1
SkillComm: Skill-Driven Semantic Communication via Workflow-Aware Token Prioritization
arXiv:2607.11972v1 [cs.NI] 13 Jul 2026
Ziyang Meng, Student Member, IEEE and Lu Lu, Member, IEEE
Abstract—As wireless visual intelligence evolves from isolated task inference to ordered skill workflows, the core communication bottleneck shifts from transmitting a single semantic representation to coordinating reusable skill states under channel constraints. Existing DeepJSCC and prompt-guided visual transmitters usually treat each task as an independent fulltoken transmission, with limited reuse of execution memory across semantic workflows. This is inefficient for workflows such as Detect→Segment→Keypoint, where later stages mainly require state-relevant semantic updates rather than repeated full-grid transmission. This letter proposes SkillComm, a skilldriven semantic communication (SemCom) framework that utilizes reusable skill states as communication context to jointly guide workflow-aware token prioritization and memory-assisted token-grid reassembly. A shared Skill-Book first synchronizes the transmitter and receiver by mapping a high-level visual intent into an executable skill sequence. Conditioned on this synchronized workflow, workflow-aware adaptive token selection exploits cross-step execution memory to prioritize state-active tokens before joint source-channel coding (JSCC) transmission. Finally, the receiver reconstructs a task-ready token grid by merging the decoded active tokens with its local historical memory. Evaluations on the MS COCO 2017 validation dataset across the Detect→Segment→Keypoint workflow show that workflowaware token prioritization enables SkillComm to achieve a 51.2% token transmission cost reduction, while retaining 99.4% upperbound-normalized average precision in the high SNR regime. These results provide quantitative evidence that reusable skill states can drive workflow-aware token prioritization, enabling selective semantic innovation delivery for future agentic and embodied visual intelligence.
I. I NTRODUCTION As wireless visual intelligence evolves from isolated task inference to ordered skill workflows, semantic communication (SemCom) must move beyond transmitting compact representations toward coordinating reusable skill states under channel constraints. Existing SemCom aims to transmit taskrelevant information rather than reconstructing every source symbol [1]–[3]. Deep joint source-channel coding (DeepJSCC) maps visual sources directly into channel symbols and achieves graceful degradation over noisy channels [4], while recent task-adaptive SemCom introduces prompts, conditional rate-distortion objectives, or tokenized representations to guide semantic coding [5]–[7]. Although these methods advance semantic feature coding, prompt conditioning, and tokenized transmission, they mainly optimize each transmission instance independently and do not explicitly exploit reusable execution states across ordered visual workflows. The authors are with the Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing, 100094, China (e-mails: {mengziyang24, lulu}@csu.ac.cn).
A high-level visual intent may unfold as an ordered workflow or modular visual program [8], e.g., detecting a person, segmenting the detected instance, and estimating its keypoints [9]. In such workflows, later stages inherit execution states such as bounding boxes, masks, or intermediate token memory, and therefore require mainly state-relevant semantic updates rather than another full-token transmission [10]. Repeatedly sending the full token grid wastes wireless resources on redundant background and already available context, while compressing a compound instruction into a single prompt may cause task interference and attention drift. Therefore, the key gap is a workflow-level SemCom mechanism that uses skill states to decide which tokens should be transmitted, which context can be reused, and how task-ready representations should be reassembled at the receiver. To address this challenge, this letter proposes SkillComm, a skill-driven SemCom framework for workflow-aware token prioritization. To the best of our knowledge, this is the first work that introduces skill-driven workflow control into visual SemCom. As summarized in Tab. I, SkillComm organically couples skill-based workflow control with workflow-aware token prioritization, memory-assisted reassembly, and visual JSCC. SkillComm decomposes a high-level visual intent into an executable skill workflow synchronized by a shared SkillBook. Unlike prompt-conditioned SemCom that mainly treats task descriptions as conditioning signals, SkillComm uses reusable skill states as communication context to prioritize state-active tokens before JSCC transmission and to restore task-ready token grids through memory-assisted reassembly. By advancing visual SemCom from prompt-conditioned full-token delivery to workflow-aware semantic information delivery, SkillComm makes the following contributions. Skill-Driven SemCom Architecture. We propose SkillComm, a visual SemCom framework for sequential workflows. A shared Skill-Compiler and Skill-Book decompose intents into executable skills, enabling coordinated workflow execution across the wireless link. • Workflow-Aware Token Prioritization. A memorydriven selection mechanism exploits cross-step states to transmit only active semantic innovations. The receiver then reconstructs task-ready token grids by merging decoded features with cached context. • Experimental Validation. On the MS COCO 2017 workflow, SkillComm achieves a 51.2% token transmission cost reduction and retains 99.4% upper-boundnormalized AP at 15 dB Rayleigh fading. •
2
TABLE I: Comparison with representative approaches.
