DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Specification
The disclosure is objected to because of the following informalities:
In ¶[0028], “specialist module 220d” should be “specialist model 220d”. See Figure 2.
In ¶[0029], “selection model 204” should be “selection module 204”. See Figure 2.
In ¶[0031], “suitable for execution each step” should be “suitable for execution of each step” or “suitable for executing each step”.
In ¶[0039], “related model accuracy” should be “related to model accuracy”.
In ¶[0041], “step 306 can be repeated” should be “step 304 can be repeated” because it is the training that is repeated after testing the model in step 306 of Figure 3.
In ¶[0047], “is needed from use 108” should be “is needed from user 108”.
In ¶[0049], “approaches 508i and 508j” are not illustrated in Figure 5, but could be “approaches 508a and 508c”.
In ¶[0050], “Specific” should be “Specifically”.
In ¶[0050], “descriptor 220a-n” should be “descriptor 222a-n”. See Figure 2.
In ¶[0051], “to steps 504” appears that it should be “to steps 506” because reference numeral 504 corresponds to a plan in Figure 5.
In ¶[0052], “solicit addition information” should be “solicit additional information”.
In ¶[0054], “step output 212” should be “step output 612”. See Figure 6.
In ¶[0056], Figure 6 includes model 210b, which is not described in the Specification. Here, model 210b should have some description in the Specification, if this can be done without introducing new matter, or model 210b should be canceled from Figure 6.
In ¶[0057], Figure 7 includes Step 712 to assemble step output, but there is no description of this step in the Specification. Here, Step 712 should be described in the Specification, if this can be done without introducing new matter, or Step 712 should be cancelled from Figure 7.
Appropriate correction is required.
Election/Restrictions
Applicants’ election without traverse of Invention I, Claims 1 to 12, in the reply filed on 03 February 2026 is acknowledged.
Claims 13 to 20 are withdrawn from further consideration pursuant to 37 CFR 1.142(b) as being drawn to a nonelected invention, there being no allowable generic or linking claim. Election was made without traverse in the reply filed on 03 February 2026.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1 and 3 to 8 are rejected under 35 U.S.C. 103 as being unpatentable over Pack, III et al. (U.S. Patent Publication 2025/0086011) in view of Mandal et al. (U.S. Patent No. 11,200,885).
Concerning independent claim 1, Pack, III et al. discloses a method of automation with composable asynchronous tasks of a prompt, comprising:
“decomposing the compound prompt into a plan having a plurality of steps via a planner module instantiated in machine-readable memory and operable by a processor” – a prompt may be received at a computing device; using a first large language model (LLM); composable asynchronous tasks may be determined from the prompt (“a compound prompt”) (Abstract); agent LLM may decompose the prompt into any number of tasks (“decomposing the compound prompt into a plan having a plurality of steps”) and may generate text-based prompts to be input into LLMs used by any of the selected composable asynchronous tasks (¶[0012]); agent LLM 110 may be any suitable combination of hardware and software of computing device 100 for running a large language model; agent LLM 110 may generate input data that may input any of the composable asynchronous tasks including prompts that may be used as input to other LLMs (¶[0036]: Figure 1); storage 170 may store composable asynchronous tasks (¶[0038]: Figure 1); agent LLM 110 (“a planner module”) may decompose a larger job posed by a prompt into any suitable number of smaller jobs that may correspond to the capabilities of composable asynchronous tasks 180 (¶[0044]: Figure 1); computer 20 includes central processor 24 and memory 27 (“instantiated in machine-readable memory and operable by a processor”) (¶[0052] - ¶[0054]: Figure 6); here, agent LLM 110 is “a planner module” that generates “a plan having a plurality of steps” comprising asynchronous tasks represented by sub-prompts; that is, each task is a step in “a plan” generated by agent LLM 110;
“associating a domain descriptor to each of a plurality of distinct machine learning models” – machine learning models 120 may be any suitable machine learning models used by composable asynchronous tasks including large language models; tools used by composable asynchronous tasks 180 and any sub-tasks may include machine learning models 120 including LLMs (“a plurality of distinct machine learning models”) (¶[0037] - ¶[0038]); LLM task 201 may be a call into a specified LLM; a specified LLM task 201 may call the specified LLM using an API of the specified LLM (¶[0039]: Figure 4); a first sub-prompt may include a request to generate a gallery container component, a second sub-prompt may include a request to generate a gallery item component, a third sub-prompt may include a request to generate a gallery image component, and a fourth sub-prompt may include a request to generate a gallery user interface component (¶[0044]: Figure 4); here, an API call for a large language model provides “a domain descriptor to each of the plurality of distinct machine learning models”; Compare Specification, ¶[0028], which describes specialist models 220 for specific domains as being accessed through APIs;
