PaperlayerView exact revision on arXiv

Source-linked AI summary

From Semantic to Token Communication: The Next Paradigm for Large-Model-Driven 6G Intelligent Connectivity

Yu Ma, Zhen Gao, Li Qiao, Xiaoyuan Zhang, Mahdi Boloursaz Mashhadi, Yin Xu, Wenjun Xu, Xiaodong Xu, Kaibin Huang, Jiangzhou Wang, Rahim Tafazolli, Sheng Chen, Tony Q. S. Quek, Ping Zhang

arXiv:2609.10714v1eess.SPcs.AIcs.CVcs.ITcs.LGmath.IT

TL;DR

6G services demand high-capacity, low-latency communication, while conventional systems prioritize accurate symbol reconstruction. This survey frames TokenCom around LM-native tokens above bits and reports that, for image transmission, computation replaces retransmission only in sufficiently poor channels, with no general online choice policy identified.

  • Problem

    6G services demand extreme traffic density, ultra-high data rates, and sub-millisecond latency, while conventional systems prioritize accurate symbol reconstruction.

    Source: §1 · p. 1

  • Method

    The survey frames TokenCom as using LM-native tokens as a fundamental communication unit above bits, linking data, inference, and transmission.

    Source: §1 · p. 2

  • Results

    For image transmission, computation replaces retransmission only when the channel is sufficiently poor, and surveyed studies offer no general online policy for choosing between them.

    Source: §5.8 · p. 32

  • Takeaways & Limitations

    Context-bearing token streams can tolerate less reliable UDP and relaxed flow and congestion control, as outlined by Qiao et al. [12].

    Source: §5.8 · p. 32

    Realizing TokenCom still requires advances in privacy and security, efficiency, interoperability, robustness, token-level QoS, error handling, and service-aware scheduling.

    Source: §7 · p. 35

Abstract

from arXiv · show

The ambitious requirements of sixth-generation (6G) networks are driving communication systems from reliable bit delivery toward meaning-aware and task-oriented connectivity. Large models (LMs), with strong multimodal understanding and generation capabilities, have accelerated this shift and made semantic communication (SemCom) increasingly practical. Yet current LM-driven SemCom remains fragmented: semantic representations are typically tied to specific modalities, models, or tasks. While the bit provides a universal unit for digital transport, there is still no analogous unit for representing and processing semantics, which limits interoperability, theoretical unification, and scalable system design. We argue that tokens provide a natural candidate for this missing abstraction. Two trends support this: unified multimodal LMs now encode text, images, audio, video, and robot actions in one token space, while distributed LM inference already generates substantial token-level traffic through expert routing, cache transfer, and speculative decoding. Token communication (TokenCom) emerges by unifying these trends, using the LM's native processing unit as a communication abstraction above the bit level and enabling importance assignment, error handling, and resource allocation directly at token granularity. This survey traces the evolution from LM-driven SemCom to TokenCom. We review three major directions of LM-driven SemCom: source-centric semantic coding, channel semantics for physical-layer tasks, and collaborative edge-device intelligence. We then examine the token abstraction, the transmission techniques it requires, and two emerging paradigms, namely TokenCom for LM services and for embodied and agentic intelligence. Finally, we identify open challenges toward unified, scalable, and AI-native 6G communication systems.

1 Introduction

The introduction argues that bit-centric communication is increasingly insufficient for meaning-aware, task-oriented 6G services, while LM-driven SemCom lacks a unified representation unit. It presents LM-native tokens as a candidate shared interface and frames the survey’s review of established SemCom directions and emerging TokenCom paradigms.

1 Introduction

  • TokenCom adopts LM-native tokens as a modality-agnostic communication unit above bits, linking data, inference, and transmission in a unified framework.The introduction presents this as a candidate abstraction for representing and processing multimodal semantics.Source: §1 · p. 2
  • LM-driven SemCom spans source-centric semantic coding, channel semantics for physical-layer tasks, and collaborative edge-device intelligence.Source-centric methods shift reconstruction from bit fidelity to semantic synthesis; channel-semantic methods treat CSI as a semantic source, while edge-device work studies model selection, offloading, and collaborative inference.Source: §1 · pp. 2–3
  • For LM services, token delivery supports distributed inference through KV-cache transfer, MoE expert routing, and speculative-decoding exchanges.The survey describes efficient token scheduling and delivery as important to end-to-end LM service quality.Source: §1 · p. 3
  • For embodied and agentic intelligence, communication forms a perception-action loop and is assessed by task execution and control stability rather than reconstruction fidelity.Transmission is triggered by what the receiving agent still needs to act.Source: §1 · p. 3
  • Existing SemCom surveys leave the token-centric perspective and both emerging TokenCom paradigms substantively uncovered, despite extensive coverage of source-centric semantic coding.The introduction also identifies only scattered attention to channel semantics for physical-layer tasks.Source: §1 · p. 3
  • The survey reviews LM fundamentals and tokens, systematizes LM-driven SemCom, and discusses TokenCom’s potential and limitations as a scalable, model-aware paradigm.Its stated contributions emphasize the lack of a unified representation unit and identify research opportunities.Source: §1 · p. 4

