DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
All objections/rejections not mentioned in this Office Action have been withdrawn by the Examiner.
Status of the Claims
Prior to entry of the amendment(s) and/or consideration of the argument(s), the status of the claims is as follows.
Claim(s) 1-20 is/are pending.
Claim(s) 1-2, 11-12, and 20 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Gelfenbeyn (U.S. Pat. App. Pub. No. 2024/0221264, hereinafter Gelfenbeyn).
Claims 3-5 and 13-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn as applied to claim 2 and 12 above, and further in view of Khyatti (U.S. Pat. App. Pub. No. 2025/0356218, hereinafter Khyatti).
Claims 6-9 and 16-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn as applied to claim 1 and 11 above, and further in view of Li (CN117556802A, hereinafter Li).
Claims 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn as applied to claim 1 above, and further in view of Poulis (U.S. Pat. No. 12,182,678, hereinafter Poulis).
Response to Amendments
Applicant’s amendment filed on 02 April 2026 has been entered.
In view of the amendment to the claim(s), the amendment of claim(s) 1, 11, and 20 have been acknowledged and entered.
In view of the amendment to claim(s) 1, 11, and 20, the rejection of claims 1-20 under 35 U.S.C. §102 and 103 is withdrawn.
In light of the amended/newly added claims, new grounds for rejection under 35 U.S.C. §103 and 35 U.S.C. §112 are provided in the action below.
Response to Arguments
Applicant’s arguments regarding the prior art rejections under 35 U.S.C. §102/103, see pages 8-9 of the Response to Non-Final Office Action dated 02 January 2026, which was received on 02 April 2026 (hereinafter Response and Office Action, respectively), have been fully considered.
With respect to the rejection(s) of claim(s) 1, 11, and 20 under 35 U.S.C. §102(a)(2) as being anticipated by Gelfenbeyn, applicant asserts that Gelfenbeyn (1) fails to teach or suggest all limitations of claims 1, 11, and 20 as originally presented; and (2) fails to teach or suggest all limitations of claims 1, 11, and 20 as currently amended. Applicant’s arguments are addressed below.
Regarding the first argument, applicant's arguments fail to comply with 37 CFR 1.111(b) because they amount to a general allegation that the claims define a patentable invention without specifically pointing out how the language of the claims patentably distinguishes them from the references. Applicant “traverses the §102 rejection on the ground that Gelfenbeyn fails to expressly or inherently teach each and every limitation of claims 1, 2, 11, 12 and 20 as originally presented, arranged as recited in those claims.” (Response, pg. 8). It is noted that specific portions of the Gelfenbeyn were indicated and mapped to each and every limitation of claims 1-2, 11-12, and 20. (Office Action, pgs. 2-6). However, applicant provides no further explanation as to the evidentiary basis for this traversal. Specifically, applicant has failed to indicate which mapped components of Gelfenbeyn are believed to be improperly mapped to a specific limitation of the claim and the logical basis for said argument, such that the applicant’s argument can be properly considered. Therefore, the traversal is without substance and is improper under 37 CFR 1.111(b).
Regarding the second argument, applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 1, 11, and 20 under 35 U.S.C. §102 are withdrawn.
Applicant further argues that the rejection(s) of dependent claims 2-10 and 12-19 should be withdrawn for at least the same reasons as independent claims 1, 11, and 20. Applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 2-10 and 12-19 under 35 U.S.C. §102 and 35 U.S.C. §103 are withdrawn.
However, upon further consideration, new ground(s) of rejection under 35 U.S.C. §103 are made in light of combinations of Gelfenbeyn, Khyatti, Li, Poulis, and newly cited reference Non-patent Literature to Hazra (Hazra, R., Dos Martires, P.Z. and De Raedt, L., 2024, March. Saycanpay: Heuristic planning with large language models using learnable domain knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 18, pp. 20123-20133), hereinafter Hazra).
The Applicant has not provided any further statement and therefore, the Examiner directs the Applicant to the below rationale.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 3-7, 9, 13-17, and 19 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Regarding claim 3, and mutatis mutandis claim 13, the combination of a role query and a role prompt together constitute new matter. As amended, claim 1, from which claim 3 depends, recites “generat[ing] a role query…, applying the role query… to the LLM, and obtaining at least portions of the respective role information… from the LLM responsive to the respective role query” at lines 9-11. Applicant asserts that “Support for these amendments can be found in the specification at, for example, page 6, line 3, to page 7, line 22; page 8, lines 7-22; page 10, lines 3-8; and page 12, lines 14-20; and in the non-limiting illustrative embodiment of FIG. 2” (Response, pg. 9). Upon review of the cited sections, the specification does describe a role query and a role prompt, as different terms for the same component, as indicated by the alternating use within the same paragraph. The specification describes “the automatic role configuration module 210 can request the LLM 230 to generate the role information by using the user input as part of a query (for example, using a structured template)” which is understood as a restatement of a more specific embodiment “the automatic role configuration module 210 can extract semantic information, such as an entity, from the user input, and input the extracted language information into the LLM 230 as a role prompt and determine a response generated by the LLM 230 as the role information.” (Instant Application, pg. 6, line 22 – pg. 7, line 6). Claim 3, however, describes “generating a role prompt… and generating the role information by providing the role prompt to the LLM” at lines 2-3. As presented in the claims, the role query and the role prompt are not related components, and each, independently, results in the role information.
The specification, either at the indicated location or more generally, does not provide support for both a “role query” and a “role prompt” as separate entities, each involved in the generation of the role information. Therefore, claims 3 and 13 in light of amended claims 1 and 11, respectively, contains limitations which are not supported by the specification as filed and the claims are rejected as containing new matter.
Regarding claims 4-5 and 14-15, claims 4-5 and 14-15 depend from claims 3 and 13 and incorporate all limitations therefrom. Therefore, claims 4-5 and 14-15 are rejected for at least the same reasons as claims 3 and 13.
