⚡ electrical engineering

MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

The paper introduces MultiPathFormer, a wireless foundation model that leverages multipath propagation as its core pretraining objective through autoregressive next-path prediction and environment-aware retrieval mechanisms, achieving superior performance in localization, beam prediction, and channel estimation compared to existing channel-based approaches.

Original authors: Blessed Guda, Kayley Sze, Carlee Joe-Wong

Published 2026-08-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Blessed Guda, Kayley Sze, Carlee Joe-Wong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand a crowded room just by listening to the total hum of voices, without being able to pick out individual speakers. That is roughly how traditional wireless networks have tried to understand the airwaves: they look at the "channel," which is a messy, tangled mix of all the radio waves bouncing off walls, cars, and people. For decades, scientists have tried to build "foundation models"—super-smart AI brains trained on massive amounts of data—to predict how these signals behave. These models are like the "GPTs" of the wireless world, designed to help with tasks like finding your location, connecting your phone to the fastest signal beam, or figuring out if you are in a direct line of sight with a tower. But there's a catch: these models have been trying to learn the "hum" instead of the "speakers." They treat the signal as a giant, confusing grid of numbers, ignoring the fact that the signal is actually made of distinct paths, like individual beams of light reflecting off mirrors.

This is where the story gets interesting. The researchers at Carnegie Mellon University asked a simple question: What if we stopped trying to memorize the messy hum and started teaching the AI to understand the individual paths instead? They realized that radio waves don't just appear; they travel, bounce, and scatter in specific ways. By treating each path as a distinct character in a story, they hoped to build a model that understands the physics of the room, not just the math of the noise. This matters because as our world gets more crowded with 5G and 6G networks, we need signals that are smarter, faster, and more accurate at navigating complex environments like cities and stadiums.

Enter MultiPathFormer, a new kind of AI foundation model that changes the rules of the game. Instead of looking at the wireless signal as a giant, tangled spreadsheet, the authors decided to treat it like a story. Imagine you are a radio wave traveling from a cell tower to your phone. You don't just go in a straight line; you might bounce off a glass building, skim the top of a car, or diffract around a corner. Each of these bounces is a "path." The MultiPathFormer model looks at a connection between a tower and a phone and sees it not as a blob of data, but as an ordered list of these paths, sorted from the strongest to the weakest.

The model is trained using a technique called "next-path prediction." Think of it like a game of "Mad Libs" or a story completion game. The AI is shown the first few paths (the strongest ones) and has to guess what the next path in the sequence will be. It has to figure out: How long did it take to get there? How strong is the signal? Did it bounce off a wall or scatter through a window? To do this, the model uses a special "backbone" (a type of neural network called a transformer) that is excellent at understanding sequences, similar to how language models understand sentences.

However, the authors realized that just guessing the next path isn't enough; the AI needs to know the "scene." If you are in a city, the paths will be different than if you are in a forest. To solve this, they added two clever tricks. First, they built an "Environmental RAG" system. This is like giving the AI a map and a list of nearby objects. Before it guesses the next path, it looks up the specific geometry of the area—where the buildings are, how tall they are, and what they are made of—to make a smarter guess. Second, they added a "First-Path Codebook." Since the very first path is usually the strongest and most important (like the main character in a story), they created a reference guide. This codebook groups users into clusters based on their location and the type of signal they usually get, helping the AI make a very accurate guess for that first, most critical path before filling in the rest of the details.

The results of this approach are quite impressive, at least in the simulations the authors ran. They trained MultiPathFormer on data from 27 different environments, using over 23 million path tokens. When they tested it, the model didn't just guess; it learned the underlying rules of how radio waves move.

In terms of accuracy, the model showed it could predict the delay and power of these paths much better than previous methods. Specifically, the combination of the environmental map and the first-path reference guide reduced the error in predicting signal delay by about 38.7% and the error in predicting signal power by a massive 59.2% compared to models that didn't use these tricks.

But the real test was whether this "path-based" brain could actually help with real-world tasks. The researchers put MultiPathFormer to work on four different challenges:

Beam Prediction: Figuring out which direction to point the antenna to get the best signal. MultiPathFormer got the top 3 correct beams right 91.4% of the time, beating the previous best models.
Localization: Figuring out exactly where a user is standing. The model made a mean error of only 5.57 meters, which is a huge improvement over the 68 to 82 meters of error seen in other models.
Line-of-Sight Classification: Determining if the user has a direct view of the tower or if something is blocking the way. The model was correct 99.4% of the time.
Channel Estimation: Reconstructing the full signal picture from partial data. The model achieved a score of 0.561, outperforming the standard models.

The authors also found that this model is very flexible. Even when they gave it a completely new environment it had never seen before, it could still perform well, especially if they gave it a little bit of extra training (fine-tuning) for that specific place. It was also fast, making predictions in just a few milliseconds, which is fast enough for real-time use.

The paper suggests that by focusing on the individual "paths" rather than the messy "channel" as a whole, we can build wireless AI that is more interpretable, more accurate, and better at understanding the physical world. The authors note that while these results are based on high-fidelity simulations (using ray-tracing software to mimic real radio waves), the next step would be to see if these gains hold up in the messy, unpredictable reality of actual city streets. For now, however, MultiPathFormer offers a promising new way to teach machines how to listen to the invisible world around us.

Technical Summary: MultiPathFormer

Problem Statement

Current wireless foundation models typically rely on Masked Channel Modelling (MCM) as their pretraining objective. In this paradigm, models are trained to reconstruct masked elements of a channel tensor (spanning subcarriers, antennas, or time). The authors argue that the channel tensor is a suboptimal pretraining object because it is a superposition of propagation paths entangled with array responses, subcarrier sampling, and system configurations. This "collapsing" of physical phenomena obscures the underlying environment-dependent structure, limiting model interpretability and its ability to disambiguate users with similar aggregate channels but different physical locations.