2 Large Models and Tokens

This section introduces large-model architectures and token representations as background for treating tokens as a communication unit, including how tokens unify modalities. It distinguishes discrete and continuous tokens and the corresponding transmission approaches: digital bit streams and analog JSCC-style transmission, respectively.

2.1 Model Architecture

  • LMs learn patterns from large-scale pretraining, and in SemCom, representation models encode compact multimodal semantics while generative models decode or synthesize content.The overview identifies massive data, transformer architectures, and pretraining followed by task-specific fine-tuning as factors underpinning LM effectiveness.Source: §2.1 · p. 5

2.1.1 Transformer-Based Models

  • Transformer self-attention forms context-aware representations by weighting information across positions, while decoder-only models generate tokens autoregressively using past context.Multi-head attention captures relationships across representation subspaces; causal masking ensures each token attends only to preceding context.Source: §2.1.1 · pp. 6–7
  • Multimodal transformers project modality-specific features into a shared token space and jointly process them with text, enabling integration of heterogeneous inputs.LLaVA [37] is an example of this approach; its visual features are aligned with the text token space.Source: §2.1.1 · p. 7

2.1.2 Other Generative AI Models

  • GANs, VAEs, and diffusion models offer alternative generation paradigms, with diffusion models serving SemCom as decoders that reconstruct content from compact received semantics.Diffusion iteratively denoises; DDIMs [41] reduce sampling steps, while latent diffusion [42] operates in a compact latent space and uses text conditioning.Source: §2.1.2 · pp. 7–8

2.2 What Are Tokens?

  • Tokens unify modalities in a shared semantic space, tolerate some channel errors through model recovery, and can be extended to support diverse applications.They can be transmitted as symbol indices or embedding vectors, and extended through modality-specific projectors or added vocabulary symbols.Source: §2.2 · p. 9

2.2.1 Discrete Tokens

  • Discrete tokens are enumerable and compatible with digital bit streams, enabling entropy coding, per-token modeling and retransmission, while LM context can conceal corrupted or erased tokens.For non-text inputs, modality-specific features can be quantized to learned codebook entries and mapped to vectors by an embedding lookup.Source: §2.2.1 · p. 9

2.2.2 Continuous Tokens

  • Continuous tokens preserve dense semantic information and can degrade gracefully under analog channel noise, while shared embedding spaces ease multimodal integration.Different modality-specific projectors can map audio, vision, and sensor data into the same semantic space.Source: §2.2.2 · pp. 10–11
  • Discrete tokens transmit naturally as digital bit streams, while continuous-token embedding elements can map directly to channel symbols for analog JSCC transmission.The contrast is described from a communication perspective.Source: §2.2.2 · p. 11

3 Ubiquitous Intelligence in 6G Networks

6G networks are envisioned as collaborative platforms where LMs support perception, decision-making, and execution across connected devices. Heterogeneous device capabilities lead to coexisting bit and token traffic, while TokenCom brings distinctive traffic and differentiated service requirements.

3 Ubiquitous Intelligence in 6G Networks

  • 6G networks are envisioned to connect people, objects, and data through computing-power networks that integrate perception, decision-making, and execution.As connected devices proliferate, analytics are expected to move from centralized clouds toward the network edge, with base stations processing multimodal sensor data for resource-limited devices.Source: §3 · pp. 11–12
  • Heterogeneous AI capabilities mean raw-data bit streams and semantic-data token streams coexist, with base stations or cloud clusters processing bits into or refining semantic information.Low-capability devices may lack semantic processing, others extract partial features, and powerful nodes can complete semantic inference; edge-cloud cooperation supports flexible processing across resources.Source: §3 · pp. 11–12, Figure 6 · p. 12
  • TokenCom is an emerging traffic paradigm for 6G AI-native services, driven particularly by generative AI and intelligent-agent interactions.A company contribution to 3GPP RAN1 describes token-based services as having traffic patterns distinct from conventional packet-oriented models [60].Source: §3 · p. 12
  • Text tokens typically require success rates up to 99.9%, while visual tokens may tolerate 80–99%, motivating token-level quality differentiation in 6G design.Token traffic also combines extremely high arrival rates with small payloads, challenging control overhead, scheduling efficiency, and small-packet transmission.Source: §3 · p. 12

