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

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

Problem Statement
While Large Language Models (LLMs) have emerged as powerful semantic encoders and decoders, current LLM-based semantic communication systems largely remain monolithic. In these systems, a single prompted model or a tightly coupled transmitter/receiver pair jointly performs semantic abstraction, channel adaptation, and semantic decoding without explicit stage separation. This architecture presents three critical limitations:

1. Conflicting Objectives: The model must simultaneously optimize source compression, channel encoding, and task fidelity within a single prompt.
2. Fragility to Channel Impairment: Transmission typically occurs as a single compressed text block. Consequently, localized channel errors can disrupt global semantic coherence, leading to catastrophic failure.
3. Lack of Diagnosability: Intermediate processing decisions are hidden within the prompt's "black box," making it difficult to attribute failures to specific stages, diagnose issues, or replace components.

Methodology: The SkillCom Framework
To address these limitations, the authors propose SkillCom, a modular framework that decomposes the semantic communication pipeline into four explicit, independently replaceable skills connected via typed semantic-unit interfaces. The core design principle is independent semantic-unit transmission, where the source is abstracted into structured units rather than a monolithic text block.

The framework operates through four sequential skills:

1. Semantic Abstraction Skill ($ABS$): Maps raw source text ($x$) and task descriptors ($T$) into a set of typed semantic units ($U$). The paper proposes three realizations:

   • Heuristic: Deterministic extraction using TF-IDF and entity cues.
   • LLM-Enriched: Augments heuristics with LLM-generated keywords and summaries.
   • Structured LLM: Uses JSON-schema-constrained LLM calls to directly output unit types.
     Each unit $u_i$ is defined by a six-field interface: semantic payload ($\xi$), type ($\tau$), task relevance ($r$), source importance ($s$), channel robustness ($g$), and token cost ($\kappa$).

2. Channel-Adaptive Transmission Skill ($TRANS$): Selects a subset of units ($S$) for transmission based on channel state (SNR) and communication budget ($B$). This is formulated as a constrained utility maximization problem, prioritizing units that are task-relevant, source-important, and robust to noise, while penalizing transmission cost. Realizations include Greedy Selection and Greedy Selection with LLM Deduplication.

3. Receiver Repair Skill ($REPAIR$): Reconstructs a usable semantic context from successfully received units ($S_{rx}$) and the task descriptor. It identifies erased units ($S_{\emptyset}$) and attempts recovery via:

   • Generative Repair: Synthesizing replacement units via an LLM.
   • Guided Repair: Deriving structured guidance (e.g., active domains, confirmed slots) from surviving context without synthesizing new content.

4. Task Execution Skill ($EXEC$): A task-conditioned LLM decoder that maps the repaired semantic representation ($\hat{U}$) to the final output ($\hat{y}$). Because the input interface remains fixed, changes in upstream skills do not require modifying the execution stage.

Key Contributions

1. Explicit Skill Decomposition: SkillCom is presented as the first framework to decompose LLM-based semantic communication into four distinct, replaceable skills with typed interfaces, enabling stage-wise ablation and targeted diagnosis.
2. Independent Semantic-Unit Transmission: By transmitting typed units individually rather than as a single block, the framework localizes channel impairment. Packet loss affecting one unit does not invalidate others, enabling targeted receiver-side repair.
3. Task-Dependent Optimization: The framework demonstrates that optimal skill realizations vary by task, allowing for modular optimization of individual components.

Experimental Results
The authors evaluated SkillCom on HotpotQA (multi-hop question answering) and MultiWOZ 2.4 (dialogue state tracking) under varying Signal-to-Noise Ratio (SNR) and communication budgets.

• Performance vs. Monolithic Baseline: SkillCom consistently outperformed the monolithic LLM baseline. On HotpotQA at SNR=7 dB, the best SkillCom variant (Struct+Dedup) improved Exact Match (EM) from 0.42 to 0.56 and F1 from 0.51 to 0.68. On MultiWOZ, the monolithic baseline achieved near-zero Joint Goal Accuracy (JGA=0.02), whereas SkillCom variants recovered meaningful performance (up to JGA=0.08).
• Noise Robustness: As SNR decreased, the monolithic baseline suffered catastrophic degradation. In contrast, SkillCom variants maintained substantially higher utility, particularly at low SNR (4 dB), confirming that unit-level transmission prevents total message failure.
• Ablation Studies:
  • For HotpotQA, channel-aware transmission (unit prioritization) was the most critical component; removing it caused the largest performance drop ($\Delta F1 = -0.20$).
  • For MultiWOZ, LLM-enriched abstraction was the dominant factor; removing it significantly reduced Slot F1 and JGA, while transmission and repair had minor effects.
• Budget Sensitivity: SkillCom demonstrated superior efficiency across all budget levels. On MultiWOZ, increasing the budget significantly improved SkillCom's performance (Slot F1 from 0.33 to 0.47), whereas the monolithic baseline remained nearly flat.

Significance and Claims
The paper claims that explicit skill decomposition provides a more robust and diagnosable foundation for LLM-based semantic communication compared to monolithic methods. By converting channel impairment from whole-message failures into localized unit-level erasures, the framework enables:

• Targeted Repair: Recovery from successfully received units without re-transmitting the entire message.
• Modularity: The ability to replace or ablate specific skills (e.g., swapping a heuristic abstraction for a structured LLM) without redesigning the entire system.
• Task-Specific Tuning: The identification of task-dependent bottlenecks (e.g., transmission for QA vs. abstraction for DST) allows for precise optimization that is difficult to achieve in end-to-end monolithic systems.

The authors conclude that while LLMs possess the capability for semantic communication, their deployment as monolithic blocks limits their robustness and controllability; SkillCom addresses this by organizing complex behaviors into explicit, reusable modules.