- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This work proposes "Neurosymbolic Multi-head attention",. A "role embedding" is computed as , where and are learned. A "role-conditioned" representation is computed as , where is the elementwise product. Then, "neurosymbolic multi-head attention" is computed as a softmax of of inner products between added to inner products of . This is framed as capturing a "role-conditioned structural bias". The proposed architecture is evaluated on a number of benchmarks and tassks.
Critical strengths
- The paper seeks to contribute to the literature on structural inductive biases in neural networks and attention mechanisms, which is an area that is of interest to the ICLR community.
Critical weaknesses
There are several significant conceptual and experimental issues.
- The paper presents an incomplete and inaccurate view of the prior literature. It frames prior work as being limited to “external modules” or “fixed structural signals”. This is inaccurate and ignores a large literature on learned, in-attention structural biases. See the list below, generated with the help of AI, for a selection of relevant work that is not discussed or address. Of relevance, [1,2] propose a mechanism which is essentially the role embedding of Eqn 2. These works are neither discussed nor compared against in the experiments.
- The experimental methodology and results suggest significant issues regarding soundness. The proposed architecture appears to have over twice the parameter count and compute requirements of the standard attention baseline. Many claimed improvements are very small (e.g., in SCAN vs , and in LAMBDADA vs ). Additionally values reported are not comparable to the literature, and suggest methodological issues. For example, the perplexity reported for WikiText-2 is on the order of 500 whereas typical values in the literature are 20-100. This suggests severe implementation or methodological issues.
- Several choices in the architecture design are generally not very well motivated or justified, and it is unclear in what formal sense the architecture is "neuro-symbolic". In several places, an architectural component is described then given a label that implies a neurosymbolic function (e.g, "role", etc.), but there isn't a clear argument for why this should be interpretted as functionally analogous to neurosymbolic computation, and there isn't mechanistic interpretability evidence that it learns such a functional role.
References
[Generated with the help of AI when prompted to find the most relevant work on relational and symbolic inductive biases in transformers. Listed most-recent first.]
- Altabaa, A., & Lafferty, J. (2025). Disentangling and Integrating Relational and Sensory Information in Transformer Architectures. ICML 2025.
- Altabaa, A., Webb, T. W., Cohen, J. D., & Lafferty, J. (2024). Abstractors and Relational Cross-Attention: An Inductive Bias for Explicit Relational Reasoning in Transformers. ICLR 2024. arXiv:2304.00195.
- Mittal, S., Raparthy, S. C., Rish, I., Bengio, Y., & Lajoie, G. (2022). Compositional Attention: Disentangling Search and Retrieval. ICLR 2022. arXiv:2110.09419.
- Lamb, A., He, D., Goyal, A., Ke, G., Liao, C.-F., Ravanelli, M., & Bengio, Y. (2021). Transformers with Competitive Ensembles of Independent Mechanisms. arXiv:2103.00336.
Questions for authors
- Please compare your experimental results to results appearing in the literature and explain the discrepency.
- Please discuss and compare your work to prior work on structural inductive biases in transformers.
3 comments
Thank you for the authors for their rebuttal. Please seem my responses to each point below.
Question 1: results compared with the literature
Our perplexities are not comparable with published WikiText-2 numbers, and the revised paper says so (Section 4.1). Ours are per token over GPT-2 subword tokens, from a single layer model with no positional encoding, trained on separate lines cut at 128 tokens. Published values of 20 to 100 come from much stronger models, longer context and usually word level perplexity. We also report the final epoch, when validation loss has already started to rise (Figure 2), so our number is higher than the best epoch. These differences explain the gap. Both standard MHA and proposed NS-MHA models use the same evaluation code and settings, so the comparison between them, which is what the paper is about, is not affected. We would also like to clarify that we do not claim a competitive language model.
This seems like a significant issue with the experimental methodology. You should use an experimental design that aligns with the literature so that it can be compared meaningfully. Single-layer models without positional encoding is poor experimental design for the goals of this experiment.