4 Semantic Communication

This section reviews LM-driven semantic communication across source coding, physical-layer channel semantics, and collaborative edge intelligence, tracing developments in wireless-specific models, edge-cloud inference, and task- and scenario-aware communication. It describes the move toward pre-trained LMs and motivates TokenCom as a framework for unifying these diverse approaches and supporting LM-native services.

4.1 Source Semantics vs. Channel Semantics

  • For source semantics, LMs encode task-relevant information into compact representations, while channel semantics uses LM-learned channel features across physical-layer tasks.Source representations vary by modality and can feed machine tasks or conditional generation for human receivers; channel features support estimation, CSI feedback and prediction, signal detection, and beamforming.Source: §4.1 · pp. 13–14

4.2 Source Semantic Communication

  • At 100 µbits per pixel (bpp) at 1024 × 1024 resolution, LM-based source coding demonstrates extremely low-bitrate compression alongside controllable multimodal reconstruction.The reviewed schemes incorporate channel coding separately in SSCC, jointly in JSCC, or across both paradigms in fusion schemes.Source: §4.2 · p. 14

4.2.1 Separate Source-Channel Coding

  • SSCC separates modality-specific semantic extraction, bitstream compression, and channel coding from generative-model recovery, supporting practical integration and human-oriented visual fidelity.Image schemes use embeddings, natural-language descriptions, or sketches; video and speech schemes extract frame, motion, content, and affective features, while 3D systems can prioritize a predicted viewing region. The modular design allows varied models and auxiliary modules, but impedes a unified architecture and remains vulnerable to the cliff effect because source and channel coding are optimized independently.Source: §4.2.1 · pp. 14, 15, 17
  • SSCC faces asynchronous multimodal transmission, weak semantic-level packet-loss correction, and latency from computationally intensive models at both ends.Timestamps can align modalities; masking-based training or contextual language models can complete missing semantics without retransmission, while edge-device collaboration can trade local computation against offloading.Source: §4.2.1 · pp. 16–17

4.2.2 Joint Source-Channel Coding

  • DJSCC jointly maps source signals to channel inputs, avoiding explicit quantization and error correction while degrading gracefully with SNR instead of exhibiting the abrupt quality drop of digital schemes.It offers an alternative to SSCC in low-SNR or bandwidth-limited regimes.Source: §4.2.2 · p. 17
  • DJSCC’s robustness strategies include adaptive coding, improved denoising, and corruption-aware training, while task-oriented designs optimize downstream performance rather than only reconstruction.Feedback-based DeepJSCC-f supports progressive refinement and variable-length transmission; reported applications include improved robustness and generalization under corruption or interference, and bandwidth reduction for VQA at comparable accuracy.Source: §4.2.2 · pp. 18–19

4.2.3 Fusion of Separate and Joint Coding

  • Fusion schemes combine SSCC’s compression flexibility and external knowledge with joint training’s robustness, including LM-assisted encoding and learned channel codecs co-adapted through alternating training.Other examples add color correction and super-resolution to underwater image transmission or use a shared codebook to send residual semantic information while retaining interpretability.Source: §4.2.3 · p. 19

4.3 Channel Semantics for Physical-Layer Tasks

  • Channel-semantic research seeks a unified LM that coordinates physical-layer tasks, either by fine-tuning pre-trained LLMs or pretraining wireless foundation models on channel data.The former brings wireless data in at fine-tuning, whereas the latter pretrains on channel data and needs little or no fine-tuning.Source: §4.3 · p. 19
  • Channel semantics approaches map wireless and environmental inputs into shared representations to support multiple physical-layer tasks.ChannelGPT [101] uses modality-specific embeddings and cross-modal fine-tuning for tasks including channel parameter generation, mapping, and wireless reasoning; Zheng et al. [103] use multiple encoding modules and instruction fine-tuning for channel prediction, multi-user precoding, and signal detection. Sheng et al. [102] map historical beam data into word embeddings and fine-tune GPT-2 for beam prediction.Source: §4.3 · pp. 19–20
  • Wireless intelligence’s real-time and signaling constraints expose the inefficiency of general-purpose LLMs, motivating channel models designed for wireless data and tasks.Sheng et al. [104] use granularity encoding and univariate decomposition for zero-shot generalization to unseen tasks; LWM [62] provides a task-agnostic, self-supervised channel representation through its CLS token.Source: §4.3 · p. 20