Regarding claim 6, and mutatis mutandis claim 16, the combination of an alignment query and an alignment prompt together constitute new matter. As amended, claim 1, from which claim 6 depends, recites “generat[ing]… an alignment query…, applying… the alignment query… to the LLM, and obtaining at least portions of the respective… alignment information… from the LLM responsive to the respective… alignment query” at lines 9-11. Applicant asserts that “Support for these amendments can be found in the specification at, for example, page 6, line 3, to page 7, line 22; page 8, lines 7-22; page 10, lines 3-8; and page 12, lines 14-20; and in the non-limiting illustrative embodiment of FIG. 2” (Response, pg. 9). Upon review of the cited sections, the specification, though never reciting either an alignment query, an alignment prompt, it does describe querying for an alignment using a prompt. The specification describes “At block 730, an alignment strategy is obtained…by … querying the LLM” where “the alignment target can be input into the LLM as a prompt, and then the LLM can generate the alignment strategy.” (Instant Application, pg. 13, line 12-16). Claim 6, however, describes “generating an alignment prompt for aligning with the user based on the user input; and generating the alignment information by providing the alignment prompt to the LLM” at lines 2-3. As presented in the claims, the alignment query and the alignment prompt are not related components, and each, independently, results in the role information. As such, for such an embodiment to be properly claimed, the specification as filed must support the co-existence of both an alignment query and an alignment prompt. However, no clear support such an embodiment is known to exist in the specification as filed.
The specification, either at the indicated location or more generally, does not provide support for both an “alignment query” and an “alignment prompt” as separate entities, each involved in the generation of the alignment information. Therefore, claims 6 and 16 in light of amended claims 1 and 11, respectively, contains limitations which are not supported by the specification as filed and the claims are rejected as containing new matter.
Regarding claims 6-7, 9, 16-17 and 19, claims 6-7, 9, 16-17 and 19 depend from claims 6 and 16 and incorporate all limitations therefrom. Therefore, claims 6-7, 9, 16-17 and 19 are rejected for at least the same reasons as claims 6 and 16.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-2, 11-12, and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn in view of Hazra.
Regarding claim 1, Gelfenbeyn discloses A method for a large language model (LLM) (Systems and methods for “controlling safety settings for behavior characteristics of an AI character model,” described with reference to “a platform 200 for generating AI character models” including “a studio 202, an integration interface 204, and an AI character model 300,”; Gelfenbeyn, ¶ [0030], [0040]), comprising: receiving, in a processor-based machine learning system, a user input for an LLM agent of the processor-based machine learning system (Teaches the “platform 200 {the LLM agent}” as part of the overall AI character generation system {machine learning system}, which is a “computing platform {processor based}... for providing an AI character model” and “receiving a user input 208… [which] may include voice messages of a user” for an “AI character model” such as AI character model 300, which “may utilize LLMs in conversations,” and, further, the “AI character models may include generative models”; Gelfenbeyn, ¶ [0028], [0035]-[0036], [0040], [0046], [0076]), the LLM agent comprising an automatic role configuration component and an automatic alignment configuration component, each configured to interact with the LLM (The system comprising a “heuristics models 620” including at least “the goals model 622” {automatic role configuration component}, and “the safety model 624 {automatic alignment configuration component}.”; Gelfenbeyn, ¶ [0058], [0063], [0068]-[0069]); determining role information and user-related alignment information for the LLM agent based on the user input, in the respective automatic role configuration component and automatic alignment configuration component of the LLM agent (“The goals model 622 {...in the respective automatic role configuration component} may be configured to process the text and/or embeddings 618 and recognize, based on what was said by the user or the AI character, what goals need to be activated {determining role information... for the LLM agent}” and “The safety model 624 {in the...automatic alignment configuration component} may be configured to process the text and/or embeddings 618 and filter out unsafe responses {user related alignment information},” where, as described with reference to FIG. 7, the input is processed to determine at least “input impact for goals model 704” which structures “top-level motivations” and the “identity profile 718” {role information} and determines a “level of safety” and the “safety model” to adjust “safety settings” {...alignment information for the LLM agent} based on context such as the user’s age or social group, as well as, “user behavior,” a determined intelligence level, and historical data {user-related...}, as derived from user input.; Gelfenbeyn, ¶ [0031], [0033], [0066], [0076], [0078], [0080]), at least in part by processing the user input to generate a role query and an alignment query (Discloses generating pre-processed text and/or embeddings where the “pre-processed data stream in the form of text and/or embeddings 618 may be obtained upon pre-processing of the received inputs” and “pre-processing may include converting the received inputs into a singular format,” {a role query and an alignment query}; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), respectively, applying the role query and the alignment query to… [a heuristic model] (The goals model 622 and the safety model 624 are members of the heuristic models 620, where the heuristic models 620 “may include machine learning models... [and] can be implemented as artificial neural networks,” and each receives the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), and obtaining at least portions of the respective role information and alignment information from the…[heuristic model] responsive to the respective role query and alignment query (“Upon the processing of the data, the heuristics models 620 may provide intermediate outputs,” where the goals model produces an intermediate output, which, at least, “recognizes, based on what was said by the user or the AI character, what goals need to be activated” and the safety model produces an intermediate output, which, at least, adjusts “safety settings” and “filters out unsafe responses”; Gelfenbeyn, ¶ [0066], [0068]); generating, in the processor-based machine learning system, a prompt including the role information and the alignment information (The system further includes an “orchestration step” that involves “composing the intermediate outputs received from the heuristics models into templated formats for ingestion” and “upon composing the intermediate outputs into a template, the composed outputs may be fed into primary models representing elements of multimodal expression”; Gelfenbeyn, ¶ [0066], [0068]-[0070]) at least portions of which were previously obtained from the…[heuristic model] utilizing the respective role query and alignment query (The intermediate outputs were obtained from the goals model and the safety model based on the processing of the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0063]-[0066], [0068]); and generating, in the processor-based machine learning system, an answer to the user input by providing the prompt to the LLM (The composed outputs generated based on the generated intermediate output template, are “sent to another series of AI models” including the LLM, which, in response, produce the “dialogue output 654, the client-side narrative triggers 656, the animation controls 658, and the voice parameters 644,” collectively referred to as “output data”, where the “output data obtained upon applying the text to speech conversion 660 are sent as a stream to the client 662”; Gelfenbeyn, ¶ [0068]-[0070], [0074]-[0075]). However, Gelfenbeyn fails to expressly recite wherein the heuristic model includes a large language model.
