SkillCom: Decomposing LLM-based Semantic Communication into Task and Channel Aware Skills

• Monolithic LLM systems jointly handle semantic abstraction, channel encoding, and downstream decoding, making their processing stages difficult to isolate and diagnose.A single text-block representation can also let localized channel impairment disrupt global semantic coherence.Source: §I · p. 1
• Independent semantic-unit transmission localizes channel impairment so successfully received units can support targeted repair.The design also supports stage-wise diagnosis and modular replacement while the rest of the processing chain remains fixed.Source: §I · pp. 1–2
• SkillCom decomposes the pipeline into four independently replaceable skills connected through typed interfaces.The skills are semantic abstraction, channel-adaptive transmission, receiver-side repair, and task execution.Source: §I · p. 1
• Evaluations on multi-hop question answering and dialogue state tracking show SkillCom outperforms monolithic LLM models, with preferred skill realizations depending on the task.Source: §I · p. 2

• The design goal is to maximize task utility under a communication budget, with the receiver producing an output from source text and a task descriptor.The model measures task performance using a utility function against a reference output.Source: A. Problem Formulation · p. 2
• SkillCom decomposes transmitter–receiver processing into sequential skills, with the transmitter first abstracting and then selecting semantic units.Source: A. Problem Formulation · p. 2
• A subset of semantic units is transmitted through the channel as the representation that the receiver receives in corrupted form.The transmitted representation is denoted US and the corrupted received representation U-tilde_S.Source: A. Problem Formulation · p. 2
• The receiver repairs the received representation into a repaired semantic context and then executes the task to produce its output.Source: A. Problem Formulation · p. 2

B. Channel Model

• Each selected semantic unit is treated as an independently coded packet under the AWGN channel model, with SNR as the channel state.The model uses BPSK modulation and an error-detecting code.Source: B. Channel Model · p. 2
• A packet is declared erased whenever the error-detecting code detects any bit error; otherwise, the unit is delivered intact.Source: B. Channel Model · p. 2
• The coded length depends on unit byte length and robustness score, with the robustness factor adjusting coded length and code rate accounting for error-detection overhead.Source: B. Channel Model · p. 2

˜US | US, SNR

• The channel independently erases transmitted units, and the receiver repairs received units before task execution.The packet-level transition assigns erasure probability PER_j and successful reception probability 1 − PER_j.Source: Figure 2 · p. 3, ˜US | US, SNR · p. 3

C. Semantic Unit Representation

• Each semantic unit is typed and can be selected, transmitted, and repaired independently rather than sent as one compressed text block.The six-field representation includes payload, type, task relevance, source importance, channel robustness, and token cost; the interface is fixed across downstream skills.Source: C. Semantic Unit Representation · p. 3
• Unit-level transmission leaves remaining received units available for downstream repair and task execution when one unit is erased.Units carry semantic value and transmission cost, supporting selection under explicit rate constraints.Source: C. Semantic Unit Representation · p. 3

III. SKILL MODULES AND REALIZATIONS

• SkillCom presents representative realizations of its four skills, with Fig. 2 showing one realization of the processing chain.The abstraction stage maps raw source text to typed units through a shared six-field interface, allowing its realizations to be interchangeable downstream.Source: §III · p. 3, A. Semantic Abstraction Skill · p. 3

A. Semantic Abstraction Skill

• Heuristic abstraction uses TF-IDF salience and entity cues, emphasizing evidence-bearing spans for QA and domain–slot–value structure for DST.Candidate spans are mapped to task-appropriate unit types.Source: A. Semantic Abstraction Skill · p. 3
• LLM-enriched abstraction adds generated keywords, entities, and concise summaries to broaden coverage of implicit or paraphrased information.It preserves the same typed interface as the other abstraction realizations.Source: A. Semantic Abstraction Skill · p. 3
• Structured LLM extraction uses a JSON-schema-constrained call to output target unit types while preserving the shared semantic-unit interface.The paper states that different tasks favor different abstraction modes.Source: A. Semantic Abstraction Skill · p. 3

B. Channel-Adaptive Transmission Skill

• Transmission selects a subset of semantic units under channel and budget constraints, prioritizing units for task success, communication efficiency, and robustness.The channel state is instantiated by SNR, and the unit priority score combines relevance, source importance, channel robustness, and transmission cost.Source: B. Channel-Adaptive Transmission Skill · p. 4
• Greedy selection ranks units by priority and adds them until a budget constraint would be exceeded.The LLM-deduplication variant adds semantic deduplication and reranking before final budget enforcement.Source: B. Channel-Adaptive Transmission Skill · p. 4