“selecting a subset of the plurality of distinct machine learning models [based on the respective model relevance score] of each of the plurality of distinct machine learning models” – sub-outputs of the composable asynchronous tasks 182, 184, 186, and 188 may be sent as inputs to suitable LLMs from machine learning models 120 (¶[0046]: Figure 4); composable asynchronous task 182 may, upon being called by agent LLM 110 with first sub-prompt, use the first sub-prompt as input to an appropriate LLM from machine learning models 120; composable asynchronous tasks 184, 186 may operate similarity when called by agent LLM 110, each using suitable LLMs from machine learning models 120 (¶[0047]: Figure 4); here, agent LLM 110 is “selecting a subset of the plurality of distinct machine learning model” by determining ‘suitable’ and ‘appropriate’ LLMs to perform tasks 182, 184, 186;
“generating a step output addressing the step from the selected subset of the plurality of distinct machine learning models” – specified LLM task 201 may call into the specified LLM and receive the output of the specified LLM; output generated by the specified LLM after being called by LLM task 201 may be potential output for the composable asynchronous task 182 (¶[0039]: Figure 2); return code 305 may be code that returns output from a composable asynchronous task; output may be returned to agent LLM 110 (¶[0040]: Figure 3); each sub-prompt may be suitable for input to an LLM or generative machine learning model that may request the generation of output that may be used in order to generate the output requested in the prompt received by agent LLM 110 (¶[0044]: Figure 4); agent LLM 110 may receive output from composable asynchronous tasks, receiving first sub-output from composable asynchronous task 182, second sub-output from composable asynchronous task 184, third sub-output from composable asynchronous task 186; composable asynchronous tasks 182, 184, 186, and 188 may perform any sub-tasks on first LLM output, second LLM output, third LLM output, and fourth LLM output to generate sub-outputs that may be sent back to agent LLM 110 (¶[0046]: Figure 4);
“assembling all of the step outputs into a syntactically and semantically coherent final output via an integration module utilizing a large language model” – a composable asynchronous task may be generated to handle generation of product descriptions for an e-commerce website; a desired output format for output from composable asynchronous tasks may be a product description for each product that follows a consistent structure starting with a headline followed by a brief overview of the product and then a detailed list of features and benefits of the product (¶[0029]); once agent LLM 110 has executed and received sub-outputs from all of the composable asynchronous tasks, agent LLM 110 may generate and send a final output that may be responsive to the request to the prompt received by agent LLM 110; final output may use the sub-outputs received from composable asynchronous tasks called by agent LLM 110; agent LLM 110 may use all of the sub-outputs as part of the final output, or use only a single sub-output if that sub-output already includes the other sub-outputs (¶[0048]: Figure 4); here, agent LLM 110 may generate a final output from all of the sub-outputs so as to provide “assembling all of the step outputs into a syntactically and semantically coherent final output via an integration model utilizing a large language model”; that is, agent LLM 110 is “an integration module utilizing a large language model” that ‘assembles’ all of the sub-outputs of tasks; implicitly, generating a product description that incorporates a desired output format in a consistent structure provides “syntactically and semantically coherent final output”.
Concerning independent claim 1, Pack, III et al. discloses a general concept of decomposing a prompt into a plurality of sub-prompts for machine learning models and assembling outputs from the machine learning models. Pack, III et al. discloses determining suitable and appropriate machine learning models corresponding to sub-prompts of tasks, but does not select a model based on a score in the limitation of “for each of the plurality of steps: generating a model relevance score by semantic comparison of the step to the domain descriptor of each of the plurality of distinct machine learning models”, and selecting a subset of the plurality of distinct machine learning models “based on the respective model relevance score”.
