Diffusion Language Models for Mobile Edge Agentic AI: Foundations, Applications, and Challenges

Introduction

• Edge IoT deployment is constrained by computation, energy, bandwidth, privacy, and memory, while autoregressive decoding adds sequential latency and large KV-cache demands.Long contexts can require gigabytes of RAM for KV caches, making deployment difficult on memory-limited devices.Source: §1 · pp. 2–3
• Reported literature results suggest diffusion-based LLMs approach autoregressive models on general reasoning and code benchmarks, with advantages in planning and emerging faster inference, but use non-unified evaluation settings.The cited results are a literature-level summary, not a controlled unified benchmark.Source: §1 · pp. 3–4
• DLMs iteratively denoise multiple tokens in parallel with bidirectional context, enabling flexible compute-quality trade-offs, controllable generation, and distributed inference.They can start from Gaussian noise in embedding space or an absorbing [MASK] state.Source: §1 · p. 4
• Distributed DLMs can partition denoising across device-edge-cloud tiers, use predictable fixed-length latent states, and adjust denoising depth or offloading to network, latency, and battery conditions.Keeping initial noising and final decoding on-device while offloading obfuscated intermediate states may support privacy-preserving collaboration.Source: §1 · p. 5
• Earlier diffusion and DLM surveys cover broad modeling foundations and DLM taxonomies, but do not systematically connect DLMs to mobile-edge deployment, communication cooperation, IoT or wireless applications, and agent evaluation.The introduction distinguishes this survey by centering DLMs, rather than continuous diffusion models, and taking a system-level view of edge constraints.Source: §1 · p. 6
• This survey links DLM properties to mobile-edge requirements and reviews efficient architectures, training, inference and compression, deployment and communication, IoT and wireless applications, evaluation, and open challenges.Its open challenges include long-context state management, split inference, trustworthy execution, multimodal grounding, and reproducible benchmarking.Source: §1 · pp. 6–7

2.1.1 Autoregressive Models

• Autoregressive models generate sequences through left-to-right next-token prediction, factorizing joint sequence probability so each token depends only on preceding tokens (Gruver et al., 2023).This factorization supports data compression, pattern extrapolation, and in-context learning.Source: §2.1.1 · p. 9
• GPT-3 and GPT-4 exemplify scaling through larger model capacity and training data (Touvron et al., 2023), while their zero- and few-shot generalization makes them cloud-based upper-bound references for edge agents (Gruver et al., 2023).They perform complex cognitive tasks without downstream fine-tuning.Source: §2.1.1 · pp. 9–10
• Open-source Llama models use optimized autoregressive Transformers such as GQA, and Llama 2 adds RLHF alignment, making them foundations for on-device and lightweight edge research (Touvron et al., 2023).Llama 2 combines large-scale self-supervised pre-training with reinforcement learning from human feedback.Source: §2.1.1 · p. 10
• Time-LLM reprograms frozen LLaMA/GPT-2 for forecasting and spatiotemporal reasoning, while coarse-to-fine next-scale visual autoregression improves raster-scan efficiency and surpasses DiTs in fidelity and scalability (Jin et al., 2024; Tian et al., 2024).Time-LLM maps time-series data into textual prototypes through patch reprogramming; visual autoregression uses multi-scale parallel prediction.Source: §2.1.1 · p. 10

2.1.2 Masked Language Models for Edge Intelligent Perception

• Masked language models reconstruct masked tokens from surrounding left and right contexts, providing compact semantic perception for robust edge-agent decisions under computation, memory, and bandwidth limits (Devlin et al., 2019; Deng et al., 2020).Unlike autoregressive models, MLMs exploit bidirectional context.Source: §2.1.2 · p. 10
• BERT and RoBERTa show that masked pre-training yields transferable semantic embeddings, while DeBERTaV3 uses RTD and GDES to improve pre-training efficiency and NLU (Devlin et al., 2019; Liu et al., 2019; He et al., 2023a).GDES reduces conflicts between discriminative and generative objectives.Source: §2.1.2 · p. 10
• BERT/RoBERTa embeddings support BERTopic’s latent-topic discovery, while masked completion can impute missing sensor readings for environmental awareness and makes MLMs a precursor to DLM reconstruction (Grootendorst, 2022).BERTopic combines density-based clustering with class-based TF-IDF.Source: §2.1.2 · pp. 10–11