Task 1
Workflow Multi-Task Token Visual Compiler Control Prioritization JSCC DeepJSCC [4] 2019 × — × ✓ TokenCom [5] 2025 × — × × SkillComa [11] 2026 △ — × × TASC-f [6] 2026 × Prompt × ✓ PGMT-SC [7] 2026 △ Prompt × ✓ SkillComm Ours ✓ Skill ✓ ✓ Method
Task 2
Task 3
...
Year
High-Level Intent q
“Capture the person's body mechanics”
I Skill-Guided Backbone Encoder (e.g Swin / ViT)
Shared Skill Compiler skill () Step 1 k1
Xt Token Selection Γθ ()
B = {σ1 , σ2 , . . . , σM },
(1)
where each reusable visual skill is represented as σm = (pm , em , Hm ).
(2) de
Here, pm is the textual skill description, em ∈ R is the compact skill embedding, and Hm (·) denotes the associated task head, e.g., object detection, instance segmentation, or keypoint estimation. The Skill-Book can be pre-loaded from a library or optimized offline before transmission. In this work, em is instantiated as a single feature vector, while it can be extended to multiple embedding tokens for more complex composite skills. Given a high-level instruction q, the skill compiler maps it to an ordered workflow: W(q) = Cskill (q, B) = [k1 , k2 , . . . , kT ] ,
(3)
where kt ∈ {1, . . . , M } denotes the skill index executed at stage t. We define et ≜ ekt and Ht ≜ Hkt as the scheduled skill embedding and task head. Since both B and W(q) are synchronized, the receiver activates the same skill at each stage without repeatedly receiving natural-language prompts. B. Token Selection Criteria and JSCC Encoding At the t-th skill stage, the transmitter extracts a skillconditioned token canvas: Xt = Eϕ (I, et ) = [xt,1 , . . . , xt,N ] ∈ RN ×d ,
(4)
where N and d denote the spatial token count and token dimension. To avoid repeated full-token transmission, SkillComm selects a state-active token subset It+ ⊆ {1, . . . , N } according to It+ = Γθ (Xt , MTx t−1 , et , γt ),
(5)
...
Step T kT
Xˆ t Token Reassembly ()
Shared Skill-Book
= { 1 , 2 ,..., M }，
X t+
A. Workflow Decomposition using Skill Compiler SkillComm assumes that the transmitter and receiver share a synchronized compact Skill-Book,
Step 2 k2
Sequential Workflow (q ) (e.g., Detect→Segment→Keypoint)
△ denotes partial support. a SkillCom studies text-only LLM-based SemCom, whereas SkillComm targets visual workflow-level SemCom with token prioritization.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION Fig. 1 presents the proposed SkillComm architecture. We consider a wireless visual semantic communication system where a transmitter observes an image I ∈ RH×W ×3 and has a high-level instruction q ∈ Q, e.g., “Capture the person’s body mechanics.” The receiver is expected to execute an ordered visual skill workflow, such as Detect→Segment→Keypoint, under wireless channel constraints.
oˆ t
Task Heads (e.g., Mask R-CNN / Keypoint R-CNN)
TX Memory
JSCC Encoder
M tTX −1
st+
m
= ( pm , em , H m )
Wireless Channel
Xˆ t+
yt+
JSCC Decoder
Rx Memory
M tRX −1
Transmitter
Receiver
Fig. 1: System model of SkillComm. A high-level instruction is decomposed into a synchronized sequential workflow via skill compiler. Reusable skill states guide workflow-aware token prioritization for JSCC transmission, while memoryassisted reassembly reconstructs task-ready token grids at the receiver. N ×d where MTx is the transmitter-side execution memt−1 ∈ R ory and γt is the normalized channel SNR. For the initial stage, I1+ = {1, . . . , N } is used to initialize the workflow memory. + The selected tokens are gathered as X+ t = Gather(Xt , It ) and mapped to channel symbols by a JSCC encoder: + s+ t = Φenc (Xt ).