Hazra teaches systems and methods for heuristic planning using LLMs. (Hazra, Abstract). Regarding claim 1, Hazra teaches wherein the heuristic model includes a large language model (Discloses “SayCanPay,” which “employs LLMs to generate actions (Say) guided by learnable domain knowledge, that evaluates actions’ feasibility (Can) and long-term reward/payoff (Pay), and heuristic search to select the best sequence of actions.” The use of LLMs “in the context of heuristic planning,” as read in the context of Gelfenbeyn, discloses the use of LLMs as part of the machine learning “heuristic models 620”, including the goals model 622 and the safety model 624.; Hazra, ¶ Abstract).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn to incorporate the teachings of Hazra to include wherein the heuristic model includes a large language model. The heuristic modeling via LLMs described in Hazra provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which combines “the world knowledge and generative capabilities of LLMs with the systematicity of classical planning by formulating a heuristic search-based planning framework for LLMs,” which provides the known benefit of broader knowledge input and better response to novel requests, than traditional heuristic models, as recognized in light of Hazra. (Hazra, Abstract; pg. 20131, para. 2).
Regarding claim 2, the rejection of claim 1 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. Gelfenbeyn further discloses wherein determining role information for the LLM agent comprises: extracting semantic information from the user input (Discloses the “streams of inputs” which includes the user input 208, “may be pre-processed” and “based on the pre-processing, all inputs may be transformed into text and/or embeddings 618”; Gelfenbeyn, ¶ [0060]-[0061], [0063]); and determining the role information for the LLM agent based on the extracted semantic information (“the text and/or embeddings 618 may be passed...through a goals model 622” which uses “what was said by the user” to determine “what goals need to be activated.”; Gelfenbeyn, ¶ [0065]-[0066]).
Regarding claim 11, Gelfenbeyn discloses An electronic device, comprising (Systems and methods for “controlling safety settings for behavior characteristics of an AI character model,” described with reference to “a platform 200 for generating AI character models” including “a studio 202, an integration interface 204, and an AI character model 300,” as incorporated into computer system 1100; Gelfenbeyn, ¶ [0030], [0040], [0107]-[0108]): at least one processor; and a memory coupled to the at least one processor and storing instructions, wherein the instructions, when executed by the at least one processor, cause the electronic device to perform actions (Computer system 1100 can include “memory 1104... used to store program instructions for execution by the processor(s) 1102” where said “memory 1104, in one example, is used by software (e.g., the operating system 1114 or the software applications 1116)” including “software applications suitable for implementing at least some operations of the methods for controlling safety settings for behavior characteristics of an AI character model as described herein”; Gelfenbeyn, ¶ [0107]-[0108]), comprising: receiving, in a processor-based machine learning system, a user input for an LLM agent of the processor-based machine learning system (Teaches the “platform 200 {the LLM agent}” as part of the overall AI character generation system {machine learning system}, which is a “computing platform {processor based}... for providing an AI character model” and “receiving a user input 208… [which] may include voice messages of a user” for an “AI character model” such as AI character model 300, which “may utilize LLMs in conversations,” and, further, the “AI character models may include generative models”; Gelfenbeyn, ¶ [0028], [0035]-[0036], [0040], [0046], [0076]), the LLM agent comprising an automatic role configuration component and an automatic alignment configuration component, each configured to interact with the LLM (The system comprising a “heuristics models 620” including at least “the goals model 622” {automatic role configuration component}, and “the safety model 624 {automatic alignment configuration component}.”; Gelfenbeyn, ¶ [0058], [0063], [0068]-[0069]); determining role information and user-related alignment information for the LLM agent based on the user input, in the respective automatic role configuration component and automatic alignment configuration component of the LLM agent (“The goals model 622 {...in the respective automatic role configuration component} may be configured to process the text and/or embeddings 618 and recognize, based on what was said by the user or the AI character, what goals need to be activated {determining role information... for the LLM agent}” and “The safety model 624 {in the...automatic alignment configuration component} may be configured to process the text and/or embeddings 618 and filter out unsafe responses {user related alignment information},” where, as described with reference to FIG. 7, the input is processed to determine at least “input impact for goals model 704” which structures “top-level motivations” and the “identity profile 718” {role information} and determines a “level of safety” and the “safety model” to adjust “safety settings” {...alignment information for the LLM agent} based on context such as the user’s age or social group, as well as, “user behavior,” a determined intelligence level, and historical data {user-related...}, as derived from user input.; Gelfenbeyn, ¶ [0031], [0033], [0066], [0076], [0078], [0080]), at least in part by processing the user input to generate a role query and an alignment query (Discloses generating pre-processed text and/or embeddings where the “pre-processed data stream in the form of text and/or embeddings 618 may be obtained upon pre-processing of the received inputs” and “pre-processing may include converting the received inputs into a singular format,” {a role query and an alignment query}; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), respectively, applying the role query and the alignment query to… [a heuristic model] (The goals model 622 and the safety model 624 are members of the heuristic models 620, where the heuristic models 620 “may include machine learning models... [and] can be implemented as artificial neural networks,” and each receives the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), and obtaining at least portions of the respective role information and alignment information from the…[heuristic model] responsive to the respective role query and alignment query (“Upon the processing of the data, the heuristics models 620 may provide intermediate outputs,” where the goals model produces an intermediate output, which, at least, “recognizes, based on what was said by the user or the AI character, what goals need to be activated” and the safety model produces an intermediate output, which, at least, adjusts “safety settings” and “filters out unsafe responses”; Gelfenbeyn, ¶ [0066], [0068]); generating, in the processor-based machine learning system, a prompt including the role information and the alignment information (The system further includes an “orchestration step” that involves “composing the intermediate outputs received from the heuristics models into templated formats for ingestion” and “upon composing the intermediate outputs into a template, the composed outputs may be fed into primary models representing elements of multimodal expression”; Gelfenbeyn, ¶ [0066], [0068]-[0070]) at least portions of which were previously obtained from the…[heuristic model] utilizing the respective role query and alignment query (The intermediate outputs were obtained from the goals model and the safety model based on the processing of the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0063]-[0066], [0068]); and generating, in the processor-based machine learning system, an answer to the user input by providing the prompt to the LLM (The composed outputs generated based on the generated intermediate output template, are “sent to another series of AI models” including the LLM, which, in response, produce the “dialogue output 654, the client-side narrative triggers 656, the animation controls 658, and the voice parameters 644,” collectively referred to as “output data”, where the “output data obtained upon applying the text to speech conversion 660 are sent as a stream to the client 662”; Gelfenbeyn, ¶ [0068]-[0070], [0074]-[0075]). However, Gelfenbeyn fails to expressly recite wherein the heuristic model includes a large language model.