4.4 Collaborative Edge-Device Intelligence

  • Edge-device collaboration addresses the mismatch between resource-limited devices and costly cloud inference through task partitioning, model splitting, and migration.Ren et al. [19] dynamically offload tasks based on channel state and device capabilities; NetGPT [107] pairs edge prompt augmentation with cloud response generation. Tensor parallelism and over-the-air aggregation [108], expert caching [109], and satellite-aware task and model migration [110] support distributed inference; collaborative speculative decoding [18, 111, 112] uses device-side draft tokens verified by an edge LLM.Source: §4.4 · pp. 20–21

4.5 Task-Oriented and Scenario-Specific Communication

  • Task- and scenario-aware SemCom tailors transmitted information and system design to application requirements rather than maximizing overall semantic fidelity.Examples include task-guided image reconstruction and segmentation [66], facial-keypoint video conferencing with semantic error detection [69], dynamic-vehicle-only traffic video [113], trajectory- and saliency-guided metaverse transmission [74], and security-selected codec experts [114].Source: §4.5 · p. 21
  • Model division multiple access uses shared semantic content to improve bandwidth efficiency, while fragmentation across intelligent SemCom designs motivates TokenCom.MDMA [115] separates shared and personalized multiuser semantic information so shared content can use a common physical channel; the section points to direct token routing and transmission across distributed nodes as motivation for Section 5.Source: §4.5 · p. 21

5.1 From SemCom to TokenCom: The AI-Native Unified Interface Above the Bit Level

LM-driven SemCom lacks a unified semantic unit because modalities and models produce fragmented, incompatible representations. TokenCom proposes LM-native tokens as a shared interface above bits, aligning multimodal representation with communication and distributed inference.

5.1 From SemCom to TokenCom: The AI-Native Unified Interface Above the Bit Level

  • SemCom lacks a unified basic unit: modality-specific methods and incompatible LM representations fragment solutions and hinder integration or scaling.The identified sources are modality heterogeneity and model diversity.Source: §5.1 · p. 21
  • Tokens are already transmitted between distributed nodes as part of LM inference and serving, making TokenCom a natural response to this trend.The passage connects TokenCom to token traffic in distributed LM inference and serving.Source: §5.1 · pp. 21–22
  • Unified multimodal systems show that one token space can represent text, images, audio, video, and actions for cross-modal processing and generation.Unified-IO 2 uses a shared semantic space and a unified multimodal denoising objective.Source: §5.1 · p. 22
  • Token-level importance and error handling do not replace bits: discrete tokens travel as encoded indices over conventional channel coding and modulation.Continuous tokens instead take an analog route based on DJSCC; the discussion also frames semantic evaluation through a rate-distortion-perception trade-off.Source: §5.1 · p. 22
  • Tokens unify modality representation, LM capabilities across tasks, and communication protocols for encoding, error correction, and resource allocation.The proposed protocol mechanisms can be shared across scenarios and devices.Source: §5.1 · p. 22

5.2 Token as the Fundamental Unit of Communication

TokenCom treats LM tokens as the communication system’s semantic transmission unit, while bits remain the link’s transport unit. The section describes tokenization, contextual reconstruction, and reported image-communication results.

