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 .
Response to Arguments
Applicant' s arguments, see pages 7-14, filed on 22 July 2026, with respect to the rejections of claims 1, 10, and 22 under 35 USC § 103 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Cao et al. (US 2023/0108579) for claims 1 and 22 and Cao et al. (US 2023/0108579) in view of Tunstall-Pedoe et al. (US 2023/0274094) for claim 10.
Claim Rejections - 35 USC § 102
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 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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless -
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1, 6-9, 11-13, 22, and 26-29 are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Cao et al. (US 2023/0108579, hereinafter Cao).
Regarding claim 1, Cao discloses
A method comprising:
generating prompt comprising environment information and at least one example task expressed as computer code (paragraph [0042]: the input sequence 102 can include an input prompt from a user, and the one or more prompt entities can include topics important to the user; paragraph [0045]: the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.));
identifying a task to be performed by an agent present in an environment (paragraph [0049]: the prompt entities can identify entities in an environment, the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions, e.g., natural language instructions or other instructions, to the agent to cause the agent to carry out the task); and
providing the prompt (paragraph [0042]: the input sequence 102 can include an input prompt from a user, and the one or more prompt entities can include topics important to the user; paragraph [0045]: the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.)) to at least one previously trained language model to cause the previously trained language model (paragraph [0052]: Prior to using the neural network 110 to generate output sequence 150, the system 100 or another training system trains the neural network 110 in order to cause the neural network 110 to accurately generate output sequence; paragraph [0054]: The training system can perform this training in any of a variety of ways. As one example, the first network blocks in each dual layer can be pre-trained, and then the neural network can be trained with both the first network blocks and the second network blocks included to improve the way in which the neural network handles entity mention; paragraph [0118]: the system fine-tunes the first blocks while training the second blocks) to generate a new plan for performing the task, the new plan comprising a set of operations (paragraph [0045] In another example, the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence. The output sequence 150 can include a general summary of the text sequence, and a respective sub-summary for each of the one or more prompt entities; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.). The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity) performable by the agent within the environment (paragraph [0049]: the prompt entities can identify entities in an environment, the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions, e.g., natural language instructions or other instructions, to the agent to cause the agent to carry out the task).
Regarding claim 6, Cao discloses
wherein the environment information identifies one or more actions performable by the agent (paragraph [0048]: the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.); paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent… instructions… to cause the agent to carry out the task) and one or more objects in the environment (paragraph [0049]: the prompt entities can identify entities in an environment)), and the at least one example task comprises at least one of the one or more actions (paragraph [0048]: the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies… to be used in the code”; Note: the “entities in an environment” map directly to the claimed “objects,” and the “algorithms, methodologies” or “tasks to be carried out” map directly to the claimed “actions”).
Regarding claim 7, Cao discloses
wherein a particular example task of the at least one example task comprises a first action set of the one or more actions (paragraph [0048]: the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.)),
the set of operations comprises a second action set of the one or more actions, and the first action set is different from the second action set (paragraph [0048]: The output sequence 150 can represent a continuation of the lines of computer code… or respective alternative examples of the lines of computer code rewritten using each prompt entity; Note: because the output sequence is a continuation or a rewritten alternative of the input code, the generated operations (i.e., second action set) are necessarily different from the input example (i.e., first action set)).
Regarding claim 8, Cao discloses
wherein the first action set is performed with respect to a first object set of the one or more objects, the second action is performed with respect to a second object set of the one or more objects, and the first object set is different from the second object set (paragraph [0048]: The output sequence 150 can represent… particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity”; paragraph [0049]: the prompt entities can identify entities in an environment; Note: Cao teaches providing an input sequence with lines of computer code (i.e., the first action set) and then generating an output sequence that is rewritten using different prompt entities (i.e., the second object set). Because the output sequence generates instructions applied to the specific target entities in the environment, which are “alternative” or “rewritten” compared to the initial example provided in the prompt, the objects acted upon in the generated plan (i.e., second object set) are necessarily different from the objects acted upon in the example task (i.e., first object set)).
Regarding claim 9, Cao discloses
wherein the at least one example task comprises at least one comment written in other than computer code (paragraph [0042]: the input sequence 102 can include an input prompt from a user, and the one or more prompt entities can include topics important to the user; paragraph [0045]: the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.)).
Regarding claim 11, Cao discloses
further comprising: an agent to perform the plan (paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions… to the agent to cause the agent to carry out the task; Note: Cao explicitly teaches that the system generates instructions for a mechanical agent/robot, which inherently requires the presence of the agent to receive and perform those instructions).