The relevance of Hazra is described above with relation to claim 1. Regarding claim 11, Hazra teaches wherein the heuristic model includes a large language model (Discloses “SayCanPay,” which “employs LLMs to generate actions (Say) guided by learnable domain knowledge, that evaluates actions’ feasibility (Can) and long-term reward/payoff (Pay), and heuristic search to select the best sequence of actions.” The use of LLMs “in the context of heuristic planning,” as read in the context of Gelfenbeyn, discloses the use of LLMs as part of the machine learning “heuristic models 620”, including the goals model 622 and the safety model 624.; Hazra, ¶ Abstract).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn to incorporate the teachings of Hazra to include wherein the heuristic model includes a large language model. The heuristic modeling via LLMs described in Hazra provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which combines “the world knowledge and generative capabilities of LLMs with the systematicity of classical planning by formulating a heuristic search-based planning framework for LLMs,” which provides the known benefit of broader knowledge input and better response to novel requests, than traditional heuristic models, as recognized in light of Hazra. (Hazra, Abstract; pg. 20131, para. 2).
Regarding claim 12, the rejection of claim 11 is incorporated. Claim 12 is substantially the same as claim 2 and is therefore rejected under the same rationale as above.
Regarding claim 20, Gelfenbeyn discloses A computer program product (Systems and methods for “controlling safety settings for behavior characteristics of an AI character model,” described with reference to “a platform 200 for generating AI character models” including “a studio 202, an integration interface 204, and an AI character model 300,” as included in a computer program product; Gelfenbeyn, ¶ [0030], [0040], [0108]), the computer program product being tangibly stored on a non-transitory computer-readable medium and comprising machine-executable instructions, wherein the machine-executable instructions, when executed by a machine, cause the machine to perform actions (discloses “memory 1104... used to store program instructions for execution by the processor(s) 1102” where said “memory 1104, in one example, is used by software (e.g., the operating system 1114 or the software applications 1116)” including “software applications suitable for implementing at least some operations of the methods for controlling safety settings for behavior characteristics of an AI character model as described herein”; Gelfenbeyn, ¶ [0108]), comprising: receiving, in a processor-based machine learning system, a user input for an LLM agent of the processor-based machine learning system (Teaches the “platform 200 {the LLM agent}” as part of the overall AI character generation system {machine learning system}, which is a “computing platform {processor based}... for providing an AI character model” and “receiving a user input 208… [which] may include voice messages of a user” for an “AI character model” such as AI character model 300, which “may utilize LLMs in conversations,” and, further, the “AI character models may include generative models”; Gelfenbeyn, ¶ [0028], [0035]-[0036], [0040], [0046], [0076]), the LLM agent comprising an automatic role configuration component and an automatic alignment configuration component, each configured to interact with the LLM (The system comprising a “heuristics models 620” including at least “the goals model 622” {automatic role configuration component}, and “the safety model 624 {automatic alignment configuration component}.”; Gelfenbeyn, ¶ [0058], [0063], [0068]-[0069]); determining role information and user-related alignment information for the LLM agent based on the user input, in the respective automatic role configuration component and automatic alignment configuration component of the LLM agent (“The goals model 622 {...in the respective automatic role configuration component} may be configured to process the text and/or embeddings 618 and recognize, based on what was said by the user or the AI character, what goals need to be activated {determining role information... for the LLM agent}” and “The safety model 624 {in the...automatic alignment configuration component} may be configured to process the text and/or embeddings 618 and filter out unsafe responses {user related alignment information},” where, as described with reference to FIG. 7, the input is processed to determine at least “input impact for goals model 704” which structures “top-level motivations” and the “identity profile 718” {role information} and determines a “level of safety” and the “safety model” to adjust “safety settings” {...alignment information for the LLM agent} based on context such as the user’s age or social group, as well as, “user behavior,” a determined intelligence level, and historical data {user-related...}, as derived from user input.; Gelfenbeyn, ¶ [0031], [0033], [0066], [0076], [0078], [0080]), at least in part by processing the user input to generate a role query and an alignment query (Discloses generating pre-processed text and/or embeddings where the “pre-processed data stream in the form of text and/or embeddings 618 may be obtained upon pre-processing of the received inputs” and “pre-processing may include converting the received inputs into a singular format,” {a role query and an alignment query}; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), respectively, applying the role query and the alignment query to… [a heuristic model] (The goals model 622 and the safety model 624 are members of the heuristic models 620, where the heuristic models 620 “may include machine learning models... [and] can be implemented as artificial neural networks,” and each receives the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0048], [0061], [0063]-[0065]), and obtaining at least portions of the respective role information and alignment information from the…[heuristic model] responsive to the respective role query and alignment query (“Upon the processing of the data, the heuristics models 620 may provide intermediate outputs,” where the goals model produces an intermediate output, which, at least, “recognizes, based on what was said by the user or the AI character, what goals need to be activated” and the safety model produces an intermediate output, which, at least, adjusts “safety settings” and “filters out unsafe responses”; Gelfenbeyn, ¶ [0066], [0068]); generating, in the processor-based machine learning system, a prompt including the role information and the alignment information (The system further includes an “orchestration step” that involves “composing the intermediate outputs received from the heuristics models into templated formats for ingestion” and “upon composing the intermediate outputs into a template, the composed outputs may be fed into primary models representing elements of multimodal expression”; Gelfenbeyn, ¶ [0066], [0068]-[0070]) at least portions of which were previously obtained from the…[heuristic model] utilizing the respective role query and alignment query (The intermediate outputs were obtained from the goals model and the safety model based on the processing of the pre-processed “text and/or embeddings”.; Gelfenbeyn, ¶ [0063]-[0066], [0068]); and generating, in the processor-based machine learning system, an answer to the user input by providing the prompt to the LLM (The composed outputs generated based on the generated intermediate output template, are “sent to another series of AI models” including the LLM, which, in response, produce the “dialogue output 654, the client-side narrative triggers 656, the animation controls 658, and the voice parameters 644,” collectively referred to as “output data”, where the “output data obtained upon applying the text to speech conversion 660 are sent as a stream to the client 662”; Gelfenbeyn, ¶ [0068]-[0070], [0074]-[0075]). However, Gelfenbeyn fails to expressly recite wherein the heuristic model includes a large language model.