5.2 Token as the Fundamental Unit of Communication

  • A token is the semantic reasoning unit, while a bit is the link-transport unit; TokenCom assigns importance and handles errors at token granularity.For discrete tokens, codebook indices are carried as bits through conventional channel coding and modulation.Source: §5.2 · p. 22
  • Tokenization maps a source message into a sequence of tokens that replaces modality- and task-specific semantic features as the transmission unit.The source can be text, image, audio, video, or another modality.Source: §5.2 · pp. 22–23
  • Figure 8 depicts modality-specific tokenizers mapping heterogeneous sources into a common format for network transmission.The figure presents the “everything can be tokenized” framing.Source: Figure 8 · p. 23, §5.2 · p. 23
  • LM-native token embeddings carry contextual semantic information, allowing the LM to reconstruct missing or unreliable tokens using context and channel observations.Discrete-token embeddings come from an embedding layer; continuous-token embeddings are produced by a projector.Source: §5.2 · p. 23
  • 0.15 to 0.03 at 6 dB and 0.065 to 0.019 at 7 dB token error rate; at PER 20%, TCE gains 24.4% over baseline.The setup uses ImageNet100 at 0.039 bpp; CLIP semantic similarity stays above 0.70.Source: §5.2 · p. 23
  • Up to 67.3% TCE gains over conventional codebook-based baselines are reported for TokenCom with cross-modal information as PER increases.The baselines are Cheng2020 and VQGAN; “+ R” requests retransmission of every corrupted packet.Source: Figure 9 · p. 24

5.3 Token Selection and Compression

TokenCom compresses token sequences by deciding what to send using semantic importance or receiver-side predictability. These criteria support selective transmission, layered reconstruction, and modality-specific token designs.

5.3 Token Selection and Compression

  • Semantic importance and contextual predictability are distinct selection criteria: one measures task or reconstruction value, while the other measures how reliably the receiver can infer a token.A token may be semantically central yet predictable from its neighbors.Source: §5.3 · p. 24
  • LMECM-SC [122] uses cross-modal attention to select the top k% of tokens for downstream tasks, while adaptive merging [123, 124] combines redundant tokens and DJSCC-based selection [125, 126] adjusts the transmitted count.The merging framework is training-free; the DJSCC-based method uses end-to-end training and is reported without a cliff effect.Source: §5.3 · p. 24
  • Coarse-to-fine tokenization [57] places 155 of 680 indices in its first 7 of 10 scales, enabling prefix-based layered reconstruction and unequal error protection.Protection order still requires validation against distortion and downstream-task performance because fine scales may carry critical details.Source: §5.3 · p. 24
  • Context-aware token masking [127] uses a shared masked language model to omit low-entropy positions, restore them at the receiver, and allocate the saved channel resources to remaining tokens.The rule is task-independent and derives from a generic context model rather than task-specific attention.Source: §5.3 · pp. 24–25
  • Ada-TokenCom [58] transmits an arithmetic-coded token prefix and uses an identical pretrained model to generate the tail, with truncation length controlling rate.The method relies on later tokens being more predictable under autoregressive generation.Source: §5.3 · p. 25
  • TaiChi fuses global and high-resolution visual token streams, while Video TokenCom [128] prioritizes intent-relevant video tokens and Chen and Yang [131] reconstruct full tokens from transmitted embedding prefixes.At 6 dB SNR, Video TokenCom reports PSNR of 28.82 dB versus H.265 at 11.24 dB on a representative MCL-JCV sequence [130]; the evolving-memory method reports up to 1.09 dB PSNR improvement over prior evolving-memory baselines under varying MIMO fading conditions.Source: §5.3 · p. 25

5.4 Token Transmission and Multiple Access at the Physical Layer

Token-aware physical-layer designs connect token importance to coding, modulation, power, rate, retransmission, and multiple access. They combine semantic priorities with channel conditions and, in some cases, receiver-side token recovery.