Concerning independent claim 1, Mandal et al. teaches analogous art of a dialog manager that selects one or more dialog model to produce text data by ranking the models in an N-best list and selecting the top-scoring output. (Abstract) Dialog focus data 316 may be used by dialog model predictor 320 to select one or more dialog models 314 in dialog model storage 330. Each dialog model may be associated with one or more categories of functions (“a domain”). Dialog model predictor 320 may create an N-best list by determining a score for each dialog model 314 given a dialog from focus data 316 and model data 324. Model data 324 may include a type of each dialog model 314 and APIs and corresponding entities for each dialog model 314 (“a domain descriptor”). Dialog model predictor may determine the score based on presence or absence of one or more entities determined by entity chunker 312 in model data 324 and a presence of an entity in a list of entities corresponding to a dialog model 314 (“by semantic comparison”). The dialog model predictor 320 may send input text data 302 to models 314 having the N highest scores greater than a threshold. Action selector 318 selects at least one of the outputs for further processing, and may determine a score for each of the outputs of the dialog models 314 based on each’s similarity or relevance to dialog focus data 316 (“generating a model relevance score by semantic comparison”). Natural language generator 326 may be used to generate output text data 306. Inference engine 402 may be used to generate scores for one or more outputs of a single dialog model and scores may correspond to a similarity or correlation between outputs of dialog model 404 to dialog focus data 416. Inference engine 402 determines and sends a request for N-best models 512 to one or more dialog models and returns to inference engine 402 the N-best models 514. (Column 10, Line 62 to Column 12, Line 19: Figures 3 to 5A and 5B) Mandal et al., then, teaches selecting a subset of models based on “a model relevance score” and that a model has “a domain descriptor” based on a presence of entities in an entity list associated with that model. An objective is to select a different dialog model for each turn of a dialog to thereby respond accurately to each goal expressed by a user. (Column 2, Lines 55 to 61) It would have been obvious to one having ordinary skill in the art to select a suitable and appropriate machine learning model in Pack, III et al. based on a relevance score of a model that is identified by a domain descriptor in Mandal et al. for a purpose of selecting different dialog models for each turn of a dialog to accurately respond to a plurality of goals expressed by a user.
Concerning claim 3, Mandal et al. teaches that dialog manager 260 determines a goal corresponding to an action that a user desires to be performed to execute an intent, e.g., weather or turning off the lights. (Column 6, Lines 17 to 28: Figure 1) Dialog focus data 316 uses encoders to encode some or all of received data into one or more feature vectors and a decoder to determine based on the feature vectors intent data corresponding to an intent of a user. (Column 10, Lines 53 to 57: Figure 3) Determining a similarity or relevance as a score between dialog focus data 316 and model data 324 is based on entities in a list of entities corresponding to a dialog model. (Column 11, Lines 3 to 42: Figure 3) Here, model data 324 is a list of entities describing the model that can be construed as “domain descriptors”. Mandal et al., then, teaches “classifying intent” of a dialog “and scoring similarity of the classified intent with the domain descriptors of each of the plurality of distinct machine learning models.”
Concerning claims 4 to 5, Mandal et al. teaches that dialog model predictor 320 may send input text data 302 to models 314 having the N highest scores greater than a threshold; a threshold may be a numerical value (“selecting those of the plurality of distinct machine learning models having associated model relevance scores above a threshold value”) or a number N of models 314 to be selected (“selecting those of the plurality of distinct machine learning models having the highest associated model relevance scores among the plurality of distinct machine learning models”). (Column 11, Lines 13 to 19: Figure 3).
Concerning claim 6, Pack, III et al. discloses that once agent LLM 110 has executed and received sub-outputs from all of the composable asynchronous tasks, agent LLM 110 may generate and send a final output that may be responsive to the request to the prompt received by agent LLM 110; final output may use the sub-outputs received from composable asynchronous tasks called by agent LLM 110; agent LLM 110 may use all of the sub-outputs as part of the final output (“integrating intermediate outputs from all models of the selected subset of distinct machine learning models”), or use only a single sub-output if that sub-output already includes the other sub-outputs (¶[0048]: Figure 4).