(6)
The skill embedding et conditions the backbone feature extraction, while JSCC operates only on the selected active tokens. = IndexEnc(It+ ) and The spatial indices are encoded as bidx t transmitted through a reliable control channel. The wireless channel is modeled as yt+ = ht s+ t + nt ,
(7)
where ht is the complex channel coefficient and nt ∼ CN (0, γt2 I) represents additive white Gaussian noise. The transmitter then updates its local memory as MTx t = Xt . C. JSCC Decoding and Token Reassembly Objective Upon receiving the wireless signals, the receiver decodes the spatial indices Ît+ = IndexDec(bidx t ) and applies the JSCC decoder to recover the high-priority active tokens: + X̂+ t = Φdec (yt ).
(8)
A full token canvas X̂t ∈ RN ×d is then reconstructed by a reassembly operator R(·), which scatters the decoded active tokens back to their original grid coordinates and replenishes the unselected slots utilizing the synchronized historical receiver memory MRx t−1 : + Rx X̂t = R(X̂+ t , Ît , Mt−1 ).
(9)
Due to the stochastic nature of the noisy wireless channel, Rx MTx t−1 ̸= Mt−1 generally holds true. Therefore, SkillComm operates robustly without requiring strictly identical memory
3
Transmitter Token Selection
TX Memory Update
TX memory
Cosine skill relevance
Rate MLP
Score MLP
Top-K selection
cos ( xt ,i , Weet )
M tTX −1
Token-level feature interaction
Current tokens Xt
r
t+
at, i
Skill embedding et
Cosine skill difference 1 − cos ( xt ,i , mtTX −1,i )
where ⊙ denotes element-wise multiplication. The definitive routing score synergizes this logit with explicit current-skill relevance and temporal difference priors:
Gather
Xt+
Wireless Channel Decoded selected tokens
Scatter & Reassembly
RX memory M tRX −1
x , i , xt ,i = + m tRx−1,i , i t ,
t+ X
+ t , t ( i )
+ t
st,i = at,i + α cos(xt,i , We et ) + β 1 − cos(xt,i , mTx t−1,i ) , (13) Backbone Decoder
Task Heads
...
Receiver Token Reassembly
RX Memory Update
aware token gate. For the i-th token, a lightweight Multi-Layer Perceptron (MLP) first evaluates a routing logit: Tx Tx at,i = gθ xt,i , mTx (12) t−1,i , et , |xt,i − mt−1,i |, xt,i ⊙ mt−1,i
Fig. 2: Skill-driven token prioritization and memory-assisted reassembly in SkillComm.
where We is a learnable linear projection that maps the skill embedding to the token feature dimension. A rate MLP rω receives average-pooled current, memory, and absolutedifference features, together with et and normalized SNR γt , and predicts ρt = ρmin + (ρmax − ρmin ) sigmoid rω (Xt , MTx t−1 , et , γt ) , (14)
states or receiver feedback. Because X̂t strictly preserves the original grid shape, the activated task head directly executes inference: ôt = Ht (X̂t ). (10) Finally, the receiver updates its memory as MRx t = X̂t . Let ot be the ground-truth target for the t-th skill and dt (·, ·) be the task distortion metric. SkillComm aims to minimize the cumulative workflow distortion subject to a total communication cost constraint:
minΘ E
hP
T t=1 αt dt (ot , ôt ) + λRtotal
i
,
(11)
where Θ encompasses the parameters of the shared feature backbone, JSCC codecs, task heads, and gating modules; αt balances task priorities, and λ dictates the accuracy-rate tradeoff. Here, Rtotal represents the cumulative transmission rate across the entire sequence of skills. III. S KILL -D RIVEN S EMANTIC C OMMUNICATION This section realizes the token selection policy Γθ (·) and the reassembly operator R(·) introduced in Sec. II. As shown in Fig. 2, SkillComm implements workflow-aware token prioritization through memory-aware token scoring, adaptive Top-K JSCC transmission, and memory-assisted token-grid reassembly, enabling state-active semantic innovations to be delivered across sequential skill stages. A. Token Scoring at Transmitter To facilitate cross-task information reuse without redundant transmission, both the transmitter and receiver cache their respective historical states. The transmitter memory is defined Tx Tx as MTx t−1 = [mt−1,1 , . . . , mt−1,N ], where each memory toTx ken mt−1,i is initialized as the corresponding feature xt−1,i generated at the previous stage t − 1. Similarly, the receiver Rx maintains a local memory MRx t−1 , where each token mt−1,i stores the previously reconstructed feature x̂t−1,i . A cosine-similarity heuristic can determine whether a current token resembles the historical memory, but fails to differentiate between task-irrelevant background and task-critical novel features. To resolve this, SkillComm introduces a delta-
where ρt is the retained-token fraction and (ρmin , ρmax ) = (0.2, 0.8). The transmitted token count and indices are It+ = TopK({st,i }N i=1 , Kt ). (15)