2.2 Diffusion Models

• Diffusion models are generative latent-variable models that corrupt data toward a simple prior through a parameterized Markov chain and synthesize it by learning reverse denoising; they may use continuous or discrete state spaces.This framework is inspired by non-equilibrium thermodynamics Lipman et al. (2023).Source: §2.2 · p. 11

2.2.1 Continuous Diffusion Models

• Continuous diffusion injects Gaussian noise and reverses it through an SDE; score estimation enables denoising, while LDMs use autoencoder latents and Flow Matching regresses a conditional-path vector field Song et al. (2021); Karras et al. (2022); Rombach et al. (2022); Lipman et al. (2023).LDMs reduce score-estimation cost in high-dimensional spaces and support lighter edge deployment; Flow Matching accelerates training and sampling.Source: §2.2.1 · p. 11

2.2.2 Discrete Diffusion Models

• Discrete D3PMs replace Gaussian noising with categorical Markov transitions, including absorbing states such as [MASK], connecting hierarchical reverse prediction to conventional masked language modeling Austin et al. (2021).In the described masking process, the reverse step predicts masked tokens hierarchically and layer by layer.Source: §2.2.2 · p. 11

2.3 Diffusion Language Models

• DLMs iteratively reconstruct corrupted sequences rather than decode left to right, enabling parallel token refinement, bidirectional context, controllability, and flexible quality-latency adaptation for edge agents.These properties are presented as useful under resource constraints.Source: §2.3 · p. 12

2.3.1 Generation Pipeline of Diffusion Language Models

• DLM generation pipelines include continuous embeddings, discrete state-space transitions, and masked diffusion; masked-token training can upper-bound the induced model’s negative log-likelihood Nie et al. (2025).Continuous embeddings enable gradient-guided control but may incur rounding errors, embedding sensitivity, and slow sampling Li et al. (2022); discrete state-space diffusion avoids continuous-space quantization errors Austin et al. (2021). A complementary variational derivation expresses masked diffusion training as a schedule-weighted average of masked language modeling losses Sahoo et al. (2024). DiffusionGemma-26B-A4B-it combines masked diffusion decoding with a mixture-of-experts backbone as an open-weight inference reference Google (2026a,b).Source: §2.3.1 · pp. 12–13

2.3.2 Comparison with Autoregressive LLMs

• DLMs use bidirectional refinement, adjustable reverse-sampling steps, and inference-time conditioning, while autoregressive models decode causally and can struggle with global planning, infilling, and zero-shot constraints.LLaDA is reported to improve global contextual reasoning and mitigate the Reversal Curse Nie et al. (2025); at comparable scale, SEDD outperforms GPT-2 with far fewer network evaluations Lou et al. (2024); Feng et al. (2025a). Diffusion-LM uses gradient-based guidance and SEDD uses classifier-free guidance for controllable generation without task-specific retraining Li et al. (2022); Lou et al. (2024).Source: §2.3.2 · p. 14

2.3.3 Multimodal Large Diffusion Language Models

• Representative multimodal DLMs extend diffusion to visual understanding, reasoning, and generation through vision-text architectures and, in MMaDA, shared discrete diffusion Yang et al. (2025a).Examples include Dimple’s autoregressive-then-diffusion training and Confident Decoding Yu et al. (2025b), LaViDa’s complementary masking and Prefix-DLM caching Li et al. (2025e), and LLaDA-V’s visual instruction tuning and bidirectional attention You et al. (2026). Edge deployment still depends on efficient visual tokenization, denoising acceleration, cache optimization, quantization, and adaptive offloading.Source: §2.3.3 · pp. 14–15

2.4.1 Elastic Compute Adaptation and Deployment

• Agents can adapt diffusion steps to real-time energy thresholds, using deeper denoising when resources allow and early stopping under battery constraints (Lou et al., 2024).Early stopping trades refinement for ultra-low latency while retaining core semantics.Source: §2.4.1 · p. 15

2.4.2 Alleviating Memory Bandwidth Bottlenecks via Parallel Decoding

• On mobile edge devices, DLMs such as MDLM and LLaDA predict multiple tokens in parallel during reverse diffusion, improving edge-accelerator utilization and reducing end-to-end perception-to-action latency for intelligent agents (Sahoo et al., 2024; Nie et al., 2025).Source: §2.4.2 · p. 15