The relevance of Hazra is described above with relation to claim 1. Regarding claim 20, Hazra teaches wherein the heuristic model includes a large language model (Discloses “SayCanPay,” which “employs LLMs to generate actions (Say) guided by learnable domain knowledge, that evaluates actions’ feasibility (Can) and long-term reward/payoff (Pay), and heuristic search to select the best sequence of actions.” The use of LLMs “in the context of heuristic planning,” as read in the context of Gelfenbeyn, discloses the use of LLMs as part of the machine learning “heuristic models 620”, including the goals model 622 and the safety model 624.; Hazra, ¶ Abstract).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn to incorporate the teachings of Hazra to include wherein the heuristic model includes a large language model. The heuristic modeling via LLMs described in Hazra provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which combines “the world knowledge and generative capabilities of LLMs with the systematicity of classical planning by formulating a heuristic search-based planning framework for LLMs,” which provides the known benefit of broader knowledge input and better response to novel requests, than traditional heuristic models, as recognized in light of Hazra. (Hazra, Abstract; pg. 20131, para. 2).
Claims 3-5 and 13-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn and Hazra as applied to claim 2 and 12 above, and further in view of Khyatti (U.S. Pat. App. Pub. No. 2025/0356218, hereinafter Khyatti).
Regarding claim 3, the rejection of claim 2 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. Gelfenbeyn further discloses wherein determining the role information for the LLM agent comprises: generating… [tasks] based on the extracted semantic information (“the text and/or embeddings 618 may be passed...through a goals model 622” which uses “what was said by the user” to determine “what goals {tasks} need to be activated,” in which the system generates a prompt that incorporates the personality and role data extracted from the initial semantic analysis; Gelfenbeyn, ¶ [0065]-[0066]). However, Gelfenbeyn fails to expressly recite generating a role prompt based on generated tasks; and generating the role information by providing the role prompt to the LLM.
Khyatti teaches systems and methods for “generating a chain of thought content using an artificial intelligence (AI) system.” (Khyatti, ¶ [0041]). Regarding claim 3, Khyatti teaches wherein determining the role information for the LLM agent comprises: generating a role prompt based on the extracted semantic information (Discloses the backend generating the initiation prompt {role prompt} which “includes a set of rules, guidelines, desired output, core instruction description, and a task,” where the task is based on user input of a “natural language description of a task” where “to build an initiation prompt, the backend receives text inputted from the user to the frontend, extracts important information from the text to include in the initiation prompt, and builds the prompt into a data structure and format readable by the machine learning model.”; Khyatti, ¶ [0167], [0197], [0203]); and generating the role information by providing the role prompt to the LLM (“When the user 302 performs ‘generate a prompt’, the LLM decides what ‘job’ is appropriate to solve the task,” where the task is derived from the natural language description and semantic information extracted therefrom.; Khyatti, ¶ [0295]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Khyatti to include generating a role prompt based on generated tasks; and generating the role information by providing the role prompt to the LLM. The disclosed “CoT meta-prompting system” described in Khyatti provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which provides for the known benefit of increasing the overall likelihood of success in performing complex tasks and improving the quality of output to the end user, as recognized by Khyatti. (Khyatti, ¶ [0054] ). Specifically, with regards to selecting a job or role, the relatively static selection mechanisms of Gelfenbeyn would be understandably improved by the LLM job selection of Khyatti, as the use of an LLM for job selection based on the task offers well known and real-time flexibility in selecting the job/role when dealing with new parameters and/or situations, over prior art systems, such as rules based systems described in Gelfenbeyn, while further avoiding the need for specialized training. (Khyatti, ¶ [0295]; Gelfenbeyn, ¶ [0059]).
Regarding claim 4, the rejection of claim 3 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. However, Gelfenbeyn fails to expressly recite further comprising: generating the role prompt based on meta-learning and priming, wherein the meta-learning provides an initialization parameter for the role prompt and the priming provides one or more role examples.
The relevance of Khyatti is described above with relation to claim 3. Regarding claim 4, Khyatti teaches further comprising: generating the role prompt based on meta-learning and priming, (Discloses the backend generating the initiation prompt {role prompt} based on extracted “important information from the text” of the user input {meta-learning} and “generic prompts” {priming}; Khyatti, ¶ [0067], [0197]) wherein the meta-learning provides an initialization parameter for the role prompt (Discloses the backend generating the initiation prompt {role prompt} which “includes a set of rules, guidelines, desired output, core instruction description, and a task,” where the task is based on user input of a “natural language description of a task” where “to build an initiation prompt, the backend receives text inputted from the user to the frontend, extracts important information from the text to include in the initiation prompt, and builds the prompt into a data structure and format readable by the machine learning model” which defines the “objective” and “instructions to define a plurality of agents” {initialization parameter}; Khyatti, ¶ [0167], [0197], [0203]) and the priming provides one or more role examples (Discloses providing priming via “generic prompts” retrieved from the previously generated “prompt templates in a storage”, which provide the generic concept for guiding the development of the prompt; Khyatti, ¶ [0055], [0067], [0296]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Khyatti to include further comprising: generating the role prompt based on meta-learning and priming, wherein the meta-learning provides an initialization parameter for the role prompt and the priming provides one or more role examples. The disclosed “CoT meta-prompting system” described in Khyatti provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which provides for the known benefit of increasing the overall likelihood of success in performing complex tasks and improving the quality of output to the end user, as recognized by Khyatti. (Khyatti, ¶ [0054] ). Specifically, with regards to selecting a job or role, the relatively static selection mechanisms of Gelfenbeyn would be understandably improved by the LLM job selection of Khyatti, as the use of an LLM for job selection based on the task offers well known and real-time flexibility in selecting the job/role when dealing with new parameters and/or situations, over prior art systems, such as rules based systems described in Gelfenbeyn, while further avoiding the need for specialized training. (Khyatti, ¶ [0295]; Gelfenbeyn, ¶ [0059]).
