Testing Stimulus Equivalence in Transformer-Based Agents

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

* The discussion and conclusion section of the paper are not directly supported by the results. The results are presented in the appendix and are not discussed. The authors should integrate this discussion and results and refer directly to the experiment results.
* The authors should indicate the differences among the figures A10, A11; A12, A13, A14.
* The introduction should be revised as it does not fulfil the requirements for the introduction: it does not provide introductory materials, does not explain why the research is important, what are their application and what are the contribution of this work.

Author Response

Dear reviewer, 

We would like to express our sincere gratitude for your thoughtful and constructive comments on our manuscript. Your insights and suggestions have been invaluable in improving the clarity, depth, and overall quality of our work. We have carefully considered your comments and made the necessary revisions to address the raised points. We believe these changes have significantly strengthened the paper's clarity and impact.

Comments 1: "The introduction should be revised as it does not fulfil the requirements for the introduction: it does not provide introductory materials, does not explain why the research is important, what are their application and what are the contribution of this work".
Response 1: We wrongly considered that it was implicitly stated. Thanks to the reviewer for pointing it out. As a modification, the first paragraph of the introduction was included to explicitly state the broader context. Also, the last paragraph of the introduction, which states the objective, was modified and now includes a more direct problem statement and the contribution of this paper. We consider that the literature review encompasses theoretical, methodological and empirical elements in the subsections stimulus equivalence, transformer based models, and related work. 

Comments 2: "The discussion and conclusion section of the paper are not directly supported by the results. The results are presented in the appendix and are not discussed. The authors should integrate this discussion and results and refer directly to the experiment results. The authors should indicate the differences among the figures A10, A11; A12, A13, A14."
Response 2: We consider that this observation contributes to the improvement of clarity of the paper and we thank the reviewer for that.  In discussion, we explore the implications of the findings in the context of the current debate of FLMs capabilities and how our SE approach and the findings can contribute to this topic. However, this can be misinterpreted as an extrapolation beyond results, as the reviewer points out. To solve this issue, we modified the discussion: 
- We made explicit reference to all the figures in appendix B (A4 to A15) to highlight the pattern in our findings: a heavy reliance on reject relations in the formation of decision rules in the agents. 
- We made explicit the differential use of the terms Transformer-based models (TBM) as our models implemented, and Foundational Language Models (FLM) refers to large language models. 
- We moved the implications to current larger models  (FLMs) to the general discussion as arguments that support our claim that a contribution of our paper is that SE can be used as a tool for probing abstraction and symbolic manipulation in Language Models. Paragraphs related to few-shot learning and reversal curse were moved to contributions subsection in general discussion. 

Thank you again for your valuable contributions.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents the  analysis of stimulus equivalence in transformer based models, which is an important research task for improving AI systems.  Following are few comments about the experiments:

• How do the results compare with the existing research on non-transformer models?
• The choice of minimal models could have limited the scope, and made the results not representative of actual transformer models. It would be interesting to see the comparison of what impact would the complexity in models cause.
• Not using fine-tuning sounds good for isolating inherent capabilities of SE learning, however, it is not clear what ‘pseudo-equivalence’ is and why is it not good. It would be desirable to see the explanation on why pseudo-equivalence bad but not complementary in SE.
• Explainability is one of the important applications mentioned, which is vague. It would be interesting to see more discussion on that part.

 

Overall, the results a very well coincide with the transformers typical nature and  SE can be useful in understanding complex internal decision models. 

Author Response

Dear reviewer, 

We would like to express our sincere gratitude for your thoughtful and constructive comments on our manuscript. Your insights and suggestions have been invaluable in improving the clarity, depth, and overall quality of our work. We have carefully considered your comments and implemented the necessary changes to address the raised points. We believe these revisions have significantly enhanced the paper's contribution to the field.

Comments 1: "How do the results compare with the existing research on non-transformer models?"
Response 1: The reviewer's question highlights the importance of comparing our transformer-based models (TBMs) to previous research on non-transformer approaches. We conducted a comparative analysis of TBMs with both feedforward neural networks (FFNs) and reinforcement learning-based models (EPS and E-EPS). Our results indicate that while TBMs outperformed FFNs in terms of accuracy and generalization on specific training conditions, particularly within the linear series with select-and-reject relations, they did not fully replicate the human-like performance observed in EPS and E-EPS across all experimental paradigms. This comparative analysis is detailed in a dedicated section of the paper, where we discuss the strengths and limitations of each modeling approach in relation to stimulus equivalence.

Comments 2: "The choice of minimal models could have limited the scope, and made the results not representative of actual transformer models. It would be interesting to see the comparison of what impact would the complexity in models cause."
Response 2:  We acknowledge the reviewer's concern regarding the potential limitations imposed by using minimal models. While our study focused on establishing a foundational understanding of transformer-based models' capacity for stimulus equivalence, we recognize the importance of exploring the impact of model complexity on performance. As outlined in the future research section, upscaling the models to investigate the relationship between model size and performance is a promising avenue for future research. However, based on our current findings, we anticipate that a significant increase in model size may not necessarily lead to a substantial improvement in the ability to form equivalence classes. This is because the core challenge lies in capturing the underlying processes involved in stimulus equivalence rather than simply increasing computational power.

Comments 3: "Not using fine-tuning sounds good for isolating inherent capabilities of SE learning, however, it is not clear what ‘pseudo-equivalence’ is and why is it not good. It would be desirable to see the explanation on why pseudo-equivalence bad but not complementary in SE."
Response 3: The term "pseudo-equivalence" refers to situations where a model appears to demonstrate stimulus equivalence based on superficial cues or learned biases rather than a true understanding of the underlying relationships. This can occur due to factors such as information leakage or overfitting to specific training patterns. To isolate the inherent capabilities of TBMs in forming equivalence classes, we intentionally omitted fine-tuning, as it could introduce biases and potentially obscure the model's true learning processes. While fine-tuning might enhance overall performance on SE tasks, our primary goal was to understand the foundational abilities of these models. We have modified the paper to directly address this phenomenon without using the term "pseudo-equivalence."

Comments 4: Explainability is one of the important applications mentioned, which is vague. It would be interesting to see more discussion on that part.
Response 4: This study positions stimulus equivalence as a novel approach for probing the reasoning capabilities of artificial intelligence systems. By requiring models to form complex relationships between stimuli without explicit training, we challenge systems to demonstrate a deeper level of understanding beyond simple pattern recognition. Our findings suggest a correlation between successful performance on stimulus equivalence tasks and a model's ability to reason about abstract relationships and generalize knowledge. While further research is needed to fully explore the explainability potential of this approach, these initial results indicate that stimulus equivalence can serve as a valuable diagnostic tool for assessing model interpretability. For instance, by analyzing the patterns of errors made by a model during stimulus equivalence tasks, researchers can gain insights into the model's decision-making processes and identify potential biases or weaknesses.

Thank you again for your time and expertise.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors adjusted the paper to my suggestions.