ICWLM: In-Context Wireless Models

• ICWLM is a paradigm where large, frozen models adapt to wireless tasks using contextual inputs rather than traditional fine-tuning.
• It employs techniques such as prompt-based adaptation, expert tool integration, and retrieval-augmented learning to improve tasks like localization, power control, and modulation classification.
• Empirical studies show substantial performance gains, highlighting that shifting intelligence from weight updates to context construction can enhance system adaptability and efficiency.

Searching arXiv for the provided ICWLM-related papers and closely related work to ground the article. In-Context Wireless Large Model (ICWLM) denotes a family of wireless intelligence architectures in which large models acquire task- or environment-specific capability at inference time from context rather than from repeated task-specific retraining. In the recent literature, the term is used in two closely related senses. One is a broad paradigm in which a frozen large model is conditioned by prompts, retrieved examples, external expert outputs, or learned prompt tokens to perform wireless tasks such as environment understanding, power control, localization, modulation classification, and covert communication (Liu et al., 3 May 2025). The other is a narrower, wireless-native formulation in which a transformer foundation model is trained directly on large-scale mixed wireless datasets and then adapts to new physical-layer tasks by in-context learning from a few demonstration pairs (Wen et al., 24 Jul 2025). Across both senses, the defining idea is that adaptation is shifted from weight updates to context construction, retrieval, prompting, and tool orchestration (Pan et al., 1 Jun 2026).

1. Terminological scope and research emergence

The emergence of ICWLM is tied to a recurring limitation in applying generic LLMs to wireless systems: such models have no native access to physical-layer signals, lack inductive biases for log-scale metrics, complex-valued operations, and wireless physics, and may therefore produce physically impossible values, add dB values linearly, or ignore phase when asked to reason directly over wireless quantities (Liu et al., 3 May 2025). At the same time, conventional task-specific deep learning in wireless communication often suffers from data scarcity, poor generalization under changing environments, and fragmentation across separate models for separate tasks (Wen et al., 24 Jul 2025).

Recent work uses the phrase “ICWLM” or explicitly identifies specific systems as concrete instantiations of it. The MCP-based Internet of Experts for wireless environment-aware agents is described as “exactly such an ICWLM realization” (Liu et al., 3 May 2025). ProWin is presented as a prototype in which a general-purpose LLM becomes a wireless optimizer purely through task-specific prompts and reinforced example selection (Zhou et al., 6 Jun 2025). RA-LWLM is described as a concrete instantiation of the same paradigm for cross-scene localization, with scene adaptation happening through retrieved fingerprints rather than per-scene retraining (Pan et al., 1 Jun 2026). A separate line of work uses the term in a stricter sense for a wireless-native decoder-only transformer trained from scratch on mixed wireless datasets for precoding and channel prediction (Wen et al., 24 Jul 2025).

A theoretical precursor is the formulation of wireless symbol estimation as an in-context learning problem. “In-context estimation” models the prompt as a sequence of received observations and known symbols under a shared latent context, and proves that, for a subclass of such problems, a single-layer softmax attention transformer computes the optimal solution in the limit of large prompt length (Kunde et al., 2023). This suggests that ICWLM is not only a systems-engineering metaphor but also has a formal estimation-theoretic basis in wireless communications.

2. Core architectural patterns

ICWLM systems are unified less by one canonical network topology than by a shared adaptation mechanism: the backbone remains fixed or largely fixed, while wireless competence is supplied through contextualized inputs at inference time. Representative formalisms make this explicit. In the MCP-based Internet of Experts, each wireless expert is exposed as a tool implementing $$\mathcal{E}_m(\mathbf{h}) = p(s_m \mid \mathbf{h}) \in [0,1]$$ for attributes such as line-of-sight, Doppler, Rayleigh fading, or Rician fading (Liu et al., 3 May 2025). In context-aware token communication, the receiver updates each token through an iterative MAP rule,

$$\hat{w}_i^{(l)} = \arg\max_{w_i} P(\mathbf{y}_i \mid w_i)\, P(w_i \mid \hat{\mathbf{w}_{\backslash i}^{(l-1)}}),$$

which fuses channel likelihood with a language-model prior (Shin et al., 25 Jan 2026). In RA-LWLM, a frozen wireless foundation-model encoder provides scene-agnostic CSI representations, retrieval supplies per-scene references, and a transformer in-context module predicts UE position from the query and retrieved examples (Pan et al., 1 Jun 2026).

