generating modifying prompt language utilizing a machine learning model.
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Cuomo and Gharibi as applied to claim 11 above, and further in view of Mukherjee et al. (US 2025/0294091 A1; herein “Mukherjee”).
Regarding claim 12, the combination of Cuomo and Gharibi teaches all of the elements of claim 11 (see detailed element mapping above). Although the combination of Cuomo and Gharibi discuss transformer based machine learning models, the combination of Cuomo and Gharibi fails to explicitly disclose that the transformer model includes encoder/decoder structure or how the machine learning model is trained.
Mukherjee teaches generative and adaptive encoder/decoder structured machine learning mediator model for real-time interactions with conversational agents which includes:
transmitting the input to a tokenization model (Fig. 6A, positional encodings 606 );
transmitting a tokenized input to a trained embedding model (Fig. 6A, element 604);
receiving an embedded input sequence from the trained embedding model (Fig. 6A, output from input embedding 604 );
transmitting the embedded input sequence to an encoder (Fig. 6A, Encoder 608 );
receiving a context vector from the encoder (Fig. 6A, output from encoder blocks 610a…610c );
transmitting the context vector to a decoder (Fig. 6A, Decoder 612 );
receiving a decoder output from the decoder (Fig. 6A, output from decoder blocks #1-#M ); and
evaluating the updated input (¶[0167] teaches “…an error/loss function is generated by comparing the output from the ML model 710 with the labels 704…”).
The combination of Cuomo and Gharibi differs from the claimed invention, as defined in claim 12, in that the combination fails to explicitly disclose that the trained machine learning model is an encoder/decoder structured model as claimed. Conversation machine learning models which determine intent of user prompts which utilize a trained encoder/decoder structure as known in the art as evidenced by Mukherjee. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo and Gharibi utilizing a trained encoder/decoder structured machine learning model as taught by Mukherjee as it merely constitutes the substitution of known machine learning structures to achieve the predictable result of determining the intent/type of the user prompt for use in selection of particular LLM to process the prompt.
Claims 13 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Cuomo, Gharibi and Mukherjee as applied to claim 12 above, and further in view of Chhabra et al. (JP 2019200551; herein “Chhabra”).
Regarding claim 13, the combination of Cuomo, Gharibi and Mukherjee teaches all of the elements of claim 12 (see detailed element mapping above). In addition, Mukherjee further teaches
transmitting a target sequence to a tokenization model (Fig. 6A, positional encodings 606);
transmitting a tokenized target sequence to a trained embedding model (Fig. 6A, positional encodings 606);
receiving an embedded target sequence from the trained embedding model (Fig. 6A, output from input embedding 604);
determining a similarity between the decoder output and the embedded target sequence (¶[0167] teaches “an error/loss function is generated by comparing the output from the ML model 710 with the labels 704. The coefficients of the ML model 710 are iteratively updated to reduce an error/loss function. The value of the error/loss function decreases as outputs from the ML model 710 increasingly approximate the labels 704.” );
generating a loss based on the similarity (¶[0167] teaches “an error/loss function is generated by comparing the output from the ML model 710 with the labels 704. The coefficients of the ML model 710 are iteratively updated to reduce an error/loss function. The value of the error/loss function decreases as outputs from the ML model 710 increasingly approximate the labels 704.”).
The combination of Cuomo, Gharibi and Mukherjee fails to specifically disclose generating a length loss between the decoder output and the embedded target sequence; generating a total loss score based on the loss and the length loss; and computing a gradient of the total loss score with respect to parameters of the trained machine learning model.
Chhabra teaches generating a length loss between the decoder output and the embedded target sequence (Trans. p. 6, 4th ¶ teaches “The neural network environment is trained with a combination of weighted code length loss for embedding (i.e., a set of latent codes) and reconstruction loss for output (i.e., a set of reconstructed data). The weighted code length loss generates a hierarchical order of importance with weighting. There are no restrictions on the architecture” );
generating a total loss score based on the loss and the length loss (Trans. page 8, last ¶ teaches “Total loss is the sum of reconstruction loss and code loss (ie, loss due to generating latent code from noise data” ); and
computing a gradient of the total loss score with respect to parameters of the trained machine learning model (Trans. p. 8, 1st ¶ teaches Here, the term L.sub.aux may include quasi-gradient loss and regularization loss for model weighting to aid optimization, which is the main objective. Thus, the training efficiency of machine learning can be improved by generating the latent code by imposing a loss function on the input data.” ).
The combination of Cuomo, Gharibi and Mukherjee differs from the claimed invention, as defined in claim 13, in that the combination fails to explicitly disclose that the utilizing length loss as part of total loss as claimed. Training machine learning models utilizing both length loss and total loss gradient is known in the art as evidenced by Chhabra. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo, Gharibi and Mukherjee by utilizing both length loss and a similarity loss as taught by Chhabra as it merely constitutes the substitution of known machine learning processes to achieve the predictable result of optimizing both the prompt length and similarity.
Regarding claim 14, the combination of Cuomo, Gharibi, Mukherjee and Chhabra teaches all of the elements of claim 13 (see detailed element mapping above). In addition, Mukherjee further teaches backpropagating the total loss score to adjust the machine learning model parameters (¶[0167] teaches “the ML model 170… is trained via supervised learning using a backpropagation technique to train the weighting parameters between nodes within respective layers of the ANN”).
