Collaborative In-Context Learning for Indoor Radio Map Estimation

Introduction

Indoor radio map estimation (RME) underpins many wireless services—coverage planning, indoor positioning, and network rollout. Traditional approaches rely on supervised learning, demanding extensive labeled measurements and centralized model training. In many settings, however, building layouts and signal data are sensitive and cannot be pooled. The recent paper Collaborative In‑Context Learning for Indoor Radio Map Estimation proposes a novel solution: treat RME as an in‑context learning (ICL) task and enable multiple clients to collaborate without sharing raw data or updating the underlying model.

From Supervised RME to Collaborative In‑Context Learning

Supervised RME pipelines typically follow a three‑step workflow: (1) collect dense signal measurements, (2) train a task‑specific neural network, and (3) deploy the model centrally. This pipeline suffers from two major drawbacks:

1. Label cost – Acquiring high‑resolution signal maps is labor‑intensive.
2. Privacy bottleneck – Raw measurements often reveal floor plans and occupancy patterns, making data centralization risky.

In‑context learning sidesteps these issues by prompting a large pretrained model with a handful of examples (demonstrations) and letting it generate predictions for new queries. The model itself remains frozen; only the prompt changes per client. The authors extend this paradigm to a collaborative setting, where several clients each contribute private demonstrations for the same indoor scene and jointly infer the radio map.

Poly‑Ripple: Architecture and Core Ideas

Poly‑Ripple is built around three pillars:

1. Large Autoregressive Vision Models (LAVMs)

The framework leverages vision‑oriented autoregressive transformers that have been pretrained on massive image datasets. These models excel at capturing spatial dependencies, making them suitable for representing the propagation of radio signals across a floor plan.

2. Ripple‑Ordered Autoregressive Modeling

Radio propagation radiates outward from a transmitter (Tx) in a causal fashion. Poly‑Ripple mirrors this physical process by ordering the autoregressive generation in a "ripple" pattern: the model predicts signal strength for pixels progressively farther from the Tx, preserving the causal structure of the wireless channel. This ordering improves consistency and reduces error accumulation compared with naïve raster scanning.

3. Importance‑Aware Aggregation

Each client can only include a limited number of demonstrations due to the context‑length constraint of LAVMs. Poly‑Ripple addresses this by assigning an importance weight to each client’s prompt based on the relevance of its private data (e.g., similarity of building materials, transmitter power). The server aggregates the multiple predictions using these weights, yielding a final radio map that benefits from heterogeneous knowledge while respecting privacy.

Practical Takeaways

• No model updates needed – Clients interact with a frozen LAVM, simplifying deployment and avoiding costly retraining.
• Privacy‑preserving – Only demonstration patches and query coordinates are transmitted; raw measurements stay on‑device.
• Scalable collaboration – Importance‑aware aggregation works even when clients have vastly different amounts of contextual data.
• Theoretical grounding – The authors provide bounds showing that, under reasonable assumptions, the aggregated prediction converges to the optimal map as the number of clients grows.

Closing Thoughts

Poly‑Ripple demonstrates that large vision models can be repurposed for wireless signal estimation when guided by carefully designed prompts. By framing indoor RME as a collaborative ICL problem, the approach eliminates the need for costly labeled datasets and central data collection, while still delivering state‑of‑the‑art accuracy on public benchmarks. For practitioners, the key lesson is that leveraging the generative capabilities of pretrained models—combined with privacy‑aware prompt engineering—opens a new avenue for solving domain‑specific estimation tasks without the traditional overhead of supervised learning.