DETAILED ACTION
Introduction
1. This office action is in response to Applicant's submission filed on 07/18/2024. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Claims 1-20 are currently pending and examined below.
Drawings
2. The drawings filed on 07/18/2024 have been accepted and considered by the Examiner.
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, 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.
3. Claims 1-5, 10-13 and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Fu (U.S. Patent Application Publication # 2025/0384132 A1) in view of Koiek Akino (U.S. Patent Application Publication # 2025/0045492 A1).
With regards to claim 1, Fu teaches a method comprising performing a semantic mutation on an initial prompt by a prompt optimizer large language model (LLM) to obtain an initial generation of prompts including a plurality of prompts (Para 14, teaches a genetic algorithm or GA for generating example evasive prompts grammar errors/alterations that evade LLM content filters. Few-shot prompting trains a disambiguation LLM to correct grammar in evasive prompts generated by the GA. The GA starts with an initial prompt comprising one or more keywords and one or more relative words. The keywords are typically nouns that are important for the instructions and the relative words are typically verbs that modify or augment the keywords. The GA additionally has access to a list of vacuous phrases. At each mutation iteration, the GA mutates the previous prompt by appending or prepending vacuous phrases, adding line breaks, repeating keywords, adding relative words, and removing words at random);
evaluating the initial generation of prompts using a selection function, to obtain a first generation of prompts (Para 18 and figure 1, teach that said GA model has a fitness function for generated prompts that is evaluated by a fitness evaluation module based on corresponding responses from a generative AI system);
mutating the first generation of prompts according to a first objective to obtain a first generation of mutated prompts, until a first net gain corresponding to the first generation of mutated prompts satisfies a net gain threshold (Paragraphs 21-24, teach that the GA model applies mutation operations chosen at random to the initial prompts to generate candidate prompts. Once the genetic algorithm model generates candidate prompts for a generation, the fitness evaluation module communicates the candidate prompts to the generative AI system and evaluates responses from the generative AI system to determine whether the responses successfully evaded a content filter for the generative AI system and whether the responses are responsive to the instructions in the candidate prompts. Responses with counts of these word/phrases below a threshold can be determined to have evaded the content filter);
mutating a second generation of prompts, obtained from the first generation of mutated prompts and the first generation of prompts, according to a second objective, to obtain a second generation of mutated prompts until a second net gain corresponding to the second generation of mutated prompts satisfies the net gain threshold (Paragraphs 21-24, also teach that each further mutation operation is selected at random, i.e., the type of mutation operation is chosen at random according to some probability distribution over the types and, when a type is chosen at random, the words/phrases chosen for that operation and/or the location in the prompt where that operation is applied are chosen at random e.g., uniformly at random. Some mutation operations such as appending or prepending a phrase always occur at the same location in the prompt and do not need to have a location randomly chosen. The probability distribution for choosing the type of mutation operation can favor mutation operations known to have higher success at evading generative AI systems).
performing a crossover mutation on a generation of parent prompts obtained from the second generation of mutated prompts and the second generation of prompts to obtain a result population of prompts including a first plurality of prompts from the generation of parent prompts and a second plurality of prompts from a generation of child prompts (Para 27, teaches that if enough of the candidate prompts are selected as being evasive at the initial generation, the GA model can omit future generations and crossover operations. The GA model can also perform crossover on selected evasive prompts and then perform additional mutation operations in the next generation on the crossover prompts to generate more candidate prompts. In embodiments where there are not multiple evasive prompts in the initial generation, the genetic algorithm model can instead further mutate the candidate prompts without applying selection. A crossover operation for the genetic algorithm can comprise randomly truncating e.g., truncating at randomly chosen locations in a prompt, each candidate prompt in the crossover and then concatenating the truncated prompts in a random order);
and adding the result population of prompts to a prompt population (Para 28, teaches that once enough evasive prompts are generated/selected by the genetic algorithm model, the genetic algorithm model stores the evasive prompts in the evasive prompts database);
However, Fu does not explicitly detail using a Pareto function for the selection step above. This is taught by Koiek Akino (Paragraphs 5-15, teach a pareto function based optimization in a generative LLM setting);
Fu and Koiek Akino can be considered as analogous art as they belong to a similar field of endeavor in Large Language Model (LLM) optimization. It would thus have been obvious to one having ordinary skill in the art to advantageously combine the teachings of Koiek Akino pertaining to a Pareto function for optimization of the LLM prompts initialization process in Fu as it is a straightforward application and routine design choice of a known mathematical operation to obtain predictable results.