Kt = round(ρt N ),
Because hard Top-K is non-differentiable, training uses the e t = STTopK(st , Kt ) ∈ [0, 1]N . Its straight-through mask m forward value is the binary Top-K mask, while its backward derivative is provided by a soft sigmoid relaxation. Training then minimizes
L=
PT
t=1 Lt + λ
PT
t=2
e t ∥1 ∥m N ,
(16)
where Lt is the downstream loss for skill t and λ controls the accuracy–rate tradeoff. Thus, inference still transmits exactly e t exists only to train the hard Top-K token Kt tokens; m selector. B. Memory-Assisted Token Reassembly at Receiver Following token gathering and JSCC transmission of the active subset X+ t , the receiver performs cross-step feature reassembly. Because the receiver lacks access to the untransmitted portions of the transmitter’s current state, it cannot execute dynamic fusion mechanisms based on Xt . Instead, the receiver gracefully overwrites the selected positions with the newly arrived tokens and directly inherits the unselected positions from its local historical memory: ( + x̂+ t,πt (i) , i ∈ Ît , x̂t,i = (17) mRx / Ît+ , t−1,i , i ∈ where πt (i) is an order-preserving mapping from the original global index i to its compact position in the transmitted sequence X+ t . This pure memory-overwrite mechanism ensures robust feature restoration without introducing supplementary communication overhead. For a workflow spanning T stages, a prompt-guided fulltoken benchmark requires the transmission of T N tokens. In contrast, SkillComm incurs full communication cost only at the initial stage, yielding a drastically reduced cumulative workflow token utilization of: Rtotal = N +
T X t=2
Kt .
(18)
4
Concurrently, the index transmission overhead is bounded N by Bidx,t = ⌈log2 K ⌉. This analytical formulation explict itly guarantees that by dynamically modulating Kt via the rate controller, SkillComm effectively eliminates the repeated transmission of redundant semantic tokens. IV. E XPERIMENTAL R ESULTS This section evaluates SkillComm on a representative compound intent, “Capture the person’s body mechanics,” instantiated as Detect → Segment → Keypoint: localizing the person, delineating the body, and estimating the joints. This workflow is used as a representative benchmark to validate sequential skill execution, while the proposed mechanism is not tied to this specific task chain. We report stage-specific intent conditioning, workflow-level token efficiency, and atomic-task APbox , APmask , and APkp . A. Experimental Setup Experiments are conducted on 100 person-containing images randomly sampled from the MS COCO 2017 validation set [12]. A prompt-conditioned Swin Transformer processes 448 × 448 images and produces a 14 × 14 × 256 token canvas, yielding N = 196 spatial tokens per stage. Pre-trained Mask R-CNN and Keypoint R-CNN heads consume the reassembled canvas for detection, segmentation, and keypoint estimation. Therefore, a conventional full-token execution of the Detect → Segment → Keypoint workflow requires transmitting 3N = 588 tokens. During training, the delta-aware ranking gate is optimized for 10,000 iterations. The SNR-aware rate controller is trained for 5,000 iterations with the backbone and task heads frozen, followed by 500 Rayleigh-calibration iterations. The training SNR is continuously sampled from [0, 15] dB. For evaluation, Rayleigh fading is applied at {0, 3, 6, 9, 12, 15} dB with shared channel realizations across methods. We compare SkillComm with the following references and ablation variants: • NC-UB: The no-channel upper bound, where the full workflow is executed locally without wireless degradation. • TASC-f [6]: A task-adaptive SemCom reference based on prompt/conditional transmission, included only for applicable comparisons since it does not provide automatic workflow compilation or workflow-level token reuse. • PGMT-SC [7]: A single-task prompt-guided SemCom reference. For a compound intent, it executes one favorable atomic task and is therefore excluded from per-task AP curves. • SkillComm-Full: The sequential baseline, where task prompts are manually supplied step by step and all N = 196 tokens are transmitted at each stage. • SkillComm-Fixed: An ablation that retains automatic workflow compilation and execution memory, but uses deterministic cosine Top-K selection with a fixed 50% downstream token ratio. • SkillComm-Adaptive: The complete method with automatic skill compilation, cross-step memory, learned deltaaware token gating, and adaptive SNR-aware rate control.