2.4.3 Deterministic and Safe On-Device Execution

• For physical actions through local operating systems or actuators, controllable DLM generation can enforce predefined formats and security boundaries, improving on-device robustness and safety (Li et al., 2022; Lou et al., 2024).Source: §2.4.3 · pp. 15–16

2.5 Lessons Learned

• Autoregressive LLMs remain preferable for open-ended and long-form generation, but sequential decoding and KV-cache dependence create edge latency and memory bottlenecks (Li et al. (2025a); Peng et al. (2025)).The cited passage identifies these bottlenecks as limitations at the edge.Source: §2.5 · p. 16
• DLMs better support infilling, structured generation, global planning, and constraint-aware action synthesis, but practical gains depend on stable sampling, cache compatibility, and hardware-aware parallel decoding (Nie et al. (2025); Wu et al. (2025)).These implementation conditions qualify DLMs’ edge advantages.Source: §2.5 · p. 16
• Edge systems should treat autoregressive models and DLMs as complementary backbones, selecting between them by task structure, latency budget, and ecosystem maturity.The passage frames backbone choice as dependent on these three factors.Source: §2.5 · p. 16

3.1 Efficient Model Architectures / 3.1.1 Lightweight Diffusion Transformers

• DiT-based lightweight designs, token merging and pruning, and low-bit quantization target attention and memory costs, while SnapGen++ demonstrates 0.4B parameters and 1.8-second generation latency.U-ViT uses uniform tokens and long skip connections in place of U-Net downsampling and upsampling modules. DiT and SnapGen++ are discussed as routes to resource-efficient diffusion, with SnapGen++ combining architectural optimization and efficient training.Source: §3.1 · p. 17, §3.1.1 · p. 17

3.1.2 Block Diffusion and Hybrid Autoregressive Diffusion Models

• Block and hybrid designs combine diffusion parallelism with autoregressive length flexibility and cache reuse, including adaptive token refinement based on semantic maturity (Wu et al., 2025; Arriola et al., 2025; Wu et al., 2026).Block Diffusion supports flexible-length generation with KV caching and adaptive parallel sampling. Evo assigns tokens maturity values and uses local autoregressive-style refinement for less mature tokens and global diffusion-style planning for more mature ones.Source: §3.1.2 · pp. 17, 19

3.1.3 Sparse and MoE Diffusion Models

• Sparse MoE models increase total capacity while activating only selected experts, with timestep-aware routing intended to improve complex generation without raising single-pass computation (Park et al., 2025; Fei et al., 2024).DiT-MoE exhibits different expert preferences across early and late denoising, and sparse activation offers a configurable trade-off between model capacity and memory requirements.Source: §3.1.3 · p. 19

3.2 Efficient Training Strategies / 3.2.1 Autoregressive-to-Diffusion Adaptation and Joint Training / 3.2.2 PEFT Mechanisms

• Resource-efficient training combines pretrained autoregressive-to-diffusion adaptation with PEFT, reusing foundation weights and limiting updates through methods such as LoRA and DiffFit (Gong et al., 2025; Hu et al., 2022; Xie et al., 2023).LoRA freezes pretrained weights and adds trainable low-rank matrices; DiffFit fine-tunes bias terms and lightweight spatial scaling factors to reduce storage and gradient costs.Source: §3.2 · p. 19, §3.2.1 · p. 20, §3.2.2 · p. 20

3.2.3 Federated Diffusion Training Strategies

• Federated diffusion complements local PEFT by coordinating privacy-sensitive distributed training, but remains challenged by non-IID data, communication overhead, and heterogeneous edge resources (Sun et al., 2024; Li et al., 2025b; Mendieta et al., 2025).One-shot federated diffusion with differential privacy and Fourier magnitude filtering targets synchronization latency and privacy risks; on-demand quantization adapts compression and wireless transmission to battery and bandwidth limits.Source: §3.2.3 · pp. 20–21