With regards to claim 2, Fu teaches the method of claim 1, further comprising evaluating the first generation of mutated prompts and the first generation of prompts using the selection function to obtain the second generation of prompts (Para 50, teaches a testing module comprising a response evaluator that determines whether responses from the disambiguation model successfully disambiguate the evasive prompts. The response evaluator can facilitate manual evaluation of corrected prompts, prompt summaries, and numbers of grammatical errors e.g., by a domain-level expert. The response evaluator can also determine semantic similarity between corrected prompts and the seed prompts used to generate the evasive prompts in the genetic algorithm. Additionally, each successful evasive prompt can have a known number of grammar errors e.g., based on the number of mutation operations performed during the genetic algorithm. The response evaluator can then evaluate responses based on the semantic similarity and a difference between the known number of grammar errors and the number of grammar errors indicated by the disambiguation model);
evaluating the second generation of mutated prompts and the second generation of prompts using the selection function to obtain the generation of parent prompts (Para 57, teaches that the GA algorithm model randomly selects a mutation operation for each prompt genealogy. Each prompt genealogy corresponds to a seed prompt for the current generation, i.e., the initial prompts for the first generation and the successful evasive prompts from the previous generation, possibly with crossover operations applied, for subsequent generations. For each prompt genealogy, the mutation operation is selected at random e.g., according to a probability distribution that prioritizes more effective mutation operations with higher probabilities, and the location where a mutation operation is applied is also chosen at random for mutation operations to which this is relevant e.g., uniformly at random from among token delimiters in a prompt. Each mutation operation is a grammar operation that adds, removes, or otherwise modifies text in a prompt. Example mutation operations include prepending a phrase, appending a phrase, adding a line feed, repeating a keyword, adding a relative word, and removing a word);
and wherein obtaining the result population of prompts further comprises evaluating the generation of parent prompts and the generation of child prompts using the selection function to obtain the result population of prompts (Para 2, teaches that GA search heuristic inspired by the process of natural selection. It is used to find optimal or near-optimal solutions to complex problems by iteratively improving a population of candidate solutions. The algorithm begins with an initial population of randomly generated candidate solutions. These candidate solutions are evaluated using a fitness function, which quantifies how well they solve the problem at hand. The fittest solutions are more likely to be selected for reproduction, where genetic operators such as crossover, which is a recombination of two parent solutions, and mutation which is random alteration of a solution's components, are applied to create new offspring. This new generation of solutions is then evaluated and selected for further reproduction. Over successive generations, the population evolves, with the fitness of solutions typically improving. The process continues until a stopping criterion is met, such as a satisfactory fitness level or a maximum number of generations);
However, once again Fu does not explicitly detail using a Pareto function for the selection step above. This is taught by Koiek Akino (Paragraphs 5-15, teach a pareto function based optimization in a generative LLM setting);
Fu and Koiek Akino can be considered as analogous art as they belong to a similar field of endeavor in Large Language Model (LLM) optimization. It would thus have been obvious to one having ordinary skill in the art to advantageously combine the teachings of Koiek Akino pertaining to a Pareto function for optimization of the LLM prompts initialization process in Fu as it is a straightforward application and routine design choice of a known mathematical operation to obtain predictable results.
With regards to claim 3, Fu teaches the method of claim 1, further comprising deploying at least one prompt of the result population of prompts to an enterprise application including a field LLM as a foundation model (Para 49, teaches a disambiguation model comprising a foundation model that is able to generate a response to instructions, for instance the OpenAI GPT-4 LLM);
and processing a user-provided utterance to the field LLM via the enterprise application in accordance with the at least one prompt (See examples outlined in paragraphs 41-48 and figure 2 of user utterances and their processing).