Concerning claims 7 to 8, Pack, III et al. discloses that composable asynchronous tasks may be modular and may be executable both serially and in parallel subject to dependencies on the output from other composable asynchronous tasks (“the step output comprises serially traversing all of the selected subset of distinct machine learning models via feed-forward of an initial model output of the one of the selected subset of distinct machine learning models as input into another of the selected subset of distinct machine learning models”) (¶[0013]); first, second, and third composable asynchronous tasks may all run asynchronously, while the fourth composable asynchronous tasks may wait due to its dependency on the outputs of the first, second, and third composable asynchronous tasks (¶[0014]); composable asynchronous tasks selected by an agent LLM may be executed in parallel, sequentially, or conditionally, depending on the requirements and constraints of the problem input to the agent LLM (¶[0030]); Compare Specification, ¶[0055], which describes specialist models being prompted in series as a feed-forward from one model to the next; LLM 110 may determine an order LLMs used by composable asynchronous tasks may be executed including which composable asynchronous tasks may be executed in parallel and which may wait for output from other composable asynchronous tasks (¶[0036]: Figure 1); agent LLM 110 may execute the selected composable asynchronous tasks in any suitable order based on any dependencies between the composable asynchronous tasks (“traversing the selected subset of the distinct machine learning models in multiple orders”).
Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Pack, III et al. (U.S. Patent Publication 2025/0086011) in view of Mandal et al. (U.S. Patent No. 11,200,885) as applied to claim 1 above, and further in view of Ben David et al. (U.S. Patent No. 11,875,123).
Concerning claim 2, Mandal et al. teaches that dialog manager 260 determines a goal corresponding to an action that a user desires to be performed to execute an intent, e.g., weather or turning off the lights. (Column 6, Lines 17 to 28: Figure 1) Dialog focus data 316 uses encoders to encode some or all of received data into one or more feature vectors and a decoder to determine based on the feature vectors intent data corresponding to an intent of a user. (Column 10, Lines 53 to 57: Figure 3) Determining a similarity or relevance as a score between dialog focus data 316 and model data 324 is based on entities in a list of entities corresponding to a dialog model. (Column 11, Lines 3 to 42: Figure 3) Here, model data 324 is a list of entities describing the model that can be construed as “domain descriptors”. Mandal et al., then, teaches “generating the model relevance score comprises evaluating a vector [cosine] similarity between vectorized text of each of the domain descriptors and at least a portion of the step”, but omits “a cosine similarity”. However, cosine similarity is a well-known metric of similarity in machine learning. Specifically, Ben David et al. teaches an advice planner providing an intent generated by a large language model from text that classifies the intent into a domain corresponding to the intent and then generates a plan comprising a set of action logic associated with the domain. Each logic step is an ordered sequence for achieving a desired state, and a plan is forwarded to the large language model. (Abstract) Output of an advice planner can be a vector representation, and an intent can be classified by calculating a similarity metric, e.g., cosine similarities between the output vector and vector representations for the domain. (Column 7, Lines 43 to 53: Figure 3: Step 330) Ben David et al., then, teaches determining an intent and a domain by “evaluating vector cosine similarity”. An objective is to provide advice systems that adapt and learn from user interactions and feedback and that provide comprehensible advice in a user-friendly manner. (Column 1, Lines 26 to 39) It would have been obvious to one having ordinary skill in the art to determine a similarity to perform an intent in Mandal et al. with a cosine similarity as taught by Ben David et al. for a purpose of providing comprehensible advice in systems that adapt and learning from user interactions.
Claims 9 to 10 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Pack, III et al. (U.S. Patent Publication 2025/0086011) in view of Mandal et al. (U.S. Patent No. 11,200,885) as applied to claim 1 above, and further in view of Woolf (U.S. Patent Publication 2021/0342743).