3.3 Efficient Inference Methods / 3.3.1 Parallel Decoding / 3.3.3 Emerging Speculative Decoding Directions

• Inference acceleration spans parallel token updates and speculative draft–verify approaches, which may offload lightweight drafting to IoT devices and costly verification to edge servers (Wu et al., 2023; Chang et al., 2022; Leviathan et al., 2023).MaskGIT uses a bidirectional Transformer and dynamic mask scheduling, while speculative decoding proposes candidates for parallel target-model verification; the latter is presented as a potential diffusion extension.Source: §3.3 · p. 21, §3.3.1 · p. 21, §3.3.3 · p. 22

3.3.2 Few-Step Diffusion

• Few-step methods shorten denoising trajectories: DDIM enables 10x to 50x wall-clock acceleration, and progressive distillation reduces sampling from up to 8192 steps to as few as 4 (Song et al., 2022; Salimans and Ho, 2022).The section distinguishes fewer denoiser evaluations from parallel decoding, and also describes solver- and consistency-based approaches, including high-fidelity latent consistency generation in 2 to 4 steps or even one step (Lu et al., 2022; Song et al., 2023; Luo et al., 2023).Source: §3.3.2 · pp. 21–22

3.4 Model Compression / 3.5 Lessons Learned

• Compression strategies include timestep-aware quantization, distillation, and pruning, but edge DLMs must balance denoising depth, memory footprint, and quality while accounting for possible semantic drift and hardware complexity (Li et al., 2023a; He et al., 2023b; Salimans and Ho, 2022; Meng et al., 2023; Yang et al., 2023b; Fang et al., 2023).The section recommends prioritizing memory-bandwidth reduction, cache reuse, adaptive steps, and quality reports under fixed latency, energy, and device settings.Source: §3.4 · pp. 22–23, §3.5 · p. 23

4.1 Device-Only and Edge-Assisted DLM Agent Inference

• Device-only, edge-assisted, and cloud–edge collaboration offer complementary ways to distribute DLM inference while balancing local resource limits, latency, and privacy.Device-only designs suggest lightweight backbones, cache-aware loading, low-bit execution, fewer denoising steps, and early exits; split inference can keep private prompts and local states on-device, while cloud tiers handle heavier generation. Examples include SnapFusion Li et al. (2023b), EdgeDiT Kodavanti et al. (2026), and DiffusionX Wei et al. (2025).Source: §4.1 · pp. 23–24

4.2 Distributed and Pipeline DLM Agent Inference

• Distributed and pipeline inference can partition DLM execution across devices and edge nodes by token blocks, denoising rounds, reasoning states, or action outputs.Collaborative denoising can reduce device energy use and avoid sharing raw prompts, while adaptive step allocation responds to bandwidth and device capacity; asynchronous activation reuse can hide communication latency. Reasoning-oriented systems may also parallelize trajectory exploration before device-side response synthesis.Source: §4.2 · p. 25

4.3 Agent-State-Aware Resource Management and Adaptive Serving

• Because edge conditions and agent states vary over time, DLM serving must adapt offloading, scheduling, caching, and refinement allocation to balance latency, energy, privacy, and quality.Adaptation can account for wireless signal-to-noise ratio, device capacity, and request demand; persistent contexts and partial reasoning trajectories may need migration across nodes. Evaluation should include interaction success, state reuse, and safe action completion, not only communication cost.Source: §4.3 · pp. 25–26

4.4 Implications for Edge-Native DLM Agent Deployment

• Edge DLM systems must be redesigned for discrete or hybrid token dynamics, masked-token refinement, KV-cache overhead, long-context reasoning, and secure multi-tier state management.Existing visual-diffusion deployment schemes cannot simply be reused; the section identifies DLM-specific mobile-agent frameworks as a research direction.Source: §4.4 · pp. 26–27

4.5 Lessons Learned

• A practical edge-DLM design selects device-only, edge-assisted, or cloud–edge execution by privacy, latency, energy, and bandwidth, keeping private prompts and final actions on-device when possible.Heavy intermediate refinement can be offloaded, with partitioning adjusted to network conditions, task risk, token uncertainty, and interaction latency; visual-generation schemes do not directly address discrete tokens, long-context states, reasoning traces, or tool calls.Source: §4.5 · p. 27

5.1 Communication-Aware DLM Agent Inference

• Communication-aware inference uses model partitioning and collaboration to manage denoising, split execution, and interaction latency under limited device resources and stochastic wireless conditions.It aims to reduce centralized inference’s communication latency and bandwidth use while preserving agent goals such as time-to-first-action, privacy exposure, and action reliability.Source: §5.1 · p. 27, §5 · p. 27

