By Anonymous User
Review Details
Reviewer has chosen to be Anonymous
Overall Impression: Weak
Content:
Technical Quality of the paper: Weak
Originality of the paper: Yes, but limited
Adequacy of the bibliography: No
Presentation:
Adequacy of the abstract: No
Introduction: background and motivation: Bad
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Weak
Detailed Comments:
This paper introduces a neuro-symbolic architecture by integrating the ACT-R Cognitive Architecture with LLMs. It claims to resolve “the dichotomy between the human-like yet constrained reasoning processes of Cognitive Architectures and the broad but often noisy inference behavior of Large Language Models (LLMs)”. It does so by the fact that it “extracts and embeds knowledge of ACT-R’s internal decision-making process as latent neural representations, injects this information into trainable LLM adapter layers, and fine-tunes the LLMs for downstream prediction”.
The topic is of course interesting. However, the paper has a number of serious problems, some of which are listed below.
The writing is unnecessarily wordy and often repetitive, making it harder to read and understand the points that the authors are making. For one instance, the abstract appears to be too long and too wordy for its content.
I suggest that the authors streamline the writing of the whole paper, avoiding repeated or scattered explanations.
The paper reads like an exhaustive lab report, not a paper ready for publication. There is often no clear enough take-home message in many sections. Critical details are not well explained or missing altogether, despite its length.
The authors thus should also restructure the paper, re-organizing the most relevant materials, and removing less relevant or irrelevant materials. For example, this has very little relevance: “Dopaminergic signals are believed to transmit reinforcement information to the corpus striatum [71], traditionally signaling reward-related activities.” Or, “Neurologically, as cognitive strategies evolve …….”. Etc. etc.
Some less relevant but useful materials may be relegated to appendices.
On p.8, the temporal difference (TD) algorithm was mentioned, but the equation followed (eqn.1) seems not TD, but just time weighted updating of U, with time steps. The notations are very confusing.
In terms of the results, is there any performance advantage in fine-tuning LLMs with ACTR traces, compared with the original ACTR model from which traces were obtained? This needs to be better analyzed and discussed in detail. The paper mentioned that it “show both improved task performance as well as improved grounded decision-making capability of our approach, compared to LLM-only baselines that leverage chain-of-thought reasoning strategies”, but not comparisons with the original ACTR model.
Methodologically, is there any advances in this paper, compared with existing work such as Trieu H. Trinh, Yuhuai Wu, Quoc V. Le, He He & Thang Luong (2023)?
In terms of the scholarship and citations of relevant previous work, the authors should make some improvements in several regards. Here are just some examples:
In terms of integrating cognitive architectures and LLMs, the authors need to cite highly relevant existing work, such as:
• Integrating LLMs with Soar: arXiv:2310.06846v1 ; etc.
• Integrating LLMs with Clarion: arXiv:2401.10444 ; arXiv:2410.20037 ; etc.
• And other cognitive architectures; Etc.
So, the authors’ claim “unlike these previous efforts that incorporate LLMs into CAs, there is currently no research focusing on assimilating the advantages of CAs into LLMs” is not accurate.
In terms of a more thorough understanding of dual-process theories (beyond just their popularization), see, for example, https://content.iospress.com/articles/neurosymbolic-artificial-intellige... and references cited therein. The authors need to present a more balanced view on such theoretical issues at the beginning of the paper (even though, understandably, this is not the focus of the paper in question), as well as their relations to neuro-symbolic systems.
By the way, just FYI, here are some further readings:
• Evans, J. & K. Frankish (eds.), (2009). In Two Minds: Dual Processes and Beyond. Oxford University Press, Oxford, UK.
• Macchi, L., M. Bagassi, & R. Viale, (eds.), (2016). Cognitive Unconscious and Human Rationality. MIT Press, Cambridge, MA.
Each of the two presents a variety of views.
Finally, with regard to cognitive architectures, only focusing on two cognitive architectures misses the big picture. The authors should provide a more complete picture by reviewing a lot more of them. One way of accomplishing that is citing some existing comprehensive reviews of cognitive architectures, such as Taatgen & Anderson (2023), or Kotseruba & Tsotsos (2020), etc. etc.