Regarding claim 5, the rejection of claim 4 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. However, Gelfenbeyn fails to expressly recite wherein the one or more role examples comprise at least one of: role name, expertise, language style, and emotional expression.
The relevance of Khyatti is described above with relation to claim 3. Regarding claim 5, Khyatti teaches wherein the one or more role examples comprise at least one of: role name, expertise, language style, and emotional expression (“generic prompts” retrieved from the previously generated “prompt templates in a storage” include properties such as “name” {role name}, “title” and “description” {expertise}; Khyatti, ¶ [0055], [0067], [0291]-[0294]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Khyatti to include wherein the one or more role examples comprise at least one of: role name, expertise, language style, and emotional expression. The disclosed “CoT meta-prompting system” described in Khyatti provides for the handling of “complex tasks by breaking the task into manageable and digestible steps, each of which can be individually solved,” which provides for the known benefit of increasing the overall likelihood of success in performing complex tasks and improving the quality of output to the end user, as recognized by Khyatti. (Khyatti, ¶ [0054] ). Specifically, with regards to selecting a job or role, the relatively static selection mechanisms of Gelfenbeyn would be understandably improved by the LLM job selection of Khyatti, as the use of an LLM for job selection based on the task offers well known and real-time flexibility in selecting the job/role when dealing with new parameters and/or situations, over prior art systems, such as rules based systems described in Gelfenbeyn, while further avoiding the need for specialized training. (Khyatti, ¶ [0295]; Gelfenbeyn, ¶ [0059]).
Regarding claim 13, the rejection of claim 12 is incorporated. Claim 13 is substantially the same as claim 3 and is therefore rejected under the same rationale as above.
Regarding claim 14, the rejection of claim 13 is incorporated. Claim 14 is substantially the same as claim 4 and is therefore rejected under the same rationale as above.
Regarding claim 15, the rejection of claim 14 is incorporated. Claim 15 is substantially the same as claim 5 and is therefore rejected under the same rationale as above.
Claims 6-9 and 16-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn and Hazra as applied to claims 1 and 11 above, and further in view of Li (CN117556802A, hereinafter Li ).
Regarding claim 6, the rejection of claim 1 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. Gelfenbeyn further discloses wherein determining user-related alignment information comprises: generating... [alignment information] for aligning with the user based on the user input (Discloses a “safety model 312” used for “determining a level of safety” and “generating the safety settings 316” {alignment information} based on “the desired safety settings… set by the system in the course of the interaction with the user”; Gelfenbeyn, ¶ [0031], [0034]-[0035]). However, Gelfenbeyn fails to expressly recite generating an alignment prompt for aligning with the user based on the user input; and generating the alignment information by providing the alignment prompt to the LLM.
Li teaches “a user portrait method, device, equipment and medium based on a large language model.” (Li, ¶ [0002]). Regarding claim 6, Li teaches generating an alignment prompt for aligning with the user based on the user input (“automatically matching and generating an implication prompt word template according to the personality characteristic label Cj of the user related to the j-th dialogue”; Li, ¶ [0050]); and generating the alignment information by providing the alignment prompt to the LLM (Discloses “importing the implication prompt word template into the large language model to obtain the following implication relation between the characteristic values Vj and Vnew” and “updating the dialogue generation strategy” in light of alignment between Vj and Vnew; Li, ¶ [0050]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Li to include generating an alignment prompt for aligning with the user based on the user input; and generating the alignment information by providing the alignment prompt to the LLM. Gelfenbeyn discloses that the alignment is to be “adapted dynamically based on real-time context” (Gelfenbeyn, ¶ [0033]). However, Gelfenbeyn fails to provide a clear mechanism for achieving this functional goal. Li describes the use of a large language model to iteratively generate the indicated alignment information for a user, as part of a user portrait generation method which captures the user’s evolving inner state, thus providing a clear mechanism to achieve the dynamic adaptation referenced in Gelfenbeyn, and providing the known benefit of “greatly improving the accuracy and instantaneity of user portraits,” while maintaining convenience “for practical application and popularization”, as recognized by Li. (Li, ¶ [0008]).
Regarding claim 7, the rejection of claim 6 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. Gelfenbeyn further discloses wherein the alignment information comprises: at least one of a user portrait of the user, an alignment target for interaction with the user, and an alignment strategy for interaction with the user (Discloses identifying contextual behavior including “user behavior during the interaction with the AI character, a maturity level of the user determined based on the user behavior, an intelligence level of the user determined based on the user behavior, user data associated with a user profile registered in the system (e.g., a user profile in Roblox® online game platform), historical data associated with the user”; Gelfenbeyn, ¶ [0033]), an alignment target for interaction with the user (Discloses levels of safety which correspond to content ratings as used in media entertainment “such as “G” for content intended for general audiences, “PG” for content for which the parental guidance is suggested, “PG-13” for the content for which parents are strongly cautioned, “R” for content that requires users of under 17 years old to be accompanied by parent or adult guardian, and “NC-17” for content restricted for users of under 17” where each of which is an alignment target.; Gelfenbeyn, ¶ [0101]), and an alignment strategy for interaction with the user (Discloses using a “hierarchy of safety settings” including “enabling different settings for appropriate audiences, including allowed topics, levels of profanity, and so forth”; Gelfenbeyn, ¶ [0035]).