5.1.1 DLM-Oriented Model Partitioning

• Partitioning adapts computation to heterogeneous hardware and wireless states through fine-grained subgraphs, bandwidth-aware width scaling, patch-based diffusion, split learning, and multi-exit denoising [Huo et al. (2026); Li et al. (2025c); Li et al. (2026); Tian et al. (2025a)].DLM partition boundaries should account for token blocks, denoising rounds, reasoning states, and tool or action boundaries, not only model layers.Source: §5.1.1 · pp. 27–28

5.1.2 Collaborative DLM Agent Generation

• Collaborative generation coordinates cross-tier execution and exchanges, reducing wireless traffic through cloud semantic planning with edge refinement, on-device previews before cloud refinement, and adaptive offloading [Yan et al. (2024); Wei et al. (2025); Tian et al. (2025a)].Multi-stage relay diffusion supports task migration across heterogeneous nodes, while DLM agents should enable draft-and-revise interaction without exposing permission-critical outputs before verification [Li et al. (2026)].Source: §5.1.2 · p. 28

5.2 Agent-Aware Computation–Communication Co-Optimization

• DLM inference is memory-bound as repeated accesses across denoising rounds raise bandwidth overhead, creating joint latency-memory constraints that motivate adaptive schedules and energy-aware offloading [Peng et al. (2025)].A closed loop links runtime observation, policy control, serving decisions, tiered execution, and feedback to adapt denoising depth, early exit, routing, and cache reuse to confidence, resources, privacy risk, and outcomes.Source: §5.2 · pp. 28–29, Figure 4 · p. 29

5.2.1 Adaptive Denoising Schedules

• ETE reduces decoding rounds without compromising fidelity by exploring high-information tokens to trigger confidence cascades, while semantic-aware truncation adapts refinement to task and network conditions [Fu et al. (2025)].EoTP and learnable filtering halt work on padding or redundant iterations; early skipping bypasses expensive layers for low-importance tokens, and communication-aware guidance supports robustness to packet loss or bandwidth fluctuations [Bao et al. (2025); Zhu et al. (2026); He et al. (2025)].Source: §5.2.1 · pp. 29–30

5.2.2 Energy-Aware DLM Agent Inference

• Energy-aware serving continuously repositions split points to balance device lifetime, privacy exposure, interaction latency, and generation quality through local noising and distributed execution [Ai et al. (2026); Liu et al. (2024a)].DRL can optimize split points against battery, transmission, and privacy constraints [Zhou et al. (2022)], while multi-node routing and MoE-style collaborative decoding distribute token refinement across edge experts [Zeng et al. (2025b); Qin et al. (2025); Jin et al. (2026); Gao et al. (2026)].Source: §5.2.2 · pp. 30–32

5.3 Lessons Learned

• Adaptive serving should jointly manage offloading, denoising depth, latency, energy, and privacy because fluctuating channels, mobility, contention, packet loss, and heterogeneous hardware make fixed splits fragile.Future systems should optimize time-to-first-action, selectively refine uncertain or risky tokens, and retain permission-critical steps locally.Source: §5.3 · p. 32

6.1 Continuous DMs as Precedents for Wireless Intelligence

• Continuous DMs provide precedents in wireless and IoT security, aerial operations, and network resource management.Examples include scarce attack-sample synthesis for intrusion detection (S´anchez-Carballo et al., 2026), load forecasting, aerial-image restoration, and UAV trajectory generation (Pan et al., 2025), and O-RAN and network-slicing resource management (Yan et al., 2026). A survey of wireless GDMs provides broader context (Fan et al., 2026).Source: §6.1 · p. 32

