1
Token Communications (TokCom): A Unified AI-Native Communication Framework
arXiv:2607.17628v1 [cs.NI] 20 Jul 2026
Yaru Fu, Member, IEEE, Liang Ji, Student Member, IEEE Sabita Maharjan, Senior Member, IEEE, and Tony Q. S. Quek, Fellow, IEEE
Abstract—As artificial intelligence (AI) evolves from static perception to generative reasoning and autonomous agency, the fundamental principles of wireless communications are undergoing a paradigm shift. The classical Shannon paradigm, centered on reliable bit-level reconstruction for users, is increasingly misaligned with an emerging scenario in which the primary users of the network are interconnected AI agents. This article introduces token communications (TokCom), a novel framework that elevates tokens, i.e., the fundamental processing units of large language models (LLMs), to first-class entities for information exchange in the sixth generation wireless cellular networks (6G). We first examine the architectural transition from conventional communication systems to TokCom and identify the key challenges in implementing this transition, along with potential solution approaches. Thereafter, we present a practical case study to demonstrate the effectiveness of token sharing among heterogeneous language models. Finally, we outline promising future research directions toward realizing an AI-native, tokendriven communication paradigm suitable for 6G. Index Terms—6G, Token communications, intelligent applications, KV-cache.
I. I NTRODUCTION The evolution of wireless cellular communications has been driven primarily by the pursuit of higher data rates to satisfy ever-increasing human sensory demands [1]. From the narrowband voice services of first-generation wireless networks (1G) to the immersive 4K video capabilities of fifth-generation communication systems (5G), the fundamental optimization objective has remained largely unchanged, i.e., achieving accurate, high-fidelity reconstruction of source data at the receiver. However, the emerging vision for 6G points to a substantially different paradigm, in which the network is expected to natively support a wide range of intelligent applications [2]. In other words, we are entering an era This work was supported in part by the Team-based Research Fund under Reference No. TBRF/2024/1.10, in part by the Research Matching Grant under Reference No. CP/2025/1.1, in part by the grant from the Research Grant Council (RGC) of the Hong Kong Special Administrative Region, China, under Project Reference No. UGC/FDS16/E06/25, and in part by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Communications and Connectivity Bridging Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore. Y. Fu and L. Ji are with the School of Science and Technology, Hong Kong Metropolitan University, Hong Kong, 999077, China (e-mail: [email protected], [email protected]). S. Maharjan is with the Department of Informatics, University of Oslo, Oslo, Norway (e-mail: [email protected]). Tony Q. S. Quek is with the Department of Information Systems Technology and Design, Singapore University of Technology and Design, Singapore 487372 (e-mail: [email protected]).
characterized by agentic artificial intelligence (AI), in which autonomous entities are capable of reasoning, planning, and executing complex tasks collaboratively. In this context, the user of the network is no longer solely a human operating a handheld device, but increasingly an AI agent interacting with other agents, edge nodes, and cloud infrastructures, thereby imposing fundamentally new requirements on the design of future wireless cellular networks. The foregoing shift calls for a departure from traditional bitlevel-accurate transmission. In conventional wireless systems, the objective is to reproduce the source signal at the receiver with minimum distortion, without considering how that information is ultimately used [3]. However, for an AI agent tasked with decision-making or control, raw video frames or speech waveforms are often highly redundant intermediate representations. The agent does not act directly on individual pixels or waveform samples; instead, it operates on abstracted concepts such as object identities, scene layouts, action intents, or dialog states [4]. From this perspective, enforcing bit-accurate reconstruction of sensor data can be wasteful in both spectrum and energy, as it optimizes for human perception rather than machine utility. What such agents primarily require is the exchange of high-level knowledge, context, and intent that is directly relevant to their tasks. This observation is consistent with the motivation behind semantic communications (SemCom), which has laid an important foundation by emphasizing the transmission of meaning rather than raw data. Nonetheless, most existing SemCom approaches [4]–[6] are tightly coupled to specific tasks or modalities, relying on pre-designed encoders that extract task-dependent features. This design often limits generalizability, complicates interoperability, and makes it difficult to share or reuse learned representations across heterogeneous applications and intelligent devices [7]–[9]. Token communications (TokCom) is proposed as a practical realization of semantic communication principles tailored to the era of large language models (LLMs). In LLMs and multimodal LLMs, tokens serve as a common abstraction layer that maps diverse inputs such as text, images, audio, and even structured sensor data into a unified, high-dimensional embedding space. By treating these tokens as a universal representation of machine intelligence, TokCom introduces a standardized, modality-agnostic unit of exchange that is already well aligned with state-of-the-art AI models. This yields several concrete benefits: (i) it allows communication protocols to directly interface with pretrained foundation models without relying on feature engineering; (ii) it supports cross-modal fusion and reasoning, since different data types
2
Fig. 1: A high-level illustration of the TokCom framework.