C. Receiver Repair Skill

• Receiver repair identifies erased units and uses successfully received units with the task descriptor to recover missing information or support downstream inference.The repaired output stays in the semantic-unit space, keeping communication recovery and task inference separately diagnosable.Source: C. Receiver Repair Skill · p. 4
• Generative repair synthesizes replacement units from received units and the query, while guided repair derives structured guidance without synthesizing new units.Generative repair suits tasks where partial recovery can improve inference; guided repair suits conservative structured outputs sensitive to hallucinations.Source: C. Receiver Repair Skill · p. 4

D. Task Execution Skill

• The execution skill maps repaired semantic representations to final outputs and is the only stage that commits to the downstream output space.It produces short answers for QA and canonical slot–value pairs for DST; its shared input interface means upstream skill changes do not require changing execution.Source: D. Task Execution Skill · p. 4

1) Tasks:

• Evaluation uses HotpotQA for multi-hop question answering and MultiWOZ 2.4 for multi-domain dialogue state tracking.Source: 1) Tasks · p. 4

2) Methods Under Comparison:

• SkillCom-Heuristic uses heuristic abstraction, Enrich adds LLM-generated keywords, entities, and summaries, Struct extracts units with an LLM, and Struct+Dedup adds LLM deduplication.The baseline compresses the full source into one text block and performs task decoding without explicit stage decomposition.Source: 2) Methods Under Comparison · p. 4

3) Channel and Budget:

• Comparisons use a coded packet erasure channel at coding rate R=0.5 with matched budgets: (4, 48, 300) for HotpotQA and (5, 56, 350) for MultiWOZ.The budget tuple is (Bu, Bκ, Bc).Source: 3) Channel and Budget · p. 4

4) Metrics and Implementation:

• HotpotQA is evaluated with EM and token-level F1, while MultiWOZ uses JGA and slot-level F1; each task uses 100 sampled examples and cached DeepSeek-Chat calls at temperature 0.Repair uses the generative strategy for HotpotQA and guided strategy for MultiWOZ.Source: 4) Metrics and Implementation · p. 4

B. Main Comparison

• At 7 dB, SkillCom-Struct+Dedup improves HotpotQA EM from 0.42 to 0.56 and F1 from 0.51 to 0.68; SkillCom-Enrich reaches 0.08 JGA and 0.42 Slot F1 on MultiWOZ, versus baseline values of 0.02 and 0.03.All SkillCom variants recover meaningful dialogue state on MultiWOZ.Source: B. Main Comparison · pp. 4–5, Table I · p. 5, Table II · p. 5

C. Noise Robustness

• Across 4–14 dB, SkillCom degrades more gracefully than the monolithic baseline; at 4 dB, SkillCom preserves meaningful performance while the baseline approaches zero.On HotpotQA, Struct+Dedup remains effective even near the baseline's catastrophic-degradation regime around 7–8 dB; skill preferences persist across noise levels.Source: C. Noise Robustness · p. 5

D. Skill Ablation

• The ablation removes one component at a time from each task's best-performing variant while holding the other skills fixed at the main comparison's 7 dB setting.Tables III and IV report the results.Source: D. Skill Ablation · p. 5
• On HotpotQA at 7 dB, removing channel-aware transmission causes the largest F1 drop (−0.20), followed by removing LLM abstraction (−0.16) and deduplication (−0.09).The full model is SkillCom-Struct+Dedup.Source: D. Skill Ablation · p. 5, Table III · p. 5
• On MultiWOZ at 7 dB, removing LLM-enriched abstraction lowers Slot F1 from 0.42 to 0.36 and JGA from 0.08 to 0.05; removing transmission or repair has minor effect.The full model is SkillCom-Enrich.Source: D. Skill Ablation · p. 5, Table IV · p. 5

E. Budget Sensitivity

• At the tightest HotpotQA budget (Bu=2), SkillCom-Struct+Dedup attains F1 = 0.65, compared with 0.48 for the monolithic baseline.SkillCom maintains an advantage at every evaluated budget level.Source: E. Budget Sensitivity · p. 5, Table V · p. 6
• On MultiWOZ, SkillCom-Enrich Slot F1 rises from 0.33 to 0.47 as budget increases, while the monolithic baseline stays nearly flat at 0.04–0.09.The authors report that SkillCom converts additional budget into task performance more effectively.Source: E. Budget Sensitivity · p. 5, Table VI · p. 6

V. CONCLUSION

• The conclusion reports that SkillCom outperforms the monolithic baseline in task accuracy, noise robustness, and budget efficiency, while ablations reveal task-dependent bottlenecks.The authors describe skill decomposition and unit-level transmission as a more robust and diagnosable foundation for LLM-based semantic communication.Source: §V · pp. 5–6