Regarding claim 12, Cao discloses
further comprising: an agent to perform the plan (paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions… to the agent to cause the agent to carry out the task; Note: A robot or mechanical agent that carries out tasks based on generated computer code or instructions is, by definition, an autonomous or semi-autonomous machine).
Regarding claim 13, Cao discloses
further comprising: an agent to perform the plan (paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions… to the agent to cause the agent to carry out the task; Note: A robot or mechanical agent that carries out tasks based on generated computer code or instructions is, by definition, an autonomous or semi-autonomous machine).
Regarding claim 22, Cao discloses
A processor comprising: one or more circuits to (paragraph [0127]):
generate a prompt to comprise operating information and an identifier of a task to be performed by an agent present in an environment (paragraph [0042]: the input sequence 102 can include an input prompt from a user; paragraph [0049]: the prompt entities can identify entities in an environment, the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent), the task to be expressed in a manner that renders the task unperformable by the agent (paragraph [0048]: the prompt entities include “semantic entities” or high-level “methodologies” which are abstract intents and not directly performable by a computer agent until rewritten), the operating information to comprise at least one task example expressed as computer code (paragraph [0048]: the input sequence 102 can include lines of computer code); and
provide the prompt to at least one machine learning process to cause the at least one machine learning process to generate a plan comprising one or more tasks to be performable by the agent (paragraph [0048]: The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity; paragraph [0049]: the output sequence can be instructions… to the agent to cause the agent to carry out the task).
Regarding claim 26, Cao discloses
wherein the one or more circuits are to perform the at least one machine learning process, and the at least one machine learning process comprises at least one Large Language Model ("LLM") (paragraph [0114]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task; paragraph [0120]: There are many different possible attention mechanisms … Exploring the limits of transfer learning with a unified text-to-text transformer … Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020; Note: the cited papers are written for modern LLMs as examples of the neural networks used in this system, specifically citing GPT-3 and T5).
Regarding claim 27, Cao discloses
wherein the LLM (paragraph [0114]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task; paragraph [0120]: There are many different possible attention mechanisms … Exploring the limits of transfer learning with a unified text-to-text transformer … Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020; Note: the cited papers are written for modern LLMs as examples of the neural networks used in this system, specifically citing GPT-3 and T5) was trained on a corpus of text comprising text written in at least one natural language (paragraph [0014]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task. For example, the different neural network can have been trained through unsupervised learning on a large corpus of unlabeled text data; paragraph [0039]: the input tokens can represent characters, word fragments, and words from human languages (e.g., English, Korean, etc.); paragraph [0049]: the prompt entities can identify entities in an environment, the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions, e.g., natural language instructions or other instructions, to the agent to cause the agent to carry out the task).
Regarding claim 28, Cao discloses
wherein both the corpus of text and the plan comprise computer code (paragraph [0042]: the input sequence 102 can include an input prompt from a user, and the one or more prompt entities can include topics important to the user; paragraph [0045]: the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.); paragraph [0114]: the different neural network can have been trained through unsupervised learning on a large corpus of unlabeled text data) and the plan comprise computer code (paragraph [0045] In another example, the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence. The output sequence 150 can include a general summary of the text sequence, and a respective sub-summary for each of the one or more prompt entities; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.). The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity).
Regarding claim 29, Cao discloses
wherein the instructions, when executed by the at least one processor, cause the at least one processor to provide data representing one or more actions (paragraph [0048]: prompt entities can include desired code segments, algorithms, methodologies (i.e., actions to be performed); paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent… instructions… to cause the agent to carry out the task), data representing one or more objects in an environment in which the plan is to be performed (paragraph [0049]: the prompt entities can identify entities in an environment), and the at least one task example comprising at least one of the one or more actions to the at least one machine learning process to cause the at least one machine learning process to produce the plan (paragraph 0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.). The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity. The system 100 can then provide the generated computer code for execution by one or more computers to carry out some computing task; Note: Providing the input sequence to the neural network to generate the output sequence maps directly to providing the data to the machine learning process to produce the plan. Furthermore, the “entities in an environment” map directly to the claimed “objects,” and the “algorithms, methodologies” or “tasks to be carried out” map directly to the claimed “actions”).
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 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 of this title, 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 2-5, 10, 14-21, and 23-25 are rejected under 35 U.S.C. 103 as being unpatentable over Cao et al. (US 2023/0108579, hereinafter Cao) in view of Tunstall-Pedoe et al. (US 2023/0274094 hereinafter Tunstall-Pedoe).