Regarding claim 8, the rejection of claim 1 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. Gelfenbeyn further discloses wherein determining user-related alignment information comprises: generating... a user portrait based on the user input (“receiving... context associated with the AI character model” which “may be associated with one or more of the following: an age of the user, a gender of the user, a geographical location associated with the user, a country of residence of the user, an ethnical group of the user, a social group of the user, other parameters associated with the user, and so forth” and where the context may be determined based on user input.; Gelfenbeyn, ¶ [0099]); based on the user portrait... generating the alignment target for interaction with the user (“determining, by the at least one processor and based on the context, a level of safety {alignment target} of a content generated by the AI character model” for “content to be provided to the user”; Gelfenbeyn, ¶ [0100]); based on the alignment target... generating the alignment strategy for interaction with the user (further discloses “adjusting, by the at least one processor and based on the level of safety, the safety settings associated with the content generated by the AI character model” which can include “selecting topics that are allowed or disallowed to be discussed by the AI character with the user, selecting a type/level of profanity language that is allowed or disallowed for use by the AI character when interacting with the user, allowing or disallowing specific words or phrases, and so forth,” all of which are strategies for the level of safety {alignment targets}; Gelfenbeyn, ¶ [0102]); and generating the alignment information by combining the user portrait, the alignment target, and the alignment strategy (“the heuristics models 620” including “the goals model 622, the safety model 624, the intent recognition model 626, the emotion model 628, and the events model 630” may “provide intermediate outputs” and “the orchestration step (step C) may include formatting and representation 634 of the intermediate outputs received from the heuristics models” for incorporation into “dialogue prompts 636” which “may be provided to an LLM 646.”; Gelfenbeyn, ¶ [0068], [0070]). However, Gelfenbeyn fails to expressly recite the use of one or more prompts for the generation of the above user portrait, alignment target and alignment strategy.
The relevance of Li is described above with relation to claim 6. Regarding claim 8, Li teaches generating a first prompt for a user portrait based on the user input (Discloses including “filling a prompt word template” {generating a prompt}, using “dialogue parameters” and a “prompt word template library”, where, at runtime, the prompt template is filled with user information by the “dialogue robot... by using natural language... [and] performing semantic analysis”; Li, ¶ [0037]-[0038], [0044], [0065]); generating the user portrait by providing the first prompt to the LLM (Further teaches “importing the prompt into the large language model” to obtain the “dialogue related user personality characteristic label set”, where the feature values are understood as the user portrait in the context of instant application.; Li, ¶ [0048]); based on the user portrait, generating a second prompt for an alignment target for interaction with the user (The system uses the determined feature values to select and generate further “topic tags” and related questions of “a multi-round dialogue corpus,” which, in combination, discloses using the “obtain[ed] portrait traits” (Who) to generate a second prompt to determine the user’s specific safety needs (the target, or “what”); Li, ¶ [0040], [0060]); generating the alignment target by providing the second prompt to the LLM (Includes “calling the large language model to simulate different roles for mutual dialogue” to generate the next round of user-specific data, which discloses using the LLM to generatively produce the specific safety/alignment targets.; Li, ¶ [0059]); based on the alignment target, generating a third prompt for an alignment strategy for interaction with the user (Discloses “automatically matching and generating an implication prompt word template according to the personality characteristic label Cj of the user related to the j-th dialogue” which uses the “implication [or entailment] prompt” to “update the dialogue generation strategy” based on previously determined values, which is understood as using the determined needs (e.g., safety/alignment targets) to build a third prompt, that produces the safety/alignment strategy.; Li, ¶ [0050]); generating the alignment strategy by providing the third prompt to the LLM (Discloses “importing the implication prompt word template into the large language model to obtain the following implication relation between the characteristic values Vj and Vnew” and “updating the dialogue generation strategy {alignment strategy}” in light of alignment between Vj and Vnew; Li, ¶ [0050], [0061]); and generating the alignment information by combining the user portrait, the alignment target, and the alignment strategy (Discloses “establishing personality label set” to “complete the user portrait” for the target user, which is understood as generating the alignment information by combining the user portrait, the alignment target, and the alignment strategy.; Li, ¶ [0065]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Li to include generating a first prompt for a user portrait based on the user input; generating the user portrait by providing the first prompt to the LLM; based on the user portrait, generating a second prompt for an alignment target for interaction with the user; generating the alignment target by providing the second prompt to the LLM; based on the alignment target, generating a third prompt for an alignment strategy for interaction with the user; generating the alignment strategy by providing the third prompt to the LLM; and generating the alignment information by combining the user portrait, the alignment target, and the alignment strategy. Gelfenbeyn discloses that the alignment is to be “adapted dynamically based on real-time context” (Gelfenbeyn, ¶ [0033]). However, Gelfenbeyn fails to provide a clear mechanism for achieving this functional goal. Li describes the use of a large language model to iteratively generate the indicated alignment information for a user, as part of a user portrait generation method which captures the user’s evolving inner state, thus providing a clear mechanism to achieve the dynamic adaptation referenced in Gelfenbeyn, and providing the known benefit of “greatly improving the accuracy and instantaneity of user portraits,” while maintaining convenience “for practical application and popularization”, as recognized by Li. (Li, ¶ [0008]).
Regarding claim 9, the rejection of claim 6 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. However, Gelfenbeyn fails to expressly recite further comprising: adjusting the alignment prompt based on feedback to the answer generated by the LLM or an evaluation on the alignment information; and updating the alignment information using the adjusted alignment prompt.