6.2 DLMs for Semantic Agentic Intelligence

• DLMs extend diffusion from continuous signals and resource variables to discrete semantic and executable representations for edge-agent tasks.Potential uses include instruction repair, action-token synthesis, protocol configuration, executable-plan generation, and agent decision-making; a unified token-denoising interface can also support multimodal observations, trajectories, actions, and semantic codewords (Li et al., 2025f; Reyes-Ram´ırez et al., 2024).Source: §6.2 · pp. 32–33
• Robotics studies use DLMs to generate revisable, semantically grounded action sequences in simulation and real-robot tasks.LLaDA-VLA uses localized action-token classification and hierarchical decoding (Wen et al., 2025b); other approaches refine tokenized action chunks (Liang et al., 2025c), jointly generate visual subgoals, reasoning, and actions (Wen et al., 2025a), or synchronously denoise future observations and action chunks in a shared token space (Liu et al., 2026a; Chen et al., 2026).Source: §6.2 · p. 34
• Vehicular DLM applications include trajectory planning and safety correction as well as password-strength assessment for transportation systems.ReflectDrive regenerates unsafe trajectory tokens through inpainting (Li et al., 2025d); ReflectDrive-2 adds decision–draft–reflect editing, reinforcement learning, and on-device efficiency techniques (Wang et al., 2026); DiffGuess models password modification as iterative diffusion (Xiong et al., 2025).Source: §6.2 · p. 34
• A preliminary aircraft-design study illustrates DLM-related discrete diffusion for topology choices and variable-dimensional structures, but not real-time UAV networking.It combines a Riemannian DLM with masked diffusion to sample aircraft topologies and continuous parameters under simulation-based inference (Ghiglino et al., 2026).Source: §6.2 · p. 34

6.3 Lessons Learned

• Because edge DLM applications remain early and often indirect, practical agents should complement continuous DMs and integrate safety, verification, feedback, and real-device validation.Potential DLM uses include command repair, structured API and tool generation, semantic coordination, protocol configuration, and intent translation (Li et al., 2025f; Huh et al., 2026; Habib et al., 2026).Source: §6.3 · p. 35

7.1 Structure-Aware Evaluation with Emphasis on DLMs

• DLM evaluation should couple model quality with denoising efficiency, schedule sensitivity, parallel decoding gain, early-exit quality, cache compatibility, and quality-latency-energy trade-offs.LLaDA and DiffuLLaMA show language and reasoning potential, but repeated denoising, remasking instability, limited cache compatibility, and inconsistent latency protocols can weaken efficiency (Nie et al., 2025; Gong et al., 2025; Peng et al., 2025).Source: §7.1 · pp. 35–36
• DiffusionGemma exhibits task- and granularity-dependent left-to-right bias, large commit batches, remasking, and entropy-based commitment, motivating process-level edge reporting.Reports should include commit-order statistics, same-call commit ratios, entropy or confidence signals, denoising-step budgets, and task-regime differences (Asaria et al., 2026).Source: §7.1 · p. 36

7.2 Metrics and Benchmarks

• Metrics should address reasoning, generation stability, interaction success, and efficiency by grouping capability, agent/mobile, and edge-constraint benchmarks around deployment questions.Reports should specify denoising steps, mask schedule, output and prompt lengths, batch size, cache strategy, quantization, runtime, accelerator, and power measurement method; these settings affect fair comparisons and can hide sample-level instability (Fang et al., 2026).Source: §7.2 · p. 36

7.3 Evaluation Framework

• Table 8 presents a five-layer checklist—backbone capability, generation dynamics, mobile interaction, edge deployment, and safety and reliability—for comparing autoregressive models, general DMs, and DLMs.CDLM and R2-dLLM studies motivate joint reporting of accuracy, denoising trajectory, caching behavior, latency, and hardware efficiency (Kim et al., 2026; Du et al., 2026).Source: §7.3 · pp. 36–37, Table 8 · p. 37

7.4 Lessons Learned

• High benchmark scores alone do not establish deployability; evaluations should jointly measure generation dynamics, interaction reliability, system efficiency, and safety on real hardware.Future benchmarks should standardize fixed-latency and fixed-energy settings, time-to-first-action, invalid-action rates, wireless dynamics, and reproducible hardware measurements (Peng et al., 2025; Fang et al., 2026; Rawles et al., 2025; Janapa Reddi et al., 2022).Source: §7.4 · p. 37

8.1 Edge-Native DLM Architectures

• Edge-native DLMs must reduce repeated memory access that can offset parallel token updates, jointly optimizing denoising depth, update ratio, masks, cache compatibility, and hardware-friendly operators [Peng et al. (2025); Wu et al. (2025)].Lightweight backbones, block-wise diffusion, sparse routing, early exit, and anytime generation could adapt quality to real-time resource budgets [Arriola et al. (2025); Zhu et al. (2026)].Source: §8.1 · p. 38

