RESUBMIT of tracking number: 791-1782 Jan 30th, 2025 Dear editors-in-chief of Neurosymbolic Artificial Intelligence: We appreciate the insightful feedback provided by the reviewers on our submitted paper titled "Cognitive LLMs: Toward Human-Like Artificial Intelligence by Integrating Cognitive Architectures and Large Language Models for Manufacturing Decision-Making." The comments were very helpful to enhance the quality of our work. In this letter, we first address the general feedback provided by each reviewer. In the revision log message titled "Changes Made to Paper", we present lists of each comment from the reviewers alongside the corresponding changes we have implemented in the paper. Responses to General Comments from Reviewer 1 (Major Revision): First, we acknowledge the feedback regarding the unclear and repetitive writing, as well as the lack of organization. We have taken the following actions, resulting in a reduction of the paper’s length from 29 pages to 23 pages (15 main + 3 references + 5 appendix): (1) Delete redundant writing: We ensured that no same content appears more than once. (2) Delete repetitive figures: Figures with unclear annotations or limited information were either removed or redesigned. For example, the newly created Fig. 1 synthesizes the previous Fig. 1 and Fig. 7, with a detailed description provided in the introduction. This redesigned Fig.1 offers a clear, concise overview of the Cognitive-LLMs architecture at the outset of the paper. (3) Streamline and reorganize the paper: We streamlined and rewrote the abstract. Additionally, the paper now has a streamlined organization comprising the following sections in sequential order: introduction, research questions, related work, Cognitive-LLMs architecture, LLM-ACTR knowledge transfer framework, experiments conducted to address research questions, results, conclusions, and discussions. (4) Revision of unclear notations: We have rewritten sections with unclear notations, as highlighted by Reviewer 1. For example, the section on Reinforcement Learning in Production Systems has been rewritten and retitled to include a clearer explanation of utility update theory incorporating metacognition. Second, we understand the reviewer's concerns on the comparative performance benefits of fine-tuning LLMs with ACT-R traces versus using ACT-R models alone. To address this, we have clarified why ACT-R alone is not the baseline in this study by explaining the role of ACT-R in Cognitive LLMs and the rationale for our choice of using pre trained LLMs as the baseline. (1) ACT-R as a synthetic agent to instruct LLMs through training: We have clarified the motivation behind creating Cognitive LLMs, we aim to develop a hybrid architecture, Cognitive-LLMs, that leverages the natural language processing and generative capabilities of LLMs, complemented by the human-like learning and reasoning offered by ACT-R. Therefore, we propose a synergistic approach where ACT-R models serve as synthetic agents guiding the training of LLMs. In this context, ACT-R is used as a grounding tool for enhancing LLMs' trustworthy inference rather than as a baseline. (2) Using pre-trained LLMs as baseline: To assess the Cognitive LLMs' ability to make human-like decisions akin to ACT-R, the baseline for comparisons in this study hence is pre trained LLM to demonstrate the enhancements brought by integrating cognitive architectures into LLMs. Third, we understand the reviewer's critique regarding the value of our methodology and have improved content to explain how our approach of integrating CAs and LLMs differs from others, as well as the contributions it offers. (1) How cognitive LLMs differ from other integration approaches: Following recent findings that LLMs’ embeddings can be trained to predict human behaviors, this paper adopts an approach by leveraging CAs to ground the decisions of LLMs in a data-driven manner using machine learning and deep learning methods. Our aim is to examine the properties of a neural network representation of the decision-making process in CAs and to investigate whether knowledge from CAs can be preserved in an embedding space and infused into LLMs through transfer learning. (2) Why it matters: Transfer of learning has proven effective in applications such as text sentiment and image classification. Our experimental results show that the knowledge of CAs in decisions such as learning can be transferred to LLMs through fine-tuning, and the holistic cognitive decision process has the potential to be transferred through finetuning and activation engineering. The results open up new research directions for equipping LLMs with the necessary knowledge to computationally model and replicate the internal mechanisms of human cognitive decision-making from a data-driven perspective. Last but not least, we appreciate the referenced literature provided by the reviewer. We have carefully reviewed these sources and integrated selected ones, as listed in the appendix. Responses to General Comments from Reviewer 2 (Accept): First, we revised the confused reference to VSM-ACTR 2.0 as VSM-ACTR uniformly in the paper and instead cite the previous version of the model as VSM-ACTR 1.0 to avoid confusion. Second, the scope of this work has been defined as toward trustworthy decision-making by LLMs in manufacturing. We ask whether LLMs can replicate cognition from Cognitive Architectures (CAs) to make human-like decisions. Third, the empirical results primarily use negative log-likelihood, which is a common chosen measurement of goodness of fit in machine learning. Some of the results are empirically significant, e.g., the LLM with fine-tuning compared to the pre-trained LLM. However, preliminary results show limited improvement and warrant further investigation. We candidly discuss this, proposing possible reasons and pointing out potential solutions. Last but not least, we appreciate the reviewer's suggestion on discussing the approach's theoretical limitations. We addressed this through rewriting limitations and further work section. Thank you very much and please let us know if you have any questions! Sincerely Yours, Siyu, Alessandro, Jonathan, Lee, and Frank