5.4 Token Transmission and Multiple Access at the Physical Layer

  • TokenCom-UEP [133] protects tokens according to semantic significance, while H-TokCom [134] prioritizes semantic-cluster prefix bits and TONIC [136] combines utility-aware protection with receiver confidence-gated recovery.On COCO captions at 3 dB SNR, H-TokCom reports a 35.4% Sentence-BERT score improvement, from 0.206 to 0.279; TONIC outperforms listed baselines across AWGN, Rayleigh, and Rician image-classification channels.Source: §5.4 · p. 25
  • Ada-TokenCom [58] jointly selects truncation length and modulation and coding scheme under a long-term symbol budget, while a multi-user downlink jointly optimizes tokenizer, sub-channel assignment, and beamforming [137].The latter reports about 68% fewer video freezing events than H.265 for 1080p video with 16 users.Source: §5.4 · p. 25
  • A token-aware physical layer minimizes importance-weighted semantic distortion across token power, modulation, and coding choices, subject to power and symbol budgets.The formulation uses token priority weights and semantic distortion; related formulations jointly optimize channel matching, feature selection, modulation, and power [140], or quantization, subcarrier mapping, and power [141].Source: §5.4 · p. 26, Table 9 · p. 26
  • Importance-weighted waterfilling [142] makes power grow with the logarithm of token importance and inversely with channel gain, truncating allocation at zero for sufficiently poor links.Above 10 dB SNR, it lowers normalized importance-weighted mean squared error by over 7 dB and 10 dB against margin-adaptive waterfilling and equal power allocation, respectively, under joint sub-pixel and segment-level partitioning.Source: §5.4 · p. 26
  • Semantic-aware QAM design [143] places constellation points using concept co-occurrence and task relevance, and proves Gray-coded M-QAM strictly suboptimal under its criticality-weighted criterion when semantic importance and co-occurrence are non-uniform.Source: §5.4 · p. 26
  • Channel-driven adaptation pairs coding options with QPSK or 16-QAM, while Sun et al. [139] jointly use semantic importance and channel condition to set transmitted feature count and define an integrity index.The described channel adaptation derives packet error rate from offline-fitted waterfall SNR thresholds that grow logarithmically with packet length.Source: §5.4 · p. 27
  • Semantic-feature retransmission [144] replaces CRC-triggered HARQ with distortion-based selective retransmission, while a transformer variational autoencoder [145] adds regeneration as an alternative to retransmitting.In the low-SNR regime, the selective scheme raises vehicle re-identification rank-1 accuracy by more than 20% and vehicle color-classification accuracy by 10% over earlier schemes.Source: §5.4 · p. 27
  • ToDMA [146] detects token-index signatures without grants and separates device streams through channel clustering, using masked-token prediction to resolve ambiguous recovered positions.At 20 active devices its token error rate stays below the version without contextual recovery across 0–25 dB; at 10 dB the margin grows as device count increases to 80.Source: §5.4 · p. 27

5.5 Robust Token Transmission and Error Correction

Token-level recovery uses context, side information, and channel observations to conceal or repair missing and unreliable tokens, potentially reducing retransmission while retaining it for low-confidence cases. The reviewed methods range from shared language-model priors to token re-encoding and packet-loss concealment.

5.5 Robust Token Transmission and Error Correction

  • LM-based token recovery is semantic-level error concealment rather than exact bit correction, reducing reliance on retransmission while retaining retransmission for low-confidence cases.The approach is especially relevant where retransmission is costly or impractical, such as real-time or long-delay links [147, 148].Source: §5.5 · p. 27
  • Masked VQ-VAE [85] reduces overhead by transmitting discrete codebook indices and improves semantic-noise robustness with adversarial training and weight perturbation, while later token methods exploit pretrained-model context.Text-guided image TokenCom [150] uses text descriptions as side information and a masked token-generation model to restore lost tokens; shared-MLM detection [127] combines channel likelihood with token predictions and reports up to 1.77× and 1.63× reconstruction performance gains on Europarl and WikiText-103, respectively.Source: §5.5 · p. 27
  • TokCode uses soft foundational model adaptation (SFMA) to align a pre-trained generative foundation model with its re-encoded token space, preserving semantic fidelity in prompt-driven image transmission despite random erasure of 40–60% of tokens.Source: §5.5 · p. 28
  • Dual-functional masked visual token modeling [148] combines entropy modeling with feature-domain packet-loss concealment to improve neural image-codec resilience under packet losses.Source: §5.5 · p. 28

5.6 Token Communication for LM Services

Token traffic is intrinsic to distributed LM inference, spanning expert routing, KV-cache transfer, and draft-and-verify decoding. Token communication also supports multi-agent protocols and requires link-dependent choices between token identifiers and KV-cache states.

5.6 Token Communication for LM Services

  • MoE expert routing and disaggregated prefill-to-decode serving make token-associated communication payloads part of LM service infrastructure, with cache placement and network bandwidth affecting end-to-end latency.KV-cache tensors are not tokens themselves, but their dimensions and transmission patterns are determined by the token sequence.Source: §5.6 · pp. 28–29
  • In speculative decoding, a device-side draft model sends γ draft identifiers and per-step top-K logits for parallel edge verification, which returns a single token while preserving the target model's verified output distribution.Top-K truncation keeps uplink cost low at the cost of a modified acceptance rate.Source: §5.6 · p. 29
  • Distributed token inference can use tensor parallelism with over-the-air aggregation, while multi-agent systems exchange compact observation tokens, hybrid latent-text signals, or adaptively selected token and KV-cache media.The JMSRA algorithm selects communication medium and bandwidth according to channel conditions and per-agent computation budgets to reduce end-to-end collaboration latency.Source: §5.6 · p. 29
  • For edge LLM handover, cache transfer is more than 3.1 times slower than a joint design at 2 Gbps backhaul but converges with it at high rates, while cache transfer remains appropriate within a cluster and over high-rate backhaul.The comparison also reports 88 KB per token for cache exchange versus 16 bytes for text-based exchange, aggregated over four sharer models, at an accuracy gain of about 15% [158].Source: §5.6 · p. 29
  • Token identifiers and KV cache trade receiver prefill computation against bandwidth: tokens cost less bandwidth but require a full prefill, whereas cache transfer avoids prefill at higher bandwidth cost.Token identifiers are reported as preferable at low SNR because they can be repaired from the receiver's generative prior; cache transfer is favored within a cluster and over high-rate backhaul.Source: §5.6 · p. 29