TABLE II: Stage-conditioning alignment and workflow-level token efficiency. Method Effective Skill Sequence SCA Tokens† Saving NC-UB D→S→K (channel-free reference) 1 – – TASC-f† [6] D→D→D 33.3% 588.0 0.0% PGMT-SC† [7] M→M→M 66.3% 588.0 0.0% SkillComm-Full D→S→K 100.0% 588.0 0.0% SkillComm-Fixed D→S→K 100.0% 392.0 33.3% SkillComm-Adaptive D→S→K 100.0% 287.1 51.2% D, S, and K denote detection, segmentation, and keypoint prompts; M denotes the fixed MOP weights (0.30, 0.30, 0.40). † TASC-f and PGMT-SC are evaluation-unrolled over three task heads only to assess stage-conditioning mismatch under the same workflow, without native workflow execution.
B. Stage-Conditioning Alignment and Token Utilization This sub-section evaluates the overall intention compiler performance as well as token savings. Since TASC-f requires explicit task feedback, we instantiate its no-feedback adaptation by selecting one known atomic prompt through frozen multilingual MiniLM similarity [13]. The intent has similarities [0.447, 0.319, 0.357] to detection, segmentation, and keypoint prompts, respectively, and therefore selects detection for the entire request. Following the mixture-of-prompts (MOP) protocol of PGMT-SC, its unseen-intent embedding is 0.30edet + 0.30eseg + 0.40ekp . SkillComm first compiles the intent and then switches to the corresponding atomic embedding at each workflow stage. For the compound intent, let wt denote the effective routing weights over the atomic prompts at stage t, and let wt⋆ denote the target one-hot routing vector of the required skill. We measure stage-conditioning alignment (SCA) using a normalized Brier-style score [14]: PT 1 ⋆ 2 SCA = 1 − 2T (19) t=1 ∥wt − wt ∥2 . For the target D→S→K workflow, the desired routing vectors are w1⋆ = (1, 0, 0), w2⋆ = (0, 1, 0), and w3⋆ = (0, 0, 1). Both wt and wt⋆ lie on the probability simplex, whose maximum squared Euclidean distance is two, e.g., ∥(1, 0, 0) − (0, 1, 0)∥22 = 2. Hence, the factor 2T normalizes SCA to [0, 1]: a score of one means that every stage receives its correct atomic condition, while a mismatch decreases the score according to its full routing-vector distance. This score is evaluated in the discrete skill-routing space rather than by cosine similarity between prompt embeddings. A hard-selected prompt has one-hot routing weights, whereas PGMT-SC uses its normalized MOP coefficients directly. Tab. II separates stage-wise intent resolution from transmission cost for the same compound intent. TASC-f repeats wt = (1, 0, 0) at all stages; its three squared errors are 0, 2, and 2, giving 1 − 4/6 = 33.3% SCA. PGMT-SC repeats wt = (0.30, 0.30, 0.40); its stage-wise errors are 0.74, 0.74, and 0.54, giving 1 − 2.02/6 = 66.3%. This higher value reflects partial soft alignment, but not stage-specific decomposition. SkillComm compiles the intent into the correct one-hot sequence, incurs zero routing error, and obtains 100% SCA. Based on this workflow, its learned gate uses 287.1 rather than 588 tokens, leading to a token saving of 51.2%.
5
0.45 0.52
0.6 0.51
0.4
0.5
0.35
0.49
0.5
APkp
APmask
APbox
0.55
0.3
0.45
0.48
0.25 0.47
NC-UB TASC-f [6] PGMT-SC [7] SkillComm-Full SkillComm-Fixed SkillComm-Adaptive
0.46
0.4 NC-UB TASC-f [6] PGMT-SC [7] SkillComm-Full SkillComm-Fixed SkillComm-Adaptive
0.2
0.15 0
3
6
9
12
15
0.3 0
3
6
9
12
15
0
3
SNR (dB)
SNR (dB)
(a) Stage 1: Object Detection
NC-UB TASC-f [6] PGMT-SC [7] SkillComm-Full SkillComm-Fixed SkillComm-Adaptive
0.35
(b) Stage 2: Instance Segmentation
6
9
12
15
SNR (dB)
(c) Stage 3: Keypoint Detection
Fig. 3: Per-task COCO Average Precision (AP) over Rayleigh fading channels for the Detect→Segment→Keypoint workflow.
