- Rating
- 2(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper examines whether reconstruction quality reliably ranks discrete tokenizers for autoregressive wearable/physiological signal generation.
Across three tokenizers and five datasets, it reports inconsistent reconstruction–generation rankings and a positive association between erroneous codeword distance and single-token decoded distortion. HierCode freezes the tokenizer, clusters its codebook embeddings, autoregressively predicts cluster tokens and uses a second model to recover fine tokens in parallel. Perturbed cluster sequences during training and a soft cluster prior at inference are intended to improve robustness to coarse prediction errors.
Critical strengths
- A relevant representation-selection question. Reconstruction and token predictability are different objectives. The investigation across tokenizer types and codebook sizes provides useful evidence that selecting representations solely by reconstruction error can be misleading for the tested generation task.
- A plausible generation-side intervention with repeated evaluation. Freezing the tokenizer isolates the generation side redesign, while the soft prior permits recovery from a wrong coarse group rather than imposing an irreversible hard restriction.
Critical weaknesses
Core experimental definitions and conclusions are internally inconsistent. Appendix Eq. (8) averages the absolute difference of 50-bin normalized histograms. If these are probability masses summing to one, MDD cannot exceed 2/50 = 0.04, whereas Table 2 reports values such as 0.1049 and 0.2595. Histogram densities or an omitted scaling factor could explain this, but the implementation must be specified. In addition, Sec. 5.2 claims that perturbation improves VQ-VAE gMSE, while Table 3 changes from 0.9573 for Soft Posterior to 0.9871 for HierCode, a deterioration. Appendix Table 9 reports RQ-VAE HierCode average MDD of 0.1021, although its five entries average to 0.09874, consistent with Table 2’s 0.0987. These issues affect interpretation and component attribution, not merely wording.
Questions for authors
- Implementation/reporting clarification: What exactly are the histogram normalization, bin edges/widths and scaling used for MDD? Please reconcile Eq. (8), Table 2 and Appendix Table 9, and correct the VQ-VAE perturbation interpretation in Sec. 5.2 or identify which results should replace the conflicting entries. These explanations could materially change the soundness assessment.
- Requested attribution control: With the same tokenizer, two models and training budget, does geometric clustering outperform a random or non-geometric grouping? A limited comparison on representative existing settings would help separate geometry from the additional recovery stage and capacity.