8.2 Long-Context Memory and Agent State Management

• Long-context agents need to avoid refining every token at each step, instead identifying stable tokens, selectively remasking uncertain regions, compressing history, and caching reusable states [Li et al. (2025a); Peng et al. (2025)].Full-sequence refinement, repeated remasking, and multi-step updates increase memory-bandwidth pressure; cache- and confidence-aware decoding offer starting points for efficiency [Wu et al. (2025); Du et al. (2026)].Source: §8.2 · p. 38

8.3 Communication-Aware Split and Distributed DLM Inference

• Distributed DLMs should allocate denoising steps, token blocks, or experts across devices, edge servers, and cloud nodes based on bandwidth, channel quality, latency, and privacy [Du et al. (2024); Zheng (2025)].One proposed split keeps private prompt encoding and final action decoding on-device while offloading intermediate refinement or sparse experts to nearby edge nodes [Qin et al. (2025); Jin et al. (2026); Liu et al. (2024a)].Source: §8.3 · p. 38

8.4 Privacy, Security, and Trustworthy Edge DLM Agents

• Split inference may obscure raw prompts through noising and intermediate states, but token distributions, latent states, reasoning traces, and tool-call intentions can still leak sensitive information [Allmendinger et al. (2026); Dockhorn et al. (2023)].Future pipelines should combine privacy-preserving noising, secure split inference, trusted execution, and formal leakage analysis; iterative refinement also enables step-wise safety checks, constraint enforcement, and rollback before actions [Deng et al. (2024); Rawles et al. (2025)].Source: §8.4 · p. 39

8.5 Multimodal and Physical-World DLM Agents

• Multimodal DLMs can unify text reasoning, multimodal understanding, and text-to-image generation, but real-time mobile and IoT deployment remains underexplored [Yang et al. (2025a); Tian et al. (2025b)].A proposed direction represents instructions, observations, sensor readings, and actions as structured tokens refined through shared denoising, with real-device validation, closed-loop evaluation, latency control, robustness, and safety guarantees still required [Huh et al. (2026); Habib et al. (2026)].Source: §8.5 · p. 39

8.6 Standardized Evaluation and Reproducibility

• Benchmarks should jointly assess DLM generation dynamics and system constraints, reporting denoising steps, mask schedules, remasking stability, parallel decoding gain, cache behavior, energy, communication, and closed-loop success [Rawles et al. (2025); Janapa Reddi et al. (2022); Peng et al. (2025)].They should also measure time-to-first-action, invalid actions, privacy leakage risk, and safety failures; reproducibility needs fixed-latency and fixed-energy protocols plus hardware, accelerator, quantization, prompt/output lengths, batch size, and power details [Fang et al. (2026); Janapa Reddi et al. (2022)].Source: §8.6 · p. 39

9 Conclusion

• DLMs offer mobile edge agents parallel token refinement, adaptive denoising, constrained generation, and privacy-aware split execution.The survey also reviews their foundations, resource-efficient techniques, deployment architectures, communication-aware serving, IoT/wireless applications, and evaluation criteria.Source: §9 · p. 39
• DLMs should complement rather than replace autoregressive LLMs, suiting structured, constraint-aware, and latency-adaptive tasks.The survey connects diffusion language modeling, edge computing, wireless networking, and trustworthy agentic AI.Source: §9 · p. 40
• DLMs still face challenges in long-context state management, hardware-efficient inference, distributed collaboration, safety, multimodal grounding, and reproducibility.Source: §9 · p. 40

• The work was supported by National Natural Science Foundation of China Grants 62571529 and U25A20388.Source: Declarations · p. 40
• C.L. led conceptualization, investigation, data curation, visualization, and original drafting, while M.M., D.N., and W.N. shared conceptualization, review, and supervision.M.M. also handled project administration and funding acquisition; all authors approved the final manuscript.Source: Declarations · p. 40
• The authors declare no competing interests relevant to the article, and no datasets were generated or analysed during the study.Source: Declarations · p. 40
• Ethics approval, consent to participate, consent for publication, materials availability, and code availability are not applicable.Source: Declarations · p. 40