Regarding claim 2, Cao does not disclose further comprising: providing, to one or more machine learning processes, assertion related information and an assertion included in the new plan to cause the one or more machine learning processes to evaluate truthfulness of the assertion. Tunstall-Pedoe discloses further comprising: providing, to one or more machine learning processes, assertion related information and an assertion included in the new plan (paragraph [0107]: extracting the assertions in text generated by the LLM; paragraph [0858]: LLMs can also extract assertions from large blocks of text with a suitable prompt; paragraph [0875]: By translating each of these assertions into UL and turning them into Yes/No questions the system can check each of these facts for accuracy against other things it knows [i.e., the assertion related information]) to cause the one or more machine learning processes to evaluate truthfulness of the assertion (paragraph [0034]: Checking the one or more factual assertions for factual accuracy; paragraphs [0878]-[0883]: evaluating assertions to determine if they are “Yes” [true] or “No” [false]). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the generative AI system of Cao by incorporating the assertion extraction and truthfulness evaluation system of Tunstall-Pedoe. The motivation would have been to provide new context data for the LLM, in order to improve the output, such as continuation text output, generated by the LLM in response to a prompt (Tunstall-Pedoe paragraphs [0022]).
Regarding claim 3, Cao does not disclose wherein the assertion related information comprises state information related to the environment and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion. Tunstall-Pedoe discloses wherein the assertion related information comprises state information related to the environment (paragraph [0797]: The inputs to this are an initial state of the environment; paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real. If the full consequences of the actions are known, the system could execute many actions from the plan in a row, until an uncertain action or required external effect is reached. In these cases of uncertainty, the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute.) and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion (paragraphs [0877]-[0883]: evaluating specific assertions against known states and outputting a truthfulness result, e.g., evaluating the assertion “Brie is hard” against known state facts to output the result “No” [false])). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the particular use-case examples in the input sequence to condition the neural network’s output of Cao by incorporating Tunstall-Pedoe’s assertion-checking inputs to include examples of state information, example assertions, and corresponding truthfulness results (e.g., True/False or Yes/No). The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 4, Cao does not disclose further comprising: receiving, from the agent or a separate process, an identification of the assertion as the agent is performing the new plan. Tunstall-Pedoe discloses further comprising: receiving, from the agent or a separate process, an identification of the assertion as the agent is performing the new plan (paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real. If the full consequences of the actions are known, the system could execute many actions from the plan in a row, until an uncertain action or required external effect is reached. In these cases of uncertainty, the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute; paragraph [0885]: Note that this original text can come from anywhere. For example, the fact checking could be text generated by an LLM in a chat or other application where the purpose of the fact checking was to minimise hallucination or otherwise incorrect information given to the user. It could be a transcription of speech using ASR e.g. such a system could enable the real-time fact checking of a politician during an interview. The text could also come from a web page, a book, a news article or any other source of natural language). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to configure the combined system of Cao and Tunstall-Pedoe such that the identification of the assertion is received from the agent or a separate monitoring process as the agent is performing the new plan. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 5, Cao does not disclose further comprising: after the agent has performed the new plan, evaluating effectiveness of the new plan based at least in part on a comparison of a final state of the environment and a goal state of the environment. Tunstall-Pedoe discloses further comprising: after the agent has performed the new plan (paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real. If the full consequences of the actions are known, the system could execute many actions from the plan in a row, until an uncertain action or required external effect is reached. In these cases of uncertainty, the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute), evaluating effectiveness of the new plan (paragraph [0796]: the example's core thought loop can now be based around trying to find a plan, which can be an ordered series of actions, that can be executed to try and achieve the desired motives or goals found within the tenets; paragraph [0797]: The inputs to this are an initial state of the environment, encoded in UL, a set of goals, encoded in UL, as well as the core UL store including reasoning passages and knowledge of actions, along with their requirements and consequences) based at least in part on a comparison of a final state of the environment and a goal state of the environment (paragraph [0801]: Check again if the goals are met using the new state; paragraph [0803] These states can be looped over, following the process above, to calculate the environmental state after multiple actions have been executed. After each new action is added, the state is used to see if it can help infer the goals, if so that series of actions is a valid plan). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teaching of Cao by incorporating an evaluation loop that executes the generated plan in the real environment and compares the resulting final state against the initial goal state to determine if the goals were met. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 10, Cao discloses
A system comprising: at least one processor; and memory storing instructions that when executed by the at least one processor cause the at least one processor to (paragraph [0127]):
cause at least one machine learning process trained to process language to produce a plan based at least in part on input comprising at least one example task expressed as computer code (paragraph [0114]: language modeling task; paragraph [0048]: the input sequence 102 can include lines of computer code… The output sequence 150 can represent a continuation of the lines of computer code… or respective alternative examples; paragraph [0049]: the output sequence can be instructions… to the agent to cause the agent to carry out the task).