5.7 Token Communication for Embodied and Agentic Intelligence

Embodied TokenCom closes the loop through the physical world: decoded tokens drive actions that alter the environment and shape later observations. This changes the cost of errors, the scope of token representation, and how transmission value is judged.

Embodied feedback and token errors

  • Unlike earlier exchanges whose feedback remains within information exchange, embodied feedback changes the physical source of the next observation.Robots, autonomous vehicles, and extended-reality devices exemplify this physical feedback loop.Source: §5.7 · pp. 29–30
  • An incorrect action can perturb the environment and compound errors along a trajectory, beyond the autoregressive propagation within a generated sequence.The surveyed robotic-grasping simulation found bit-level image reconstruction failed at lower SNR, while semantic transmission maintained task success with lower end-to-end latency.Source: §5.7 · p. 30

Action and state representation

  • A shared token vocabulary can carry perception uplinks and action downlinks, while encoding robot state and commands within one budget.The cited discrete-action scheme maps action dimensions into 256 vocabulary entries; its 7B-parameter model was trained on 970k real-world demonstrations.Source: §5.7 · p. 30

Task-aware transmission and agent coordination

  • Closed-loop transmission can score tokens by their effect on long-term return, while agent coordination can trigger transmission when internal reasoning states diverge.Wang et al. use imagined rollouts from a world-model causal digital twin; Seo et al. use predicted belief divergence to decide whether transmission is needed.Source: §5.7 · p. 31

5.8 Token-Level Protocol Design

Token-level protocols must account for semantic importance and dependence while preserving transport functions such as framing, retransmission, ordering, synchronization, and congestion control. The surveyed designs adapt these functions to generative receivers, but shared-bottleneck congestion control remains underdeveloped.

5.8 Token-Level Protocol Design

  • Deployment-oriented designs place semantic extraction at the application layer and importance-aware reliability and resource allocation in lower layers, or separate source and channel mappings through a reliability interface.The latter design reflects separate application and network providers communicating over general-purpose TCP/IP links without timely channel state at the source.Source: §5.8 · p. 31
  • Token-dependent packet values make grouping a joint optimization problem, while randomizing token positions can prevent burst losses from removing consecutive context.The cited grouping method obtains near-optimal groups at a fraction of exhaustive-search cost.Source: §5.8 · p. 31
  • Framing must signal masked token positions and account for their overhead, while control signaling can indicate semantic and conventional transmissions on one physical channel.The masked positions affect receiver detection and the per-token power assumed by the transmitter and receiver.Source: §5.8 · p. 31
  • Token retransmission can be triggered by insufficient semantic confidence, operate on tokens or groups, and combine information in the latent domain rather than the bit domain.The surveyed feature-level schemes provide feedback signals but are not tied to selective-repeat sequence numbering and window management; TONIC adds receiver-side confidence gating and erasure completion.Source: §5.8 · p. 31
  • 43.7 to 26.7 ms: communication time per image falls across SNR 6 to 9 dB, while receiver tokenization and context processing are estimated at 19.2 and 34.9 ms per image.For this workload, computation replaces retransmission only when the channel is poor enough; the surveyed studies offer no general online policy for choosing between them.Source: §5.8 · p. 32
  • Token ordering may be partial: coarse scales must arrive before finer scales, but indices within a scale can arrive in any order, unlike TCP’s total order or UDP’s lack of ordering.Neither conventional transport protocol natively exposes this tokenizer-defined partial order.Source: §5.8 · p. 32
  • Handover requires scheduling context restoration through token forwarding and prefill or KV-cache transfer, while agent coordination may trigger transmission on predicted reasoning-state misalignment.The latter trigger is based on internal-state alignment rather than channel conditions or data novelty alone.Source: §5.8 · p. 32
  • Congestion control is the least developed function: quantitative approaches remain single-link, leaving shared-bottleneck fairness and whether to design or learn protocols open.A sender may lower channel rate while shifting message completion to the receiver’s generative model; fairness depends on receiver priors not observed by the network.Source: §5.8 · p. 32

