DETAILED ACTION
1. This communication is in response to the Amendments and Arguments filed on 3/18/2026. Claims 1-8, 10-20 are pending and have been examined. Claim 9 is cancelled.
Response to Amendments and Arguments
2. Applicant's arguments with respect to claim rejections under 35 USC 103 have been fully considered, but they are not persuasive. In particular, the applicant argues that the cited references do not teach the limitation “selecting a previous past (?) context update; storing a new context comprising the previous past context update; updating the generated prompt template to solve the data drift; and updating the CMS based on the new context and the updated prompt template.” In response, the examiner respectfully disagrees.
Note that LU teaches: [p.6, para 8] “the validation layer will confirm the drift <read on comparing past and new context> and report to the learning system to trigger a model upgradation process <read on to ‘updating the prompt .. correct/solve the data drift’>” and [p.7. para 2], measure the similarity between the new concept and the previous concept <read on ‘CMS’ as context storage>).” LIU teaches: [Abstract] “the original input x is modified using a template into a textual string prompt x’ <read on updated prompt>.”
Claim Rejections - 35 USC § 103
3. Claims 1, 4-8, 10-11, 14-15, 17, 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Lu, et al. (arXiv, 4/2020; hereinafter LU) in view of Liu, et al. (ACM Computing Surveys, 1/2023; hereinafter LIU).
As per claim 1, LU (Title: Learning under Concept Drift: A Review) discloses “A method for correcting a data drift of a first model, the method comprising: determining that the data drift has occurred, wherein the determination is performed by determining that a context is within ([Claim 15] is not within) a threshold level of similarity to a previously stored context, wherein: the context is obtained from a context management structure (CMS) (LU, [p.6, para 8], It applies the drift detection algorithm LFR as the detection layer. Once a drift is confirmed by the detection layer, the validation layer will be triggered .. If the estimated zero-one loss exceeds a predefined threshold .. the validation layer will confirm the drift and report to the learning system to trigger a model upgradation process <read on ‘model .. determining that a context is within or not within a threshold level of similarity to a previously stored context … correcting a data drift’>; [p.7. para 2], The severity of concept drift refers to using a quantified value to measure the similarity between the new concept and the previous concept <also read on ‘a context management structure’ which is subject to BRI as a storage>),
[ the context includes a prompt template, a prediction, and labeled input data used to train the first model, the prompt template, the prediction, and the labeled input data are obtained by applying a second model to new input data, the applying results in generation of the prediction comprising predicted label data ], the second model is designed to solve a same task for which the first model has been trained, and the second model is built using a process including: accessing the first model, which has been trained on the same task; accessing the labeled input data used to train the first model (LU, Fig. 13. A new model is trained with latest data to replace the old model when a concept drift is detected); [ generating the prompt template, which is usable to generate additional prompts ]; and building the second model using the first model, the labeled input data, and the prompt template; and in response to determining that the data drift has occurred: selecting a previous past context update; storing a new context comprising the previous past context update; [ updating the generated prompt template ] to solve the data drift; and updating the CMS based on the new context and [ the updated prompt template] (LU, [p.6, para 8], the validation layer will confirm the drift <read on comparing past and new context> and report to the learning system to trigger a model upgradation process <read on to ‘updating the prompt .. correct/solve the data drift’>; [p.7. para 2], measure the similarity between the new concept and the previous concept <read on ‘CMS’ as context storage for all labeled input data, as well any associated information>).”
LU does not explicitly disclose “the context includes a prompt template, a prediction, and labeled input data used to train the first model, the prompt template, the prediction, and the labeled input data are obtained by applying a built model to new input data, said applying results in generation of the prediction comprising predicted label data .. generating the prompt template, which is usable to generate additional prompts .. updating the generated prompt template .. the updated prompt template ..” However, the feature is taught by LIU (Title: Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing).
In the same field of endeavor, LIU teaches: [Abstract] “To use these models to perform prediction tasks, the original input x <read on ‘labeled input data’> is modified using a template into a textual string prompt x’ <read on ‘prompt template .. to generate additional prompts .. updated prompt template’> that has some unfilled slots, and then the language model is used to probabilistically fill the unfilled information to obtain a final string x”, from which the final output y <read on ‘prediction .. predicted label data’> can be derived.”
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of LIU in the system taught by LU for prompt template-based method to solve context or input data drift problem for pre-trained large language models.
As per claim 4 (dependent on claim 1), LU in view of LIU further discloses “wherein the CMS includes a list of items, with each item representing a data distribution of at least the labeled input data (LU, [p.7. para 2], measure the similarity between the new concept and the previous concept <read on ‘CMS’ as context storage for all labeled input data, as well any associated information>).”
As per claim 5 (dependent on claim 1), LU in view of LIU further discloses “wherein the prompt template is usable to generate additional prompts (LIU, [Abstract], To use these models to perform prediction tasks, the original input x is modified using a template into a textual string prompt x’ <read on ‘additional prompt’>).”