Parameters and compute:
The attention layer of NS-MHA has about 1.8 times the parameters of the standard layer, but the full model has only 1.5% more parameters on WikiText-2, LAMBADA and SCAN and 2.4% more on CLUTRR, because the embeddings and the output layer make up most of it (Table 2). The extra compute is a constant factor of about two on the attention term only, with the same asymptotic cost (Section 5.5). As a result, the model does not have twice the parameters or twice the compute overall.
Again, this highlights a significant issue in experimental design. The embedding layer should not account for >90% of the model's parameters.
** Question 2 and Prior work**
Thank you for revising the paper to improve the discussion of prior work.
Overall, this paper has major experimental design issues, confirmed by the rebuttals. These issues seem fundamental and do not seem fixable during the rebuttal stage. I encourage the authors to redesign and re-run their suite of experiments to test their ideas properly. Additionally, the paper should describe these experimental details so that the reader can evaluate its soundness and interpret the reported results in the proper context.
We would like to sincerely thank the reviewer for their time and effort put in providing us with constructive feedback for our work. We acknowledge the limitations/concerns raised. Given the lack of time to address all of them in this rebuttal, we will use them to improve our work in the future.
Thank you for the review. We answer the reviewer’s questions and then the three weaknesses. Two of the points, the parameter count and the perplexity, rest on numbers that we can clarify.
Note: It must also be noted that during the revision of the paper, we identified an error in the code for the CLUTRR experiments that affected both the MHA and NS-MHA results. Results for CLUTRR have now been updated accordingly in the revised paper.
Question 1: results compared with the literature. Our perplexities are not comparable with published WikiText-2 numbers, and the revised paper says so (Section 4.1). Ours are per token over GPT-2 subword tokens, from a single layer model with no positional encoding, trained on separate lines cut at 128 tokens. Published values of 20 to 100 come from much stronger models, longer context and usually word level perplexity. We also report the final epoch, when validation loss has already started to rise (Figure 2), so our number is higher than the best epoch. These differences explain the gap. Both standard MHA and proposed NS-MHA models use the same evaluation code and settings, so the comparison between them, which is what the paper is about, is not affected. We would also like to clarify that we do not claim a competitive language model.
Parameters and compute: The attention layer of NS-MHA has about 1.8 times the parameters of the standard layer, but the full model has only 1.5% more parameters on WikiText-2, LAMBADA and SCAN and 2.4% more on CLUTRR, because the embeddings and the output layer make up most of it (Table 2). The extra compute is a constant factor of about two on the attention term only, with the same asymptotic cost (Section 5.5). As a result, the model does not have twice the parameters or twice the compute overall.
Question 2 and Prior work: We agree with the reviewer that the mechanism in Eq. 2 has close precedent, most of all the learned library of symbols in Abstractors and the parallel pathways of the Dual Attention Transformer. Our initial version did discuss several methods that put structure inside attention (Relational Attention, the TP Transformer, the Structure Aware Transformer, Slot Attention), but it missed the papers listed by the reviewer and its summary sentence was too broad. The revised Section 2 now covers Abstractors, the Dual Attention Transformer, Compositional Attention and Transformers with Competitive Ensembles of Independent Mechanisms. The paper does not claim a learned role bank as new. Its claim is the combination of components inside one differentiable layer, and an experimental comparison with these methods is left as future work, as explained in the revised limitations section (Section 5.6).
Small improvements: We agree with the reviewer and the revised paper no longer calls them improvements. On WikiText-2 and LAMBADA it states the differences as NS-MHA matching MHA. On CLUTRR the accuracy is similar for both models (14.69 and 14.73 percent), and per token accuracy on SCAN is about 31 percent for both.
The neuro symbolic label: We agree that the paper gave no evidence for it, and the revised version uses the term as motivation only. It is now stated that we have no evidence that the learned roles correspond to symbolic or linguistic categories, and that an equally accurate name is role conditioned structured attention. Analysing what the roles capture is listed as a future work.
With the claims now limited to what the experiments show, we hope the soundness concerns are addressed.