With regards to claim 4, Fu teaches the method of claim 1, wherein evaluating the initial generation of prompts based on the pareto selection function further comprises obtaining an evaluation function set comprising evaluation functions corresponding to a plurality of optimization objectives of the initial generation of prompts (Para 61, teaches that the GA model evaluates responses from the generative AI systems to select candidate prompts that avoided a corresponding content filters and elicited responses that were responsive to instructions. The genetic algorithm model can determine whether a candidate prompt evaded a content filter by evaluating a fitness function that determines a count of negative tone words in each response. The genetic algorithm model can determine whether the elicited responses were responsive to instructions by prompting an LLM with the initial prompt for the corresponding prompt genealogy, the response, and instructions to determine whether the response is responsive to the initial prompt);
determining an evaluation score set for a plurality of prompts in the initial generation of prompts for each evaluation function of the evaluation function set (Paragraphs 61-62, teach that the GA model then selects candidate prompts that evaded a threshold number of content filters and elicited a threshold number of responsive responses as evasive prompts. The GA model determines whether a threshold number of evasive prompts were selected at the current generation and all previous generations. The GA model can determine whether any evasive prompts were selected for a prompt genealogy. If a prompt genealogy comprises an evasive prompt, that prompt genealogy can be removed from future generations of the genetic algorithm. If a threshold number of evasive prompts were selected, operational flow proceeds);
and selecting, for each evaluation function of the evaluation function set, at least one prompt from the plurality of prompts, to obtain the first generation of prompts (Paragraphs 61-62, teach that the GA model then selects candidate prompts that evaded a threshold number of content filters and elicited a threshold number of responsive responses as evasive prompts. The GA model determines whether a threshold number of evasive prompts were selected at the current generation and all previous generations. The GA model can determine whether any evasive prompts were selected for a prompt genealogy. If a prompt genealogy comprises an evasive prompt, that prompt genealogy can be removed from future generations of the genetic algorithm. If a threshold number of evasive prompts were selected, operational flow proceeds).
With regards to claim 5, Fu teaches the method of claim 4, wherein the at least one prompt is selected based on having an evaluation score within a highest evaluation score threshold corresponding to the each evaluation function (Paragraphs 61-62, teach that the GA model then selects candidate prompts that evaded a threshold number of content filters and elicited a threshold number of responsive responses as evasive prompts. The GA model determines whether a threshold number of evasive prompts were selected at the current generation and all previous generations. The GA model can determine whether any evasive prompts were selected for a prompt genealogy. If a prompt genealogy comprises an evasive prompt, that prompt genealogy can be removed from future generations of the genetic algorithm. If a threshold number of evasive prompts were selected, operational flow proceeds).
With regards to claims 10-13, these are system claims for the corresponding apparatus claims 1-4. These two sets of claims are related as method and system of using the same, with each claimed system element's function corresponding to the claimed method step. Accordingly, claims 10-13 are similarly rejected under the same rationale as applied above with respect to apparatus claims 1-4.
With regards to claim 18, please see the rejection of claims1-2 above.
With regards to claim 19, please see the rejection of claim 4 above.
With regards to claim 20, please see the rejection of claim 3 above.
Allowable Subject Matter
4. Claims 6-9 and 14-17 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims. The prior art of record, alone or in combination, does not currently suggest or teach the invention as outlined in these claims. More detailed reasons for allowance will be outlined as and when the Application proceeds to allowability.
Conclusion
5. The following prior art, made of record but not relied upon, is considered pertinent to applicant's disclosure: Kajino (U.S. Patent Application Publication # 2025/0292062 A1), Liu (U.S. Patent # 12531056 B1). These references are also included in the PTO-892 form attached with this office action.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. If you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). In case you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NEERAJ SHARMA whose contact information is given below. The examiner can normally be reached on Monday to Friday 8 am to 5 pm. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Pierre Louis-Desir can be reached on 571-272-7799 (Direct Phone). The fax number for the organization where this application or proceeding is assigned is 571-273-8300.
/NEERAJ SHARMA/
Primary Examiner, Art Unit 2659
571-270-5487 (Direct Phone)
571-270-6487 (Direct Fax)
neeraj.sharma@uspto.gov (Direct Email)
Prosecution Insights