Skip to content

Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis

Home / Papers / Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis

Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis

Authors: Bo Cui, Xiaowen Song, Yaowen Zhang, Shunzhe Zhang, B. J. F. van Beijnum, Monique Tabak, Ying Wang Date: 2026-05-13 Paper ID: openalex:2605.13248

Summary

Compact Latent Manifold Translation (CLMT) is a parameter-efficient (0.09B) foundation model designed for cross-modal and cross-frequency physiological signal synthesis. The framework employs a two-stage paradigm: first, a Universal Tokenizer using Hierarchical Residual Vector Quantization (RVQ) generates discrete latent manifolds to resolve modality entanglement. Second, a Context-Prompted Latent Translator performs signal synthesis as a discrete sequence translation task, significantly reducing computational requirements while outperforming large-scale baselines.

Key Contributions

  • Proposes CLMT, a 0.09B parameter model that achieves state-of-the-art results for PPG-to-ECG synthesis and cross-frequency super-resolution.
  • Introduces a Universal Tokenizer using Hierarchical Residual Vector Quantization (RVQ) to isolate and decouple heterogeneous signal latent manifolds.
  • Improves clinical R-peak detection F1-score from 0.37 to 0.83 in PPG-to-ECG synthesis and achieves a 0.9956 Pearson correlation in 25Hz to 100Hz super-resolution.

Open Questions & Future Work

Key Concepts

Archivist Review

Approved the CLMT framework as a novel approach to discrete signal translation and selected two open questions concerning clinical robustness and conditioning limitations. The review applied a strict filter for clinical applicability, rejecting generic baseline improvement claims and focusing on the core architectural innovations.

Approved Concepts

  • Compact Latent Manifold Translation (CLMT): It is the primary novel architecture/paradigm proposed, framing physiological signal synthesis as a discrete sequence translation task.

Approved Open Questions

  • Robustness to rare pathologies: Addressing this limitation is critical for the clinical reliability of generative physiological models, ensuring they do not suppress or blur life-saving diagnostic information during synthesis.
  • Mitigating static-data dependency: Reducing this dependency is essential for deploying these models in real-world, unpredictable clinical settings where comprehensive patient data is not always accessible.

Metadata & Links

url
https://arxiv.org/abs/2605.13248
paper_id
2605.13248
paper_source
openalex
domain
nlp
tags
time-seriesfoundation-modelscross-modal-learningsignal-processing
architectures
datasets
concept_slugs
compact-latent-manifold-translation-clmt
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-16T06:03:08Z