Cao does not disclose the plan to comprise an assertion, the assertion to identify at least one condition that must be satisfied for a task in the plan to be performed; and provide assertion related information to the at least one machine learning process to cause the at least one machine learning process to determine whether the at least one condition is satisfied. Tunstall-Pedoe discloses the plan to comprise an assertion, the assertion to identify at least one condition that must be satisfied for a task in the plan to be performed (paragraph [0107]: extracting the assertions in text generated by the LLM; paragraph [0804]: … execute it for real … the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute; paragraph [0858]: LLMs can also extract assertions from large blocks of text with a suitable prompt); and provide assertion related information to the at least one machine learning process to cause the at least one machine learning process to determine whether the at least one condition is satisfied (paragraph [0875]: By translating each of these assertions… and turning them into Yes/No questions the system can check each of these facts for accuracy against other things it knows ([i.e., providing assertion related information to determine if the condition is satisfied]; paragraphs [0878]-[0883]: evaluating specific assertions/conditions against known state facts to output a result of “Yes” (i.e., satisfied/true) or “No” (i.e., not satisfied/false)). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the generative AI system of Cao by incorporating the assertion extraction and condition evaluation system of Tunstall-Pedoe. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 14, Cao does not disclose wherein the instructions, when executed by the at least one processor, cause the at least one processor to: receive, from the agent or a separate process, an identification of the assertion as the agent performs the plan, and the at least one processor is to provide the assertion related information to the at least one machine learning process after receiving the identification. Tunstall-Pedoe discloses wherein the instructions, when executed by the at least one processor, cause the at least one processor to: receive, from the agent or a separate process, an identification of the assertion as the agent performs the plan (paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real … the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute; Note: It requires receiving state data/assertions from the agent or environment sensors as the plan is being executed), and the at least one processor is to provide the assertion related information to the at least one machine learning process after receiving the identification (paragraph [0875]: By translating each of these assertions into UL and turning them into Yes/No questions the system can check each of these facts for accuracy against other things it knows; paragraph [0886] As described above for fact checking of text generated by an LLM, alternative continuations can be generated so that the false facts are not shown or the correct versions of the false facts could be substituted and the LLM allowed to generate a continuation from that point. Fact checking can happen when the LLM has completed its continuation or on any partially generated continuation; Note: The evaluation necessarily occurs after the assertion is identified and received). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to configure the combined system of Cao and Tunstall-Pedoe such that the identification of the assertion is received from the agent or a separate monitoring process as the agent is performing the plan, and evaluated immediately thereafter. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 15, Cao does not disclose wherein the assertion related information comprises state information related to an environment in which the plan is to be performed and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion. Tunstall-Pedoe discloses wherein the assertion related information comprises state information related to an environment in which the plan is to be performed (paragraph [0797]: The inputs to this are an initial state of the environment; paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real. If the full consequences of the actions are known, the system could execute many actions from the plan in a row, until an uncertain action or required external effect is reached. In these cases of uncertainty, the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute) and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion (paragraphs [0877]-[0883]: evaluating specific assertions against known states and outputting a truthfulness result, e.g., evaluating the assertion “Brie is hard” against known state facts to output the result “No” [false])). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the particular use-case examples in the input sequence to condition the neural network’s output of Cao by incorporating Tunstall-Pedoe’s assertion-checking inputs to include examples of state information, example assertions, and corresponding truthfulness results (e.g., True/False or Yes/No). The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 16, Cao does not disclose wherein the instructions, when executed by the at least one processor, cause the at least one processor to evaluate effectiveness of the plan based at least in part on a comparison of a final state of an environment in which the plan was performed and a goal state of the environment. Tunstall-Pedoe discloses wherein the instructions, when executed by the at least one processor, cause the at least one processor to evaluate effectiveness of the plan (paragraph [0796]: the example's core thought loop can now be based around trying to find a plan, which can be an ordered series of actions, that can be executed to try and achieve the desired motives or goals found within the tenets; paragraph [0797]: The inputs to this are an initial state of the environment, encoded in UL, a set of goals, encoded in UL, as well as the core UL store including reasoning passages and knowledge of actions, along with their requirements and consequences) based at least in part on a comparison of a final state of an environment in which the plan was performed and a goal state of the environment (paragraph [0801]: Check again if the goals are met using the new state; paragraph [0803] These states can be looped over, following the process above, to calculate the environmental state after multiple actions have been executed. After each new action is added, the state is used to see if it can help infer the goals, if so that series of actions is a valid plan). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teaching of Cao by incorporating an evaluation loop that executes the generated plan in the real environment and compares the resulting final state against the initial goal state to determine if the goals were met. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 17, Cao discloses
the instructions, when executed by the at least one processor, cause the at least one processor to (paragraph [0127]) to perform the at least one machine learning process (paragraph [0114]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task; paragraph [0120]: There are many different possible attention mechanisms … Exploring the limits of transfer learning with a unified text-to-text transformer … Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020).