C. Per-Task AP Evaluation over Rayleigh Fading Fig. 3 reports the per-task AP of the three workflow stages under Rayleigh fading channels. In the first stage, all SkillComm variants transmit the full token grid and therefore achieve nearly identical detection performance. TASC-f also performs well at this stage because it interprets the compound intent as a detection prompt, whereas PGMT-SC suffers from mixed multi-task conditioning. In the segmentation and keypoint stages, however, TASC-f continues to use the detection condition and degrades significantly, while SkillComm maintains stage-specific execution through the skill compiler. Among the SkillComm variants, SkillComm-Full provides the upper performance reference by transmitting all tokens, whereas SkillComm-Fixed and SkillComm-Adaptive reduce token usage through selective transmission. SkillComm-Fixed is more robust at low SNR due to its conservative token ratio, while SkillComm-Adaptive surpasses it at high SNR by selecting higher-value tokens according to channel and memory-aware token priorities. Overall, SkillComm-Adaptive achieves 94.3% average AP retention relative to NC-UB across 0–15 dB and all three tasks, and reaches 99.4% retention at 15 dB, demonstrating that workflow-aware token prioritization preserves task-critical semantics with substantially lower communication cost. V. C ONCLUSIONS This letter proposed SkillComm, a skill-driven SemCom framework that uses reusable skill states to guide workflowaware token prioritization and memory-assisted token-grid reassembly. Instead of treating each task as an isolated fulltoken transmission problem, SkillComm synchronizes ordered skill execution through a shared Skill-Book, exploits execution memory to prioritize state-active tokens for JSCC transmission, and restores task-ready token grids at the receiver. Experiments on the MS COCO 2017 Detect→Segment→Keypoint workflow show that SkillComm reduces token transmission cost by 51.2%, while retaining 99.4% upper-bound-normalized AP in the high-SNR 15-dB Rayleigh fading regime. These results validate a general principle for workflow-level SemCom: reusable skill states can guide wireless systems to transmit semantic innovations rather than repeated full representations.
Future work will develop feedback-loop skills that adapt the next skill, token budget, and refinement transmission using receiver-side confidence, channel state information, and memory-reassembly errors. R EFERENCES [1] H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, 2021. [2] G. Zhang, Q. Hu, Z. Qin, Y. Cai, G. Yu, and X. Tao, “A unified multitask semantic communication system for multimodal data,” IEEE Trans. Commun., vol. 72, no. 7, pp. 4101–4116, 2024. [3] Z. Qin et al., “AI empowered wireless communications: From bits to semantics,” Proc. IEEE, Jul. 2024. [4] E. Bourtsoulatze, D. B. Kurka, and D. Gündüz, “Deep joint sourcechannel coding for wireless image transmission,” IEEE Trans. on Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, 2019. [5] L. Qiao, M. B. Mashhadi, Z. Gao, R. Tafazolli, M. Bennis, and D. Niyato, “Token communications: A large model-driven framework for cross-modal context-aware semantic communications,” IEEE Wireless Commun., vol. 32, no. 5, pp. 80–88, 2025. [6] J. He, Y. Deng, S. Wu, B. Cao, S. Meng, W. Zhang, and Q. Zhang, “Taskadaptive semantic communication with feedback: A conditional ratedistortion approach,” IEEE Trans. on Cogn. Commun. Netw., vol. 12, pp. 7157–7171, 2026. [7] H. Zhang, Z. Cao, L. Liang, H. Ye, S. Jin, and G. Y. Li, “Promptguided multi-task semantic communication for image transmission,” IEEE Commun. Lett., vol. 30, pp. 1111–1115, 2026. [8] T. Gupta and A. Kembhavi, “Visual programming: Compositional visual reasoning without training,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2023, pp. 14 953–14 962. [9] K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask R-CNN,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct. 2017, pp. 2961–2969. [10] W. Li, R. Zhang, R. Shao, Z. Fang, K. Zhou, Z. Tian, and L. Nie, “SemanticVLA: Semantic-aligned sparsification and enhancement for efficient robotic manipulation,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2026. [11] J. Fu, M. Xiao, and M. Skoglund, “SkillCom: Decomposing LLMbased semantic communication into task and channel aware skills,” arXiv preprint arXiv:2605.02333, 2026. [12] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft COCO: Common objects in context,” in European Conference on Computer Vision. Springer, 2014, pp. 740–755. [13] W. Wang, F. Wei, L. Dong, H. Bao, M. Yang, and M. Zhou, “MiniLM: Deep self-attention distillation for task-agnostic compression of pretrained transformers,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 33, 2020, pp. 5776–5788. [14] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in International Conference on Machine Learning (ICML), 2017, pp. 1321–1330.