Furthermore, standard MCM formulations struggle to capture the fundamental physics of wireless propagation, which is governed by a variable number of continuous-valued paths (characterized by delay, power, phase, and angles) rather than a fixed grid of discrete tokens.

Methodology: MultiPathFormer

The paper proposes MultiPathFormer, an autoregressive foundation model that shifts the pretraining object from the channel tensor to the multipath propagation process.

1. Core Formulation

Path-Level Representation: Each transmitter-receiver (tx-rx) link is modeled as an ordered sequence of continuous-valued path tokens. A path token $\pi$ includes power ( $\rho$ ), delay ( $\tau$ ), phase ( $\phi$ ), angles of arrival/departure (azimuth/elevation), and a multi-hot interaction vector (Line-of-Sight, Reflection, Diffraction, Scattering).
Autoregressive Prediction: The model is pretrained using a next-path prediction objective, analogous to next-token prediction in language models but adapted for continuous regression.
- Ordering: Paths are sorted by descending power to create a deterministic sequence.
- Termination: Instead of a special end-of-sequence token, the model explicitly predicts the total number of paths ( $L$ ) to determine when to stop generation.
Architecture: The backbone is an encoder-decoder transformer.
- Encoder: Processes tx-rx positions and environmental features to predict the total path length and generate environment-grounded representations.
- Decoder: Autoregressively generates path parameters conditioned on previous paths and the encoder's environmental representation.

2. Key Architectural Innovations

To address the challenges of continuous regression and the dominance of the first path, the authors introduce two specific mechanisms:

Environmental Retrieval Augmented Generation (RAG):
- Recognizing that path formation depends on local geometry and objects along the propagation corridor, the model retrieves specific scene context rather than using a coarse global descriptor.
- Retrieval Strategy: It retrieves three types of features:
  1. Local Dense: Top- $K$ objects near the transmitter and receiver.
  2. Corridor Dense: Top- $K$ objects lying near the tx-rx propagation line.
  3. Sparse Multi-scale: Aggregate statistics (counts, max heights) over increasing radii and corridor bins to capture broader clutter density.
- These retrieved features are concatenated to form a pair-specific context vector.
First-Path Cluster Codebook:
- The first path (dominant path) is critical as errors here propagate through the autoregressive generation.
- The authors observe that receivers near a fixed transmitter share similar first-path delay and power due to shared large-scale geometry.
- Mechanism: A K-Means codebook is trained on the first-path power/delay of training users. At inference, a receiver is assigned a cluster centroid based on its proximity to training users.
- Residual Learning: The model uses this centroid as a strong initialization and only predicts the residual correction for the first path. Subsequent paths ( $t \ge 2$ ) are predicted via standard autoregression.

Key Contributions

Path-Level Autoregressive Formulation: The first wireless foundation model to treat multipath propagation as an ordered sequence of continuous tokens, moving beyond channel-tensor masking.
Environmental RAG & Codebook: The integration of a retrieval mechanism for pair-specific geometry and a cluster-based prior for the dominant path, which significantly stabilizes generation and improves physical parameter estimation.
Transferability: Demonstration that the model learns reusable representations that transfer to unseen users within pretraining environments and adapt effectively to new environments via fine-tuning.
Downstream Performance: Validation across four tasks: beam prediction, user localization, Line-of-Sight (LoS) classification, and channel estimation.

Experimental Results

The model was pretrained on 27 environments from the DeepMIMO dataset (23.9M multipath tokens) and evaluated on 31 scenarios.

Multipath Generation Quality

Accuracy: Compared to a direct autoregressive baseline, the full MultiPathFormer (+RAG) reduced Delay MAE by 38.7% and Power MAE by 59.2%.
Interaction Classification: Improved interaction type accuracy from 0.870 to 0.925 and F1 from 0.786 to 0.890.
Data Efficiency: The model maintains low error rates even with only 10% of training users, demonstrating the robustness of the codebook prior.

Downstream Task Performance

MultiPathFormer outperformed state-of-the-art channel-based foundation models (WiFo, LWM) and task-specific baselines:

Beam Prediction: Achieved 0.914 Top-3 accuracy (vs. 0.648 for LWM).
User Localization: Achieved a mean error of 5.57 m (vs. 68.79 m for WiFo).
LoS Classification: Achieved 0.994 accuracy (vs. 0.978 for LWM).
Channel Estimation: Achieved an NMSE of 0.561 (vs. 1.321 for WiFo).

Inference Cost

The model operates with low latency (approx. 3.45 ms per inference for the full +RAG variant), with the RAG and codebook modules adding negligible overhead (<1 ms) and minimal parameter increase.

Significance and Claims

The paper claims that path-level pretraining is a superior approach for wireless foundation models because it explicitly learns the physical structure of propagation (delay, angles, interactions) rather than the entangled channel tensor.

Interpretability: The model's attention mechanisms align with physical intuition; the decoder attends to global positions for the first (LoS) paths and to specific object tokens for later multipaths.
Reusability: The learned representations are not tied to specific antenna or subcarrier configurations, allowing for transfer across different users and environments.
Foundation for 6G: By capturing the underlying multipath physics, the model provides a reusable starting point for diverse downstream tasks like beam selection and localization, which are critical for mmWave and massive MIMO deployments.

The authors conclude that while the current results are based on ray-tracing simulations, the framework establishes a new paradigm for wireless foundation models that prioritizes physical interpretability and geometric consistency.