Landscape

Warning

Pre-implementation. This page describes proposed contracts. Behavior is subject to change before code lands.

The competitive map for RF (radio-frequency) foundation-model evaluation. The page distills which surfaces the incumbents own, where the structural gap is, and why no single incumbent can fill it. It draws on the published record and cites only the literature entries.

The gap

No incumbent owns the one-backbone-many-tasks evaluation surface for raw RF. The field has bifurcated into two camps that barely interact: a CSI (channel state information) / channel-matrix foundation-model cluster led by LWM and roughly fifteen 2026 follow-ons, and a raw-I/Q (in-phase and quadrature) camp with exactly one serious entry, IQFM. Every candidate is either a single-task dataset, a tool with no evaluation, or a model whose results are self-reported. None standardizes on raw complex multi-antenna I/Q, and none makes leave-one-environment-out OOD (out-of-distribution) transfer a first-class metric with a private held-out.

The field has named this gap itself. “Against the Monolithic Wireless World Model” (2026) states that a shared, reusable RF substrate is missing. The vacuum is perceived, not just a framing applied by EMMA (Electromagnetic Multi-task Model Assessment).

The incumbents and their limits

Incumbent

What it is

Representation

Structural limit EMMA exploits

RadioML (DeepSig)

The de-facto AMC (automatic modulation classification) dataset

single-antenna I/Q

Saturated; CC BY-NC-SA (NonCommercial) license incompatible with commercial foundation-model work; publisher states the data has known errata and is not used in its products

TorchSig / Sig53

Open-source PyTorch signals toolkit

single-antenna I/Q

Single-task (classification and detection only); a toolkit, not a governed evaluation; no leaderboard or private held-out

CSRD2025 / ChangShuoRadioData

Large-scale synthetic radio dataset for spectrum sensing

I/Q rendered to COCO spectrograms

Single-task, single-org, unhosted; validates the generator-as-test-set bet while proving the multi-task governed gap

DeepMIMO

Parameterized mmWave / massive-MIMO channel datasets

CSI / channel paths

Channel-only, not raw I/Q; cannot serve an I/Q backbone’s identity or scene tasks; no multi-task protocol or held-out

PReD-Bench

Multi-task electromagnetic benchmark and model (2026)

spectrograms and constellation diagrams

Operator-as-evaluator: the same group builds model, dataset, and benchmark and reports state of the art on its own board; phase discarded; no localization pillar

IQFM

The one raw complex I/Q foundation model

raw I/Q

Self-reported results; AoA reported only as binned classification; no benchmark, leaderboard, governance, or shared held-out; “OOD” redefined as per-task adaptation

LWM

The leading CSI-camp foundation model

CSI

Wrong input modality for raw-I/Q scoring; the CSI cluster is empirical evidence a large part of the field voted against raw I/Q

RF-Analyzer / RF-GPT

The RF-language frontier (one lab)

spectrograms routed through a vision encoder

The authors concede generalization breaks under conditions outside the synthetic distribution, particularly low-SNR (signal-to-noise ratio) regimes; phase discarded

A survey of the apparent RF foundation-model names (ORACLE, RadioBERT, RadioFM, RFMixer, CLIP-for-RF) finds no genuine RF or wireless foundation models behind those labels; the populated set is the three camps above. The raw complex multi-antenna I/Q quadrant, at benchmark scale with inter-antenna phase coherence and a neutral host, is empty.

Why EMMA’s combination is uncopyable

The defensible property is the combination of four choices, not any single feature, and the choices reinforce each other.

  1. Raw phase-coherent multi-antenna I/Q as the canonical input. Any incumbent that wants the localization task must recapture its entire dataset, because inter-antenna phase coherence is enforced by physics, not branding. The input moat is the task that justifies it.

  2. OOD-first transfer as the headline metric. Orthogonal to how every incumbent scores. Each redefines “generalization” ad hoc (IQFM as per-task adaptation; fingerprinting as a different capture day; RF-GPT as unseen waveform configs); none run systematic leave-one-environment-out against a frozen, private, multi-environment held-out.

  3. Reproducible-by-generator test set. Train and dev are re-derivable from a frozen rfgen commit, config, and seed; the held-out is a frozen secret. No static-file benchmark has this property, and it sidesteps the dying-eval-server and test-set-leak failure modes that bit ImageNet and SQuAD.

  4. Neutral ring-fenced governance. The model operator (Superpose) is structurally separated from the scoring path: third-party re-scoring, pre-registered baselines, and an independent co-steward. Neutrality cannot be manufactured retroactively, so every model-builder-as-evaluator incumbent (PReD-Bench, LWM, DeepSig’s paywalled successors) is structurally barred from replicating it.

The compounding effect is the load-bearing claim. The task that justifies the input (localization) is exactly the task that most needs a neutral host, so the input moat and the governance moat compound. No single actor can copy all four at once without recapturing data, redefining its metric, rebuilding its test set, and surrendering its own scoring authority.

References

  • DeepSig, “RadioML datasets,” datasets page (deepsig.ai/datasets). Saturated, NonCommercial-licensed, and the publisher states the data has known errata and is not used in its products; the legitimacy vacuum EMMA’s permissive CC-BY-4.0 data license inverts. (verify URL)

  • Hanna and Hussain, “Robust Low-SNR Modulation Classification,” 2026, arXiv:2605.27673. Apparent model gaps on RadioML collapse under matched hyperparameter search; grounds AMC saturation.

  • Mashaal and Abou-Zeid, “IQFM: A Raw IQ Foundation Model,” 2025, arXiv:2506.06718. The closest existing analog; self-reported, no benchmark, AoA as binned classification.

  • Djordjevic, Ali, and Alkhateeb, “LWM: A Foundation Model for Wireless Channel Data,” 2024, arXiv:2411.08872. The CSI-native lineage EMMA is orthogonal to.

  • Alkhateeb, “DeepMIMO,” 2019, arXiv:1902.06435. Channel-level datasets; the data source for the CSI camp.

  • Han et al., “PReD: Pretrained Remote-sensing foundation model” and PReD-Bench, 2026, arXiv:2603.28183. Same group builds model, dataset, and benchmark; the operator-as-evaluator cautionary tale.

  • “RF-Analyzer,” 2026, arXiv:2605.04676. Concedes sim-to-real generalization breaks in low-SNR and OOD regimes; the load-bearing sim-to-real caveat.

  • Debbah and the MBZUAI group, “RF-GPT,” 2026, arXiv:2602.14833. Spectrogram-based RF-language model; the signal-to-text frontier.

  • TorchSig, github.com/TorchDSP/torchsig. The open-source signals toolkit rfgen composes for benchmark-compatible modulations.

  • ChangShuoRadioData / CSRD2025, 2025, arXiv:2508.19552. Single-task synthetic radio dataset; validates the generator-as-test-set design.

  • “Against the Monolithic Wireless World Model,” 2026, arXiv:2605.16689. The field’s own position paper naming the missing shared RF substrate.

See Also

  • Design decisions: the load-bearing choices this map explains.

  • Governance: the ring-fence mechanism no model-builder incumbent can copy.

  • Signal as a modality: why the one-backbone-many-tasks surface is the wedge.

  • Literature: the master reading list this page cites.

  • Roadmap: the release sequence that fills the gap.

  • Open questions: the adoption signals that decide whether the gap closes.