LatentWave Beats Pixel Reconstruction for Wireless AI Across 5G Tasks
LatentWave applies JEPA pretraining to wireless signals, showing that predicting in latent space builds more transferable representations than pixel-level reconstruction for 5G tasks.
Takeaways
- LatentWave is a wireless foundation model using JEPA pretraining, predicting masked signal regions in latent space rather than reconstructing raw pixels.
- Evaluated on 4 tasks: RF signal classification (80.9%), 5G NR positioning (2.32m error), beam prediction (63.1%), and LoS/NLoS classification (93.4%).
- Frequency masking boosts channel-related tasks (beam prediction +11pp, positioning error -16%) but hurts signal classification; region masking is better for classification.
- Per-channel patch embeddings with stochastic channel sampling allow the model to handle variable antenna counts without architectural changes.
- Outperforms the masked-modeling baseline WavesFM on beam prediction and positioning under linear probing with a frozen encoder.
- No public code or weights released yet; full architecture details and hyperparameters are available in the paper.
Building a single AI model that can handle every wireless task, from signal classification to 5G positioning, has been a long-standing goal in the telecom AI space. The dominant approach so far has been to pretrain on masked signal reconstruction, essentially asking the model to fill in missing pixels of a spectrogram. A new paper proposes a fundamentally different path, and the results suggest the wireless AI community may have been leaving performance on the table.
LatentWave is a wireless foundation model pretrained using a Joint-Embedding Predictive Architecture (JEPA), a self-supervised framework originally developed at Meta AI. Instead of reconstructing raw signal values, JEPA trains a model to predict the abstract representations of masked regions in a learned latent space. The key insight is that predicting in latent space forces the model to learn high-level semantic structure, rather than wasting capacity memorizing noise textures and fine-grained amplitude variations that rarely matter for downstream tasks.
Why Pixel Reconstruction Falls Short
Wireless foundation models have been gaining traction as an alternative to building separate specialized models for every task and environment. The dominant pretraining strategy has been masked modeling with pixel-space reconstruction. However, by requiring the model to reconstruct raw input values, the model is forced to dedicate representational capacity to low-level spectral details, noise textures, and fine-grained amplitude variations that may carry little discriminative value for downstream tasks.
Results indicate that wireless foundation models pretrained with pixel-space masking and reconstruction require either fine-tuning of several last layers or parameter-efficient adaptation to achieve competitive performance on tasks that differ substantially from the pretraining distributions. In other words, the frozen representations alone are not rich enough to generalize out of the box.
JEPA is an advanced self-supervised learning framework where, instead of trying to reconstruct an image or signal patch pixel-by-pixel, a model learns by predicting the missing parts of an input within an abstract latent space. This allows the architecture to focus on high-level semantic meaning rather than getting distracted by irrelevant microscopic details like noise in a sensor.