Concerning claims 9 to 10, Pack, III et al. discloses selecting among a plurality of suitable and appropriate machine learning models. Mandal et al. teaches that each domain may correspond to one or more dialog models 314, and one or more candidate domains may be determined corresponding to input text. (Column 8, Lines 45 to 57: Figure 3) However, Pack, III et al. and Mandal et al. do not expressly disclose or teach that “training each of the plurality of distinct machine learning models using different training data specific to its associated domain descriptor” and “wherein at least a subset of the plurality of distinct machine learning models are trained entirely separately from others of the plurality of distinct machine learning models, without overlapping training data.” Still, Woolf teaches machine learning models that encapsulate domain expertise. (Abstract) A primary machine learning model is trained with one or more sets of training data from a database, described as a reference subset. (¶[0076] - ¶[0077]) In an iterative fashion, a user-directed machine learning model is trained with a separate set of training data from a database, e.g., a second subset of the database, e.g., non-overlapping, or a separate database, referred to as a second reference subset. (¶[0083]) Woolf, then, teaches “training each of the plurality of distinct machine learning models using different training data” and “wherein at least a subset of the plurality of distinct machine learning models are trained entirely separately from others of the plurality of distinct machine learning models, without overlapping training data.” An objective is to provide model aggregation tools that can easily create and modify machine learning models to collectively improve the models while maintaining privacy of training data. (¶[0006]) It would have been obvious to one having ordinary skill in the art to train a machine learning model to a domain in Mandal et al. using different and non-overlapping training data as taught by Woolf for a purpose of collectively improving machine learning models while maintaining privacy of training data.
Concerning claim 12, Pack, III et al. discloses machine learning models 120 may be any suitable machine learning models used by composable asynchronous tasks including large language models; tools used by composable asynchronous tasks 180 and any sub-tasks may include machine learning models 120 including LLMs (“wherein each of the plurality of distinct machine learning models is a large language model”). (¶[00037] - ¶[0038])
Claims 9 and 11 is rejected under 35 U.S.C. 103 as being unpatentable over Pack, III et al. (U.S. Patent Publication 2025/0086011) in view of Mandal et al. (U.S. Patent No. 11,200,885) as applied to claim 1 above, and further in view of Moradi et al. (U.S. Patent Publication 2025/0013921).
Pack, III et al. discloses selecting among a plurality of suitable and appropriate machine learning models. Mandal et al. teaches that each domain may correspond to one or more dialog models 314, and one or more candidate domains may be determined corresponding to input text. (Column 8, Lines 45 to 57: Figure 3) However, Pack, III et al. and Mandal et al. do not expressly disclose or teach “training each of the plurality of distinct machine learning models using different training data specific to its associated domain descriptor” and that machine learning models “are specialized in a respective domain via fine-tuning or transfer learning.” Still, Moradi et al. teaches machine learning model selection for transfer learning that sends a selected candidate source machine learning model to a target domain and receives fine-tuned model weights for fine-tuned machine learning models. (Abstract) One or more fined-tuned ML models are determined by re-training the one or more source ML models with data samples in a target dataset. (¶[0033]) A target model is trained solely using data samples in a target dataset. (¶[0036]) Process 300 may include a step in which target domain 104 fine-tunes the received source ML models with available data samples in target dataset 114 (“training each of the plurality of distinct machine learning models using different training data specific to the associated domain descriptor”). (¶[0069]: Figure 3) Process 400 may include a step which fine-tunes the received one or more source ML models with available data samples in target dataset 114. (¶[0079]: Figure 4) An objective of transfer learning is in scenarios where not enough data samples are available in a target and/or data collection is expensive or time consuming so that the limited number of target samples are extremely valuable for fine-tuning a source model. (¶[0005]) It would have been obvious to one having ordinary skill in the art to perform fine-tuning or transfer learning as taught by Moradi et al. on machine learning models in Pack, III et al. for a purpose of training a source machine learning model to a target domain in scenarios where not enough data samples are available or data collection is expensive.
Conclusion
The prior art made of record and not relied upon is considered pertinent to Applicants’ disclosure.
Shacham et al. and Schall disclose related prior art.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MARTIN LERNER whose telephone number is (571) 272-7608. The examiner can normally be reached Monday-Thursday 8:30 AM-6:00 PM.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Richemond Dorvil can be reached at (571) 272-7602. 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.
/MARTIN LERNER/Primary Examiner
Art Unit 2658 March 25, 2026
Prosecution Insights