Realization	Context source	Adaptation mechanism
MCP-based IoX	Expert outputs over wireless attributes	Tool calls and context augmentation
ProWin	Rewarded state–action–reward examples	Reinforced in-context prompting
RA-LWLM	Retrieved CSI–position fingerprints	Retrieval-augmented ICL
RFPrompt	Learnable deep prompt tokens	Prompt-based adaptation of frozen LWM
WirelessAgent	Memory, RAG, tools, workflow state	Agentic planning and action
Wireless-native ICWLM	Demonstration pairs over tasks	In-context multi-task inference

These patterns span several distinct technical mechanisms. Tool-augmented systems expose lightweight classifiers or numerical solvers through standardized interfaces, often with the large model acting as a planner and interpreter rather than a direct signal processor (Liu et al., 3 May 2025). Retrieval-augmented systems externalize scene-specific or domain-specific knowledge into databases or knowledge bases and condition inference on the retrieved context rather than on task-specific retraining (Pan et al., 1 Jun 2026). Prompt-based wireless foundation-model adaptation keeps the backbone frozen and injects expert-specific prompt tokens into each transformer layer, thereby steering attention without overwriting pretraining structure (Uddin et al., 5 May 2026). Wireless context engineering generalizes the same logic beyond textual prompting: context is any information that conditions perception, reasoning, and action beyond instantaneous signal observations, and it may be represented as structured tokens, latent states, or retrieved map entries under latency, energy, and memory constraints (Zhao et al., 7 Feb 2026).

3. Representative realizations across the wireless stack

At the physical layer, ICWLM appears in multiple forms. The MCP-based Internet of Experts equips mainstream LLMs with wireless environment-aware reasoning by registering lightweight attribute classifiers as MCP tools; the host asks the LLM which experts are relevant, invokes them through JSON-RPC, injects their confidence scores back into the context window, and lets the frozen LLM produce classification outputs or natural-language analyses (Liu et al., 3 May 2025). In symbol detection, received I/Q samples are serialized into prompts together with a few labeled examples, and LLMs such as GPT-J and LLaMA-2 act as few-shot demodulators; contextual calibration and linear probe calibration are then used to stabilize accuracy and confidence (Abbas et al., 2024). In token communication, a shared BERT-like masked LLM, i.e. BERT from Devlin et al. (2019), serves simultaneously as transmitter-side predictability estimator and receiver-side contextual prior, enabling iterative token detection and masked transmission (Shin et al., 25 Jan 2026). In wireless-native multi-task ICWLM, a decoder-only transformer trained from scratch jointly handles weighted sum-rate precoding, max-min SINR precoding, and downlink channel prediction through demonstration-based adaptation (Wen et al., 24 Jul 2025).

At the network optimization layer, the same paradigm shifts from signal interpretation to control. ProWin formulates base-station power control as inference over task descriptions, selected examples, and current state, using discrete-state matching or continuous-state ranking to select demonstrations from an experience pool without any model training or fine-tuning (Zhou et al., 6 Jun 2025). WirelessAgent organizes resource management as an agentic workflow with perception, memory, planning, and action modules, implemented in LangGraph and demonstrated on network slicing with tool use, retrieval, and reflection (Tong et al., 2 May 2025). Shadow Wireless Intelligence extends the pattern to covert communications: DeepSeek-R1 reasons over retrieved formulas, constraints, and code templates to derive symbolic optimization steps and simulation code for artificial-noise power allocation in a full-duplex covert communication setting (Xie et al., 7 May 2025).