are expressed in a shared token space; and (iii) it enables task-oriented compression, where only the most informative tokens for inference or control need to be transmitted. Under this paradigm, the network is no longer a passive bit pipe but becomes tightly coupled with computation, effectively operating as a distributed neural processing substrate. Tokens generated by one agent can be interpreted, transformed, or extended by other agents and edge/cloud models using the same underlying token semantics. This integration reduces end-to-end latency (by avoiding redundant decoding and reencoding at each hop), improves robustness (through generative reconstruction of missing tokens), and enhances resource efficiency (by prioritizing tokens that maximally impact downstream decisions). As a result, TokCom provides a technically grounded and scalable pathway for implementing AI-native, semantics-aware communication in 6G networks. However, the realization of TokCom also faces several challenges, including the design of token-aware network architectures, the integration of LLM-based processing into existing protocol stacks, and the guarantee of reliability, scalability, and security in distributed AI settings. In this article, we first present a potential end-to-end architecture for TokCom, highlighting how tokens can be incorporated as first-class entities in 6G systems. We then discuss the key technical issues and possible solutions, such as token-level resource management, generative reliability, and context-aware networking. Next, we present a case study to illustrate the effectiveness of TokCom in a representative multi-agent collaboration scenario. Finally, we conclude the article by summarizing the main insights and outlining promising directions for future research on AI-native, token-driven communications. II. A RCHITECTURE , K EY I SSUES , AND P OTENTIAL S OLUTIONS In this section, we first elaborate on the architectural structure of TokCom. We then discuss the key technical challenges
associated with its practical implementation, followed by an overview of potential solution approaches and design guidelines.
A. Architecture To fully realize the potential of TokCom, the 6G architecture must be redesigned to be token-aware rather than merely bitaware. As illustrated in Fig. 1, in a TokCom-enabled system, the end-to-end pipeline begins with the heterogeneous source layer, where multi-modal data streams such as text, vision, audio, and sensor signals are generated from the physical world. These heterogeneous inputs are first processed by multi-modal tokenization module, which maps modality-specific signals into discrete semantic tokens, thereby replacing conventional source coding. The resulting tokens are then projected into a unified embedding space, where representations from different modalities are aligned to preserve cross-modal semantic consistency. TokCom incorporates a multi-agent collaboration scheduler to support collaborative intelligence. This module selects the participating agents (e.g., LLMs and Small Language Models (SLMs)) and performs role assignment as well as other scheduling strategies. To capture contextual dependencies across token streams, the transceiver adopts a transformer architecture, and the token streams are processed by a communication encoder that prepares them for transmission over the wireless channel, taking into account both radio link conditions and task-level priorities. The token-aware communication module transmits semantically meaningful tokens rather than raw bits, so that wireless resources can be allocated according to task importance and semantic value. This tight coupling between token generation and radio interface design is essential; namely, the network is no longer transporting raw samples for human interpretation, but structured units that directly drive downstream AI reasoning.