The combination of Cuomo and Gharibi differs from the claimed invention, as defined in claim 14, in that the combination fails to explicitly disclose backpropagating the total loss score to adjust machine learning model parameters as claimed. Utilizing backpropagation to adjust model parameters is known in the art as evidenced by Mukherjee. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo and Gharibi utilizing a trained encoder/decoder structured machine learning model as taught by Mukherjee as it merely constitutes the substitution of known machine learning structures to achieve the predictable result of determining the intent/type of the user prompt for use in selection of particular LLM to process the prompt.
Claims 15 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Cuomo and Gharibi as applied to claim 11 above, and further in view of Mukherjee, still further in view of Tiong et al. (US 2023/0419652 A1; herein “Tiong”).
Regarding claim 15, the combination of Cuomo and Gharibi teaches all of the elements of claim 11 (see detailed element mapping above). Although the combination of Cuomo and Gharibi discuss transformer based machine learning models, the combination of Cuomo and Gharibi fails to explicitly disclose that the transformer model includes an encoder/decoder structure or how the machine learning model is updated.
Mukherjee teaches generative and adaptive encoder/decoder structured machine learning mediator model for real-time interactions with conversational agents which includes:
transmitting the input to a tokenization model (Fig. 6A, positional encodings 606);
transmitting the tokenized input to a trained embedding model (Fig. 6A, positional encodings 606);
receiving an embedded input sequence from the trained embedding model (Fig. 6A, output from input embedding 604);
transmitting the embedded input sequence to an encoder (Fig. 6A, Encoder 608);
receiving a context vector from the encoder (Fig. 6A, Decoder 612); and
iterating the context vector from the encoder to receive a probability distribution from a decoder (Fig. 6A, Output Probabilities 620).
The combination of Cuomo and Gharibi differs from the claimed invention, as defined in claim 15, in that the combination fails to explicitly disclose that the trained machine learning model is an encoder/decoder structured model as claimed. Conversation machine learning models which determine intent of user prompts which utilize a trained encoder/decoder structure as known in the art as evidenced by Mukherjee. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo and Gharibi utilizing a trained encoder/decoder structured machine learning model as taught by Mukherjee as it merely constitutes the substitution of known machine learning structures to achieve the predictable result of determining the intent/type of the user prompt for use in selection of particular LLM to process the prompt.
Although Cuomo teaches the updated input may include a rephrasing of the user prompt, the combination of Cuomo, Gharibi and Mukherjee fails to explicitly disclose that the input update/modification includes sampling a word from the probability distribution to generate the updated input.
Tiong teaches methods of generating text utilizing stochastic decoding, i.e., sampling a word from a probability distribution. More specifically, ¶[0059] teaches “…where a caption is generated based on the subset K’ sampled image patches 308 (e.g., using stochastic decoding in the text decoder of image captioning module 132).”
The combination of Cuomo, Gharibi, and Mukherjee differs from the claimed invention, as defined in claim 15, in that the combination fails to explicitly disclose that the trained machine learning model samples a word probability distribution to generate the updated user input/prompt. Utilizing stochastic decoding, i.e., sampling a word probability, when generating prompts is known in the art as evidenced by Tiong. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the system taught by the combination of Cuomo, Gharibi and Mukherjee utilizing utilize stochastic decoding when rephrasing the user prompt as taught by Tiong as it merely constitutes the substitution of known machine learning structures to achieve the predictable result of preventing the generation of fixed or rigid prompt patterns.
Regarding claim 16, the combination of Cuomo, Gharibi, Mukherjee and Tiong teaches all of the elements of claim 15 (see detailed element mapping above). In addition, Cuomo further teaches converting the updated input into a format readable by the at least one large language model (¶[0016] teaches “tailoring the received user prompts for a respective LLM…can further include formatting the tailored user prompts to align with the respective LLM.” ).
Claim 19 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Cuomo and Gharibi as applied to claim 11 above, and further in view of Fu et al. (CN 117370638; herein “FU”).
Regarding claim 19,the combination of Cuomo and Gharibi teaches all of the elements of claim 11 (see detailed element mapping above). However, the combination fails to disclose prompt decomposition, e.g., transmitting a first portion of the input to a first large language model and transmitting a second portion of the input to a second large language model.
FU teaches a method and device for decomposing and transmitting prompts that includes splitting the complex task into multiple sub-task and obtaining the multi-link context of each sub-task. More specifically, FU teaches transmitting a first portion of the input to a first large language model and transmitting a second portion of the input to a second large language model (splitting the complex task into multiple sub-tasks through the large language model, and obtaining the multi-link context of each sub-task in the thinking diagram; sequencing the priority of each sub-task to generate a task queue; distributing the task processing model for each sub-task based on the multi-link context; analyzing the resource depended by each task processing model processing sub-task, processing the sub-task according to the priority, and processing the sub-task not depended on the resource in the task queue in para).
The combination of Cuomo and Gharibi differs from the claimed invention, as defined by claim 19, in that the combination fails to explicitly disclose decomposing the prompt into sub-tasks and transmitting the sub-tasks for processing. Prompt decomposition is known in the art as evidenced by FU. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have modified the system taught by the combination of Cuomo and Gharibi to include prompt decomposition as taught by FU as it merely constitutes the combination of known processes to achieve the predicable result of preventing “attention drift” for user prompts including complex or multiple tasks/intents.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to PENNY L CAUDLE whose telephone number is (703)756-1432. The examiner can normally be reached M-Th 8:00 am to 5:00 pm eastern.
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 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.
/PENNY L CAUDLE/Examiner, Art Unit 2657