Localization and sensing supply a third major realization. RA-LWLM combines a frozen wireless foundation-model encoder, Euclidean KNN retrieval in representation space, and a transformer-based in-context localization module with a mixture-of-experts selector over context sizes to achieve training-free cross-scene adaptation (Pan et al., 1 Jun 2026). The broader context-engineering literature interprets beam prediction and related tasks as a question of choosing which histories, sensor modalities, maps, and side information should enter the model’s context under inference-time constraints, effectively making wireless context construction the analogue of prompt management (Zhao et al., 7 Feb 2026).

Data-centric uses also fit the ICWLM pattern. LLM-AUG generates synthetic samples directly in a learned embedding space for modulation and interference classification, avoiding training task-specific generators and relying instead on in-context generation from a few example embeddings (Gajjar et al., 20 Apr 2026). RFPrompt adapts a frozen mixture-of-experts wireless foundation model to out-of-distribution modulation classification by inserting learnable deep prompt tokens at every transformer layer and expert, keeping the backbone frozen and updating only a small prompt-plus-head parameter set (Uddin et al., 5 May 2026).

4. Empirical evidence

The empirical record is heterogeneous because different papers target different tasks, but several quantitative results recur. In wireless environment awareness, MCP-based IoX transforms mainstream LLMs from near-chance or modest attribute classifiers into high-accuracy agents (Liu et al., 3 May 2025).

LLM agent	Raw accuracy (%)	+MCP accuracy (%)
DeepSeek-Chat	46.7	95.5
DeepSeek-Reasoner	54.2	97.5
ChatGPT-3.5	46.6	95.8
ChatGPT-4	53.5	96.9
O4-Mini	59.2	98.1
QWQ-Plus	53.8	98.0
Qwen-Plus	51.2	96.1
Qwen-Turbo	45.8	95.8

These gains amount to about 40–50 percentage points and support a specific interpretation: standalone LLMs are poor at mapping raw numerical channel vectors to deterministic labels, but once they receive expert-derived perceptions as structured context, their task reduces to interpreting confidence scores, thresholding, and assembling structured outputs (Liu et al., 3 May 2025).

Other realizations show analogous improvements. In RA-LWLM, median localization error is 0.49 m on seen scenes and 0.53 m on unseen scenes, with the latter only 8.2% worse than the former, while scene-shared end-to-end baselines degrade sharply on unseen scenes (Pan et al., 1 Jun 2026). In RFPrompt, prompt-based adaptation uses 73,728 prompt parameters and about 0.34% trainable parameters overall, yet reaches 52.6% accuracy at 2 shots per class and 83.3% at 128 shots per class on the Real-World IQ benchmark, substantially outperforming from-scratch and meta-learning baselines in the OOD setting (Uddin et al., 5 May 2026). LLM-AUG reaches 90% of oracle performance with about 15% labeled data, yields a 29.4% relative gain over diffusion-based augmentation at a lower SNR value, and reports relative gains of 67.6% on RadioML and 35.7% on the interference-classification dataset over the diffusion baseline (Gajjar et al., 20 Apr 2026).

At the control layer, ProWin is reported to outperform RL-based methods on base-station power control and to converge to average reward around 4 with average power about 3 W in the considered continuous-state setup, while Llama3-8B and Llama3-70B achieve comparable reward and service quality to the DRL baseline in the discrete-state case (Zhou et al., 6 Jun 2025). WirelessAgent achieves 44.4% higher bandwidth utilization than the prompt-based method and performs only 4.3% below the rule-based optimum in network slicing, while supporting 25 users compared with 15 for the prompt-only baseline and 26 for the rule-based solution (Tong et al., 2 May 2025). In covert communications, Shadow Wireless Intelligence reports 85% symbolic derivation accuracy and 94% correctness in simulation-code generation for DeepSeek-R1 under the format-validation setting (Xie et al., 7 May 2025).