3
The most transformative element of this architecture is the generative recovery layer. In traditional communication systems, packet loss is handled primarily through rigid retransmission protocols such as automatic repeat request, which aim to restore bit-level correctness [10]. In contrast, TokCom introduces the notion of generative reliability, where the goal is to preserve task performance rather than exact symbol recovery. When a portion of the token stream is lost due to channel fading or interference, the receiver does not necessarily trigger an immediate retransmission. Instead, it invokes a local generative model, i.e., the inpainter, to infer the missing tokens from the surrounding token context and its internal priors. This process is analogous to how humans can reliably understand a sentence even when some words are muffled or missing, by leveraging linguistic structure and semantic expectations. By exploiting the predictive capabilities of modern foundation models, TokCom can maintain high task-level or collaboration performance (e.g., in reasoning, control, or coordination) even in lossy or interference-limited environments. In many AI-native applications, slight deviations in intermediate tokens that do not affect the final decision are acceptable, making strict bit-perfect delivery unnecessarily costly [11]. Generative reliability therefore allows the system to trade a small amount of symbol-level accuracy for significant gains in spectral efficiency, latency, and robustness. Conceptually, this shifts part of the reliability burden from the physical and link layers to an intelligence layer, where errors are repaired or compensated using learned world models, enabling 6G systems to operate effectively in regimes that would traditionally be considered too noisy for reliable communication. Another critical enabler is distributed Key-Value (KV) cache repository. In LLM inference, the KV cache stores the intermediate attention states associated with previously processed tokens, thereby accelerating the generation of subsequent tokens and preserving conversational or reasoning context [12]. In a distributed TokCom system, the network itself can act as a distributed KV cache. Edge nodes in an AI-enabled radio access network (AI-RAN) can maintain the KV caches of ongoing agentic sessions, allowing mobile or distributed AI agents to hand off their reasoning state as they move across cells or offload computation between edge and cloud. This form of context-aware networking ensures that AI agents do not need to restart their reasoning or dialog from scratch after each handover or offload event, significantly reducing both latency and computational redundancy. More broadly, it marks a conceptual shift from networks that simply forward packets to networks that can remember, predict, and actively assist in the execution of intelligent tasks. Such a tokenaware, cache-enabled architecture provides a concrete and technically grounded pathway for integrating communication and computation in future 6G systems. B. Key Issues and Potential Solutions 1) Semantic Integrity and the Hallucination Propagation Problem: In traditional digital communication, failure is usually measured by the bit error rate. If a bit is erroneous, the
checksum does not match, and the packet is either corrected or discarded. In TokCom, the more serious problem is not an error bit, but a semantic hallucination [13]. The Token Inpainter at the receiver may fill in a missing token that is grammatically correct but factually or logically wrong, for example, turning a negative command into a positive one in a robotic control sequence. In this case, the connection still works, but the meaning is no longer trustworthy. The danger grows in multi-agent workflows, where the output of one AI system becomes the input of the next, allowing small errors to spread and snowball across the network. To reduce this risk, token-level verification protocols become essential. Instead of just checking bits, TokCom needs a semantic checksum. One idea is to send small anchor embeddings, compact numerical summaries of the intended meaning, together with the token stream. The receiver can then compare its predicted or recovered tokens with these anchors to ensure that the overall meaning remains on a reliable path. In addition, cross-model consensus can be used; for instance, several small models at the edge independently judge critical tokens and take a majority vote. As demonstrated in [14], this lightweight voting scheme provides a practical safeguard against drift, hallucination, and unreliable behaviors in generative and multi-LLM systems. 2) KV-Cache Explosion: Tokens are an efficient abstraction for efficiently representing and communicating information, but the mechanism that enables token-based processing, namely, the KV cache, introduces a substantial resource burden. To maintain conversational or task context, an AI agent must store or exchange the attention states associated with all preceding tokens. As the context window of state-of-theart models scales from thousands to millions of tokens, the KV cache can grow to several gigabytes, which far exceeds the practical memory and bandwidth budgets of typical 6G systems [12]. This leads to a new type of memory wall, as transferring context over the network can take longer than performing inference itself. To address this bottleneck, several directions can be explored. One is dynamic KV-cache orchestration with pipeline overlapping [12]. In such schemes, KV cache is dynamically managed across heterogeneous memory hierarchies (e.g., GPU and CPU), where attention states with limited contribution to the current inference step are selectively offloaded or recomputed to balance memory pressure and execution efficiency. From an edge-network perspective, AI-native edge context management will be a promising enabler. In this paradigm, 6G edge nodes (e.g., access points and base stations) act not only as communication relays but also as context repositories. When an AI agent moves from one cell to another, the network performs a context handover, pre-loading the agent’s KV cache onto the target edge node. This avoids transmitting the full history over the air and effectively turns the wireless infrastructure into a distributed, persistent memory substrate that follows mobile AI agents. This architecture naturally fits within a cloud-edge-device collaborative framework, enabling flexible partitioning of context and computation to balance latency, bandwidth, energy, and compute resources across the system.