Regarding claim 18, Cao discloses
the at least one machine learning process comprises at least one Large Language Model (“LLM”) (paragraph [0114]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task; paragraph [0120]: There are many different possible attention mechanisms … Exploring the limits of transfer learning with a unified text-to-text transformer … Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020; Note: the cited papers are written for modern LLMs as examples of the neural networks used in this system, specifically citing GPT-3 and T5).
Regarding claim 19, Cao discloses
wherein the LLM (paragraph [0114]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task; paragraph [0120]: There are many different possible attention mechanisms … Exploring the limits of transfer learning with a unified text-to-text transformer … Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020; Note: the cited papers are written for modern LLMs as examples of the neural networks used in this system, specifically citing GPT-3 and T5) was trained on a corpus of text comprising text written in at least one natural language (paragraph [0014]: the first neural network blocks can have been pre-trained as part of a different neural network that performs a language modeling task. For example, the different neural network can have been trained through unsupervised learning on a large corpus of unlabeled text data; paragraph [0039]: the input tokens can represent characters, word fragments, and words from human languages (e.g., English, Korean, etc.); paragraph [0049]: the prompt entities can identify entities in an environment, the input sequence 102 can specify a task to be carried out by an agent in the environment, e.g., a robot or other mechanical agent, and the output sequence can be instructions, e.g., natural language instructions or other instructions, to the agent to cause the agent to carry out the task).
Regarding claim 20, Cao discloses
wherein both the corpus of text and the plan comprise computer code (paragraph [0042]: the input sequence 102 can include an input prompt from a user, and the one or more prompt entities can include topics important to the user; paragraph [0045]: the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.); paragraph [0114]: the different neural network can have been trained through unsupervised learning on a large corpus of unlabeled text data) and the plan comprise computer code (paragraph [0045] In another example, the input sequence 102 can include a text sequence, and the one or more prompt entities can include topics to be summarized from the text sequence. The output sequence 150 can include a general summary of the text sequence, and a respective sub-summary for each of the one or more prompt entities; paragraph [0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.). The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity).
Regarding claim 21, Cao discloses
wherein the instructions, when executed by the at least one processor, cause the at least one processor to provide data representing one or more actions (paragraph [0048]: prompt entities can include desired code segments, algorithms, methodologies (i.e., actions to be performed); paragraph [0049]: the input sequence 102 can specify a task to be carried out by an agent… instructions… to cause the agent to carry out the task), data representing one or more objects in an environment in which the plan is to be performed (paragraph [0049]: the prompt entities can identify entities in an environment), and the at least one task example comprising at least one of the one or more actions to the at least one machine learning process to cause the at least one machine learning process to produce the plan (paragraph 0048] In another example, the input sequence 102 can include lines of computer code, and the prompt entities can include desired code segments, algorithms, methodologies, or semantic entities to be used in the code (e.g., for-loops, while-loops, etc.). The output sequence 150 can represent a continuation of the lines of computer code, particular use-case examples of the prompt entities, or respective alternative examples of the lines of computer code rewritten using each prompt entity. The system 100 can then provide the generated computer code for execution by one or more computers to carry out some computing task; Note: Providing the input sequence to the neural network to generate the output sequence maps directly to providing the data to the machine learning process to produce the plan. Furthermore, the “entities in an environment” map directly to the claimed “objects,” and the “algorithms, methodologies” or “tasks to be carried out” map directly to the claimed “actions”).