6 Challenges and Future Research Directions

The survey identifies inherited LM-driven SemCom challenges and TokenCom-specific issues. These span privacy, efficiency, compatibility, and robustness, as well as token-level interoperability, service quality, evaluation, synchronization, error handling, and scheduling.

6.1.1 Privacy and Security Risks

  • Token streams may expose intentions or sensitive content at plaintext endpoints, where conventional in-transit encryption does not provide semantic-level privacy or access control.The survey identifies semantic differential privacy, embedding-space encryption, and lightweight semantic-aware encryption as research directions.Source: §6.1.1 · pp. 32–33

6.1.2 Model Complexity and Inference Latency; 6.1.3 Cross-Model Compatibility

  • Real-time edge deployment is constrained by LM inference cost and delay, while differing or updated models require negotiated capabilities and interfaces that remain compatible across updates.Proposed efficiency directions include collaborative inference, speculative decoding, model compression, and lightweight architectures.Source: §6.1.2 · p. 33, §6.1.3 · p. 33

6.1.4 Robustness and Adversarial Resilience

  • Pretrained-model bias and adversarially misleading semantics require debiasing, adversarial training, interpretability, and receiver-side detection of reconstruction failures.Existing multimodal benchmarks test hallucination, fairness, bias, and adversarial robustness, but TokenCom receivers must raise alarms on generated reconstructions.Source: §6.1.4 · p. 33

6.2.1 Token Vocabulary Standardization; 6.2.2 Token-Level Quality of Service

  • Heterogeneous tokenizers call for shared vocabularies, translation, or per-session tokenizer agreement, while semantic-aware QoS must prioritize tokens by importance.The survey contrasts critical subject or action tokens with less important articles or conjunctions.Source: §6.2.1 · p. 33, §6.2.2 · pp. 33–34

6.2.3 Semantic-Aware Token-Level Metrics

  • TER counts every token mismatch equally, motivating a weighted metric that reflects positional dependence and semantic difference, including downstream task effects.The passage contrasts equal TER with differing named-entity F1 scores at the same 7.44% word error rate.Source: §6.2.3 · p. 34, §6.2.6 · p. 34

6.2.4 Token Synchronization and Alignment

  • Distributed TokenCom needs synchronization protocols that preserve semantic coherence across variable-length sequences and out-of-order delivery.Misalignment can degrade semantic reconstruction quality.Source: §6.2.4 · p. 34

6.2.5 Token Error Propagation

  • A corrupted token can propagate through autoregressive generation, but a generative receiver can regenerate a detected error; reliably identifying corrupted tokens remains open.Errors in tokens that condition many later tokens, such as coarse scales in a multiscale tokenizer, can degrade more of the reconstruction.Source: §6.2.5 · p. 34

6.2.6 LM Service-Aware Token Scheduling

  • LM-service scheduling must jointly optimize inference correctness, throughput, and network-resource use over time-varying channels, rather than assume constant wireless rates.Relevant traffic includes expert routing, KV-cache transfer, and draft-verification exchanges; the survey notes that existing approaches were developed for data-center links or constant-rate wireless links.Source: §6.2.6 · pp. 34–35

7 Conclusion

The survey presents TokenCom as a token-native abstraction connecting semantic representation, transmission, and distributed inference above the bit transport layer. It concludes that realizing unified, intelligent communication requires further advances across several deployment challenges.

7 Conclusion

  • LM-driven SemCom remains fragmented across modalities, models, and tasks, while TokenCom uses LM-native tokens as a unified interface for representation, transmission, and distributed inference.The survey extends communication design beyond semantic delivery to LM services and embodied and agentic intelligence.Source: §7 · p. 35
  • TokenCom treats semantic importance, reliability, and resource allocation jointly at a model-native layer above bits, which remain the transport unit.Source: §7 · p. 35
  • Practical realization still requires progress in privacy, efficiency, interoperability, robustness, token-level QoS, error handling, and service-aware scheduling.Source: §7 · p. 35
Loading 2609.10714v1…