4
3) Non-deterministic Latency and Token-Native Scheduling: Existing wireless cellular networks are primarily designed for quasi-continuous (fluid) data streams, where packet arrivals are relatively stable and predictable. TokCom traffic, in contrast, is inherently bursty and non-deterministic. For example, a LLM may emit several tokens almost instantaneously and then pause for on the order of 200 ms to perform internal computation before generating the next token [6]. Conventional scheduling algorithms (e.g., proportional fair scheduling) are not well matched to such behavior, and may result in buffer buildup, underutilized resource blocks, or both. In addition, for TokCom, the key latency metric shifts from time-to-first-byte to time-to-first-decision, i.e., the time until a useful control or semantic action can be taken. Addressing this mismatch requires token-aware radio resource management. Future TokCom-enabled 6G systems must be able to account for the semantic priority of token streams when allocating resources. For example, in an autonomous driving scenario, tokens associated with obstacle detection should bypass standard queues and be assigned predetermined resource blocks, whereas tokens conveying less time-critical information, such as environmental description, can be deferred. We further anticipate the development of predictive scheduling mechanisms, in which the edge nodes execute lightweight language models to estimate the length and timing of upcoming token bursts based on the current interaction state. This prediction will depend on sufficiently high forecasting accuracy, as the benefit of resource pre-allocation must be balanced against the computation overhead on edge devices. By forecasting the AI agent’s traffic patterns, the network can pre-allocate resources more efficiently, thereby smoothing the inherently unpredictable traffic of generative models into a more stable, predictable flow. 4) Joint Resource Management for TokCom: Although the non-deterministic scheduling in 3) primarily addresses the temporal domain of transmission, it directly translates into spatially multidimensional resource-allocation scaling challenges across the network. Conventional resource allocation aims to balance throughput and latency for bit-accurate data streams. In contrast, multi-tiered networks operating under TokCom demand a multi-resource coupled approach where spectral, computational (inference FLOPS), and storage capabilities must be jointly optimized based on the tokenized task requirements. A key challenge is the coupled resource heterogeneity, where an agent may generate low-rate status tokens for long periods, then abruptly require substantial computational and spectral resources for complex reasoning or multimodal draft-and-verify cycles. Static or purely trafficdriven schemes are unable to account for this computationalcommunicational coupling, leading to reasoning bottlenecks or inefficient resource utilization. To tackle this problem, intelligence-aware resource orchestration (IARO) is essential to dynamically manage the multitype resources. In IARO, the network leverages a multi-tier scheduling framework that jointly manages inference FLOPS, token bandwidth, and context cache as coupled resources. The edge nodes can employ reinforcement learning-based slicing to dynamically create token slices with differentiated priorities
based on the semantic criticality of the underlying task. For example, tokens associated with safety-critical embodied AI functions (such as collision avoidance) are mapped to a highpriority reasoning slice with guaranteed low-latency computeand-forward capabilities. In addition, predictive resource preemption may be realized by lightweight SLMs that estimate ongoing token trajectories. When the predictions are sufficiently accurate, the system can pre-activate edge computing or reserve bandwidth ahead of a forthcoming reasoning burst. This shift from reactive, traffic-centric control to proactive, intelligence-aware orchestration enables 6G networks to elastically match their resources to the time-varying resource requirements of distributed AI agents. III. C ASE S TUDY The TokCom framework can play a crucial role in enabling collaboration among AI agents. In this section, we present a case study on token-level communication between LLM agents, focusing on inference quality and system efficiency. This case study illustrates how TokCom enhances cooperative inference, offering a practical solution for efficient multi-agent collaboration. A. Experimental Setup In our case study, we employ an SLM-based agent with fewer than 10 billion parameters and an LLM-based agent with more than 10 billion parameters, each assigned a distinct role [15]. Specifically, the SLM agent is instantiated with the lightweight Vicuna v1.5-7B model, whereas the LLM agent is realized using the more powerful Llama 2-13B model. Both models share an identical tokenizer with a vocabulary of 32,000 tokens, thereby enabling direct token-level communication without any intermediate alignment. Multi-agent collaboration is realized at the token-sharing level. Concretely, the input prompt is first tokenized by the shared tokenizer and fed into the SLM, which performs inference and produces draft reasoning tokens that capture the logical skeleton but may contain errors. This token sequence is then directly transmitted to the LLM. The LLM concatenates these tokens with the original prompt tokens, verifies and corrects potential errors, and finally generates the refined answer. In Fig. 2(a), we illustrate the collaborative process between the SLM and LLM agents. The SLM agent is responsible for generating a step-by-step draft solution, while the LLM agent verifies and refines each step within the same token space. Owing to its relatively small parameter size (i.e., 7B), the SLM agent offers weaker reasoning capability but lower inference overhead. In contrast, the LLM agent, with a larger parameter size (i.e., 13B), provides stronger reasoning ability at the cost of higher inference latency, making it well-suited for error correction and the generation of high-quality final answers. In our experiment, we consider three multi-agent collaborative reasoning tasks, namely broad knowledge understanding, code generation, and mathematical reasoning. These tasks are evaluated on three benchmark datasets: Massive Multitask Language Understanding (MMLU), Human Evaluation (HumanEval), and Grade School Math 8K (GSM8K), respectively.
5
(a)