The relevance of Li is described above with relation to claim 6. Regarding claim 9, Li teaches further comprising: adjusting the alignment prompt based on feedback to the answer generated by the LLM or an evaluation on the alignment information (The system includes “automatically... generating an implication prompt word template according to the personality characteristic label Cj of the user related to the j-th dialogue” where the “pre-trained model automatically adjusts dialogue strategy based on current state”; Li, ¶ [0050], [0065]); and updating the alignment information using the adjusted alignment prompt (Discloses “importing the implication prompt word template into the large language model to obtain the following implication relation between the characteristic values Vj and Vnew” and “updating the dialogue generation strategy” in light of alignment between Vj and Vnew; Li, ¶ [0050]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Li to include further comprising: adjusting the alignment prompt based on feedback to the answer generated by the LLM or an evaluation on the alignment information; and updating the alignment information using the adjusted alignment prompt. Gelfenbeyn discloses that the alignment is to be “adapted dynamically based on real-time context” (Gelfenbeyn, ¶ [0033]). However, Gelfenbeyn fails to provide a clear mechanism for achieving this functional goal. Li describes the use of a large language model to iteratively generate the indicated alignment information for a user, as part of a user portrait generation method which captures the user’s evolving inner state, thus providing a clear mechanism to achieve the dynamic adaptation referenced in Gelfenbeyn, and providing the known benefit of “greatly improving the accuracy and instantaneity of user portraits,” while maintaining convenience “for practical application and popularization”, as recognized by Li. (Li, ¶ [0008]).
Regarding claim 16, the rejection of claim 11 is incorporated. Claim 16 is substantially the same as claim 6 and is therefore rejected under the same rationale as above.
Regarding claim 17, the rejection of claim 16 is incorporated. Claim 17 is substantially the same as claim 7 and is therefore rejected under the same rationale as above.
Regarding claim 18, the rejection of claim 11 is incorporated. Claim 18 is substantially the same as claim 8 and is therefore rejected under the same rationale as above.
Regarding claim 19, the rejection of claim 16 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. However, Gelfenbeyn fails to expressly recite wherein the actions further comprise: adjusting the alignment prompt in real time based on feedback to the answer generated by the LLM or an evaluation on the alignment information; and updating the alignment information in real time using the adjusted alignment prompt.
The relevance of Li is described above with relation to claim 6. Regarding claim 19, Li teaches wherein the actions further comprise: adjusting the alignment prompt in real time based on feedback to the answer generated by the LLM or an evaluation on the alignment information (The system includes “automatically... generating an implication prompt word template according to the personality characteristic label Cj of the user related to the j-th dialogue” where the “pre-trained model automatically adjusts dialogue strategy based on current state”; Li, ¶ [0050], [0065]); and updating the alignment information in real time using the adjusted alignment prompt (Discloses “importing the implication prompt word template into the large language model to obtain the following implication relation between the characteristic values Vj and Vnew” and “updating the dialogue generation strategy” in light of alignment between Vj and Vnew; Li, ¶ [0050]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Li to include wherein the actions further comprise: adjusting the alignment prompt in real time based on feedback to the answer generated by the LLM or an evaluation on the alignment information; and updating the alignment information in real time using the adjusted alignment prompt. Gelfenbeyn discloses that the alignment is to be “adapted dynamically based on real-time context” (Gelfenbeyn, ¶ [0033]). However, Gelfenbeyn fails to provide a clear mechanism for achieving this functional goal. Li describes the use of a large language model to iteratively generate the indicated alignment information for a user, as part of a user portrait generation method which captures the user’s evolving inner state, thus providing a clear mechanism to achieve the dynamic adaptation referenced in Gelfenbeyn, and providing the known benefit of “greatly improving the accuracy and instantaneity of user portraits,” while maintaining convenience “for practical application and popularization”, as recognized by Li. (Li, ¶ [0008]).
Claims 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gelfenbeyn and Hazra as applied to claim 1 above, and further in view of Poulis (U.S. Pat. No. 12,182,678, hereinafter Poulis).
Regarding claim 10, the rejection of claim 1 is incorporated. Gelfenbeyn and Hazra disclose all of the elements of the current invention as stated above. However, Gelfenbeyn fails to expressly recite further comprising: obtaining at least a portion of the alignment information by at least one of transfer learning, multi-task learning, and knowledge graphs.
Poulis teaches “systems and method for responsible LMMs that align with domain specific principles.” (Poulis, Col. 1, lines 25-30). Regarding claim 10, Poulis teaches further comprising: obtaining at least a portion of the alignment information by at least one of transfer learning, multi-task learning, and knowledge graphs (“an LLM...may be aligned by the system… by generating instructions (See FIG. 2) that are fine tuned in order to align the LLM/LMM to a set of principles specific for each of the more than one domains,” where “domain specific data 402 may be ingested (by domain alignment data ingestion 406) to generate a domain alignment knowledge graph (KG) 408” which may be used to maintain “alignment with the principles of the specific domain that may be determined based on the query posed by the user”; Poulis, ¶ Col. 4, lines 20-30; Col. 26, lines 32-41).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the behavioral characteristic safety settings for AI models of Gelfenbeyn, as modified by the heuristic LLM models of Hazra, to incorporate the teachings of Poulis to include further comprising: obtaining at least a portion of the alignment information by at least one of transfer learning, multi-task learning, and knowledge graphs. Gelfenbeyn establishes a requirement for determining alignment information (safety levels and settings) dynamically based on real-time user context. However, Gelfenbeyn relies on the use of heuristic model to select levels of safety in achieving this goal. Poulis describes the alignment of foundation models with domain-specific principles, and uses an automated mechanism for obtaining the alignment information, by generating a “Domain Alignment Knowledge Graph (KG) 408” from ingested domain data. (Poulis, ¶ Col. 26, lines 35-47). In the context of Gelfenbeyn, such a knowledge graph improves the precision and reliability of the determined safety settings, where the use of a knowledge graph allows Gelfenbeyn’s real-time adjustments to be grounded in verified, domain-specific facts and structured principles, rather than just broad heuristics, which improves alignment quality and ensures the responses received from the LLM are both more contextually accurate and safe for a specific user.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Gelfenbeyn (U.S. Pat. App. Pub. No. 2023/0351216) discloses an artificial Intelligence (AI) character model with modifiable behavioral characteristics including modifying, in response to the determination that the event has occurred and based on information associated with the event, parameters of the AI character model to obtain further parameters associated with behavioral characteristics of the AI character model and causing the AI character model to interact with the user according to the further parameters.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Sean E. Serraguard whose telephone number is (313)446-6627. The examiner can normally be reached 07:00-17:00 M-F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Daniel C. Washburn can be reached at (571) 272-5551. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Sean E Serraguard/Primary Examiner, Art Unit 2657
Prosecution Insights