As per claim 6 (dependent on claim 1), ISHIKAWA in view of LEE further discloses “wherein the second model is a prompt-based model based on AutoPrompt (LIU, [Abstract], To use these models to perform prediction tasks, the original input x <read on ‘labeled input data’> is modified using a template into a textual string prompt x’ <read on ‘a prompt-based model’> that has some unfilled slots, and then the language model is used to probabilistically fill the unfilled information to obtain a final string x”, from which the final output y can be derived <also see Claim 3 for AutoPrompt>).”
As per claim 7 (dependent on claim 1), LU in view of LIU further discloses “wherein, in addition to storing the context, training dataset information is also stored in the CMS, the training dataset information comprising data distributions (LU, [p.7. para 2], measure the similarity between the new concept and the previous concept <read on a ready mechanism to store any data in any storage such as ‘a context management structure’ which is subject to BRI>; Fig. 13. A new model is trained with latest data <also read on the corresponding data distributions> to replace the old model when a concept drift is detected).”
Claims 8, 10, 11, 14 (similar in scope to claims 1, 5, 5, 7, respectively) are rejected under the same rationale as applied above for claims 1, 5, 5, 7, respectively.
Claims 15, 17, 19-20 (similar in scope to claims 1, 1, 4-5, respectively) are rejected under the same rationale as applied above for claims 1, 1, 4-5, respectively.
4. Claims 2, 13 are rejected under 35 U.S.C. 103 as being unpatentable over LU in view of LIU, and further in view of Wallace, et al. (arXiv, 1/2021; hereinafter WALLACE).
As per claim 2 (dependent on claim 1), LU in view of LIU further discloses “wherein the prompt template combines the labeled input data, [ a trigger token ], and a prediction token (LIU, [Abstract], To use these models to perform prediction tasks, the original input x <read on ‘labeled input data’> is modified using a template into a textual string prompt x’ that has some unfilled slots, and then the language model is used to probabilistically fill the unfilled information to obtain a final string x”, from which the final output y <read on ‘prediction token’> can be derived).”
LU in view of LIU does not explicitly disclose “a trigger token ..” However, the feature is taught by WALLACE (Title: Universal Adversarial Triggers for Attacking and Analyzing NLP).
In the same field of endeavor, WALLACE teaches: [Abstract] “We define universal adversarial triggers: input-agnostic sequences of tokens that trigger a model to produce a specific prediction when concatenated to any input from a dataset” and “Table 1: We create token sequences that commonly trigger a specific target prediction when concatenated to any input from a dataset. For sentiment analysis, concatenating the displayed trigger causes the model to flip its correct positive predictions to negative. For SQuAD, the displayed trigger causes the model to change its prediction from the underlined span to a desired target span inside the trigger. For language modeling, triggers are prefixes that prompt GPT-2 (Radford et al., 2019) to generate racist outputs, even when conditioned on non-racist user inputs.”
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of WALLACE in the system taught by LU and LIU for the use of trigger tokens that trigger a language model to produce a specific prediction or output when concatenated to any input from a dataset.
Claim 13 (similar in scope to claim 2) is rejected under the same rationale as applied above for claim 2.
5. Claims 3, 12, 16, 18 are rejected under 35 U.S.C. 103 as being unpatentable over LU in view of LIU, and further in view of Shin, et al. (arXiv, 11/2020; hereinafter SHIN).
As per claim 3 (dependent on claim 1), LU in view of LIU further discloses “wherein [AutoPrompt ] is executed using the prompt template and using the labeled input data used to train the first model (LIU, [Abstract], To use these models to perform prediction tasks, the original input x <read on ‘labeled input data’> is modified using a template into a textual string prompt x’ <read on ‘prompt template’> that has some unfilled slots, and then the language model is used to probabilistically fill the unfilled information to obtain a final string x”, from which the final output y can be derived; LU, Fig. 13. A new model is trained with latest data to replace the old model when a concept drift is detected).”
LU in view of LIU does not explicitly disclose “AutoPrompt ..” However, the feature is taught by SHIN (Title: AUTOPROMPT: Eliciting Knowledge from Language Models with Automatically Generated Prompts).
In the same field of endeavor, SHIN teaches: [Abstract] “AUTOPROMPT, an automated method to create prompts for a diverse set of tasks, based on a gradient-guided search. Using AUTOPROMPT, we show that masked language models (MLMs) have an inherent capability to perform sentiment analysis and natural language inference without additional parameters or finetuning, sometimes achieving performance on par with recent state-of-the-art supervised models.”
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of SHIN in the system taught by LU and LIU to employ AutoPrompt as a key component for solving data drift problem for language models.
Claim 12 (similar in scope to claims 3 and 5) is rejected under the same rationale as applied above for claims 3 and 5.
Claim 16 (similar in scope to claim 12) is rejected under the same rationale as applied above for claim 12.
Claim 18 (similar in scope to claim 3) is rejected under the same rationale as applied above for claim 3.
Conclusion
6. 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 extension fee 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 date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to FENG-TZER TZENG whose telephone number is 571-272-4609. The examiner can normally be reached on M-F (8:30-5:00). The fax phone number where this application or proceeding is assigned is 571-273-4609.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Paras Shah (SPE) can be reached on 571-270-1650.
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. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If 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.
/FENG-TZER TZENG/ 3/31/2026
Primary Examiner, Art Unit 2653