5. Relation to neighboring paradigms and recurrent misconceptions

A recurrent source of confusion is that ICWLM is not a single architecture. Some work uses the term for frozen general-purpose LLMs augmented with prompts, retrieval, tools, or expert models (Liu et al., 3 May 2025), whereas other work reserves it for a wireless-native transformer foundation model trained directly on mixed wireless datasets from scratch (Wen et al., 24 Jul 2025). These are not mutually exclusive; rather, they occupy different points in a design space whose axes are pretraining domain, adaptation mechanism, and degree of task externalization.

A second misconception is that ICWLM simply means “wireless fine-tuning of an LLM.” Several papers explicitly position the paradigm against fine-tuned wireless LLMs. The MCP-based IoX keeps the foundation LLM frozen and moves domain adaptation into small, replaceable experts (Liu et al., 3 May 2025). RA-LWLM externalizes scene-specific information into a per-scene fingerprint database rather than encoding it in weights (Pan et al., 1 Jun 2026). RFPrompt adapts a frozen backbone through prompt tokens rather than full parameter updates (Uddin et al., 5 May 2026). This suggests that ICWLM should be understood less as a new training recipe than as a reallocation of intelligence from weights to context.

A third misconception is that more context is always better. The context-engineering literature states explicitly that excess context can dilute task-critical cues, introduce noise or staleness, and harm performance under latency, memory, and bandwidth constraints (Zhao et al., 7 Feb 2026). LLM-AUG reports a related effect: increasing the number of in-context examples helps up to a point, but too many examples can degrade performance, which the paper associates with context overload and the “lost in the middle” phenomenon (Gajjar et al., 20 Apr 2026). Efficient ICWLM design therefore depends on context selection, compression, ordering, and retrieval quality, not only on model scale.

6. Limitations and open directions

The literature repeatedly emphasizes that current ICWLM results remain bounded by deployment realism. The MCP-based Internet of Experts is evaluated on synthetic channels rather than over-the-air measurements, and the paper explicitly identifies end-to-end latency under real-time constraints, cost-aware expert selection, and security and privacy of the tool chain as open problems (Liu et al., 3 May 2025). Token-communication schemes require transmitter and receiver to share the exact same MLM and tokenizer, assume perfect CSI, and incur multiple BERT passes per packet, making latency and synchronization critical issues (Shin et al., 25 Jan 2026). RA-LWLM points to smarter retrieval, adaptive database construction, temporal in-context localization, and real-world measurements as immediate next steps (Pan et al., 1 Jun 2026).

Wireless-native ICWLM introduces another set of challenges. The current multi-task model covers only weighted sum-rate precoding, max-min SINR precoding, and channel prediction, so extension to broader PHY and cross-layer tasks remains open (Wen et al., 24 Jul 2025). Compression, pruning, quantization, and distillation are natural next steps for edge deployment. More broadly, wireless context engineering frames the long-term problem as one of context acquisition, structuring, compression, persistence, and delivery under strict budgets, implying that future ICWLM systems will need explicit context controllers and memory hierarchies rather than ever-larger monolithic models (Zhao et al., 7 Feb 2026).

A plausible implication is that ICWLM will continue to bifurcate into two complementary tracks. One track will emphasize wireless-native foundation models that learn reusable channel and control representations directly from large wireless corpora (Alikhani et al., 2024, Wen et al., 24 Jul 2025). The other will emphasize orchestration of frozen general models through retrieval, tools, expert calls, and workflow memory (Tong et al., 2 May 2025, Xie et al., 7 May 2025). The present literature suggests that the most capable systems may combine both: a wireless-native representational backbone with explicit in-context control over experts, memories, and task-specific context.