Regarding claim 23, Cao does not disclose receive, from the agent or a separate process, an identification of an assertion as the agent performs the plan, the assertion to identify at least one condition that must be satisfied for at least one of the one or more tasks in the plan to be performed; and provide assertion related information to the at least one machine learning process to cause the at least one machine learning process to determine whether the at least one condition is satisfied. Tunstall-Pedoe discloses receive, from the agent or a separate process, an identification of an assertion as the agent performs the plan (paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real … the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute; Note: It requires receiving state data/assertions from the agent or environment sensors as the plan is being executed), the assertion to identify at least one condition that must be satisfied for at least one of the one or more tasks in the plan to be performed (paragraph [0107]: extracting the assertions in text generated by the LLM; paragraph [0804]: … execute it for real … the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute; paragraph [0858]: LLMs can also extract assertions from large blocks of text with a suitable prompt); and provide assertion related information to the at least one machine learning process to cause the at least one machine learning process to determine whether the at least one condition is satisfied (paragraph [0875]: By translating each of these assertions… and turning them into Yes/No questions the system can check each of these facts for accuracy against other things it knows ([i.e., providing assertion related information to determine if the condition is satisfied]; paragraphs [0878]-[0883]: evaluating specific assertions/conditions against known state facts to output a result of “Yes” (i.e., satisfied/true) or “No” (i.e., not satisfied/false)). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the generative AI system of Cao by incorporating the assertion extraction and condition evaluation system of Tunstall-Pedoe. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 24, Cao does not disclose wherein the assertion related information comprises state information related to the environment and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion. Tunstall-Pedoe discloses wherein the assertion related information comprises state information related to the environment (paragraph [0797]: The inputs to this are an initial state of the environment; paragraph [0804]: Once a valid plan has been executed, the system can select the first action from the plan and execute it for real. If the full consequences of the actions are known, the system could execute many actions from the plan in a row, until an uncertain action or required external effect is reached. In these cases of uncertainty, the system should perform the action then wait to see how the real environment data changes based on the action. From that point, the system can then re-plan to find the next action to execute) and one or more examples comprising example state information, at least one example assertion related to the example state information, and at least one corresponding result indicating truthfulness of the at least one example assertion (paragraphs [0877]-[0883]: evaluating specific assertions against known states and outputting a truthfulness result, e.g., evaluating the assertion “Brie is hard” against known state facts to output the result “No” (i.e., false))). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the particular use-case examples in the input sequence to condition the neural network’s output of Cao by incorporating Tunstall-Pedoe’s assertion-checking inputs to include examples of state information, example assertions, and corresponding truthfulness results (e.g., True/False or Yes/No). The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Regarding claim 25, Cao does not disclose wherein the one or more circuits are to evaluate effectiveness of the plan based at least in part on a comparison of a final state of the environment and a goal state of the environment. Tunstall-Pedoe discloses wherein the one or more circuits are to evaluate effectiveness of the plan (paragraph [0796]: the example's core thought loop can now be based around trying to find a plan, which can be an ordered series of actions, that can be executed to try and achieve the desired motives or goals found within the tenets; paragraph [0797]: The inputs to this are an initial state of the environment, encoded in UL, a set of goals, encoded in UL, as well as the core UL store including reasoning passages and knowledge of actions, along with their requirements and consequences.) based at least in part on a comparison of a final state of the environment and a goal state of the environment (paragraph [0801]: Check again if the goals are met using the new state; paragraph [0803] These states can be looped over, following the process above, to calculate the environmental state after multiple actions have been executed. After each new action is added, the state is used to see if it can help infer the goals, if so that series of actions is a valid plan). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teaching of Cao by incorporating an evaluation loop that executes the generated plan in the real environment and compares the resulting final state against the initial goal state to determine if the goals were met. The motivation would have been to check LLM output for accuracy (Tunstall-Pedoe paragraphs [0103]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Koneru et al. (US 11,461,417) discloses “The predictive model is trained by providing the predictive model with a training set of user inputs, parameters to be identified and weights of the parameters” (col. 21, lines 11-15) and “The user may provide the user input to the VA server 100(1) in natural language, for the VA server 100(1) to execute a task. In some examples, the user input may be an image, a document, a software code and so on” (col. 26, lines 38-41).
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 [0037] 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 extension fee pursuant to [0037] 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 date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SISLEY N. KIM whose telephone number is (571)270-7832. The examiner can normally be reached M-F 11:30AM -7:30PM.
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, April Y. Blair can be reached on (571)270-1014. 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.
/SISLEY N KIM/Primary Examiner, Art Unit 2196 7/25/2026
Prosecution Insights