- The paper introduces AgentComm, a framework that combines LLM-based message compression, importance-aware semantic transmission, and task-specific knowledge bases for bandwidth-limited robot–base-station communication.
- The framework cuts bandwidth use by approximately 50% while preserving task performance, with the knowledge-base-enhanced system achieving 100% success at 5–10 dB SNR and 91% at 0 dB in detailed inspection tasks.
- The results show that compression is highly task-dependent: it works well for concise household instructions but can remove critical details from inspection messages unless key information receives stronger protection and task-specific memory support.
Overview and Motivation
This paper addresses the physical-layer communication overhead incurred by LLM-driven agent-to-agent (A2A) communication in embodied AI systems. The authors observe that messages generated by LLM-based planning and reasoning agents contain substantial redundancy and task-irrelevant content, which becomes a bottleneck when exchanged over bandwidth-limited, noisy wireless channels. Their proposal, AgentComm, is a semantic agent communication framework that combines an LLM-based semantic processor, an importance-aware transmission strategy, and a task-specific knowledge base (KB), targeting a scenario in which a resource-constrained robot agent collaborates with a powerful base station (BS) agent over OFDM links.
System Model
The architecture places a full LLM and complex agents at the BS, which decomposes user requirements into executable steps, maintains user profiles and interaction histories, and validates robot responses. The robot acts as a constrained executor and lightweight preprocessor that interacts with the environment via onboard sensors. Communication proceeds over multiple rounds, with queries and responses mapped onto OFDM subcarriers; channel quality is abstracted through an effective SNR (ESNR) metric that maps frequency-selective fading to an equivalent AWGN condition. The paper identifies three sources of information loss in this pipeline: the LLM compressor, the trainable semantic encoder-decoder, and the wireless channel itself. It also notes a structural limitation of the deployment: robots have limited memory and privacy constraints prevent public devices from retaining sensitive user information, which motivates the BS-centric knowledge architecture.
Proposed Framework
The framework has three components. First, an LLM-based compressor (LC) reorganizes and condenses messages prior to transmission using prompts alone, requiring no training; the robot-side compressor deliberately does not store historical messages to reduce computational burden and protect user privacy. Second, an importance-aware transmission strategy partitions each message via an LLM-based extractor into three parts—key items with position markers, sentences containing key items, and remaining sentences—and maps these parts to subchannels grouped by SNR, assigning key items a stronger semantic encoder configuration (n′>n bits per sentence). Third, a task-specific KB at the BS functions as long-term semantic memory: an LLM-based evaluator aggregates historical messages and user feedback after a task's first execution, extracts missing details, and injects them into the robot's extraction prompt for future executions of the same task.
The paper also articulates a conceptual distinction between semantic encoders, which produce implicit, fixed-dimensional representations optimized end-to-end for reconstruction, and LLM-based methods, which produce explicit, human-readable, lossy text abstractions guided by prompts. The two are complementary: LLM processing shortens text before semantic coding, while the semantic encoder-decoder handles channel variability.
Evaluation
Simulations cover two scenarios derived from existing embodied datasets: a warehouse inspection task (Case 1, long detailed messages) and a household cleaning task (Case 2, short messages), with the BS agents built on Claude-Sonnet-4.5 and robot agents on GPT-4o. The baseline is Huffman source coding with rate-1/2 LDPC channel coding over 4-QAM OFDM. Evaluation uses task success rate (SR) and Distinct-1 diversity, in addition to bandwidth consumption.
The headline result is approximately 50% bandwidth reduction with negligible task-performance loss. In Case 1, the LC alone requires roughly 1/3 of downlink and 1/8 of uplink resources relative to direct transmission, but this compression causes detail loss: SR drops to 95% at 10 dB SNR, and the LC+SC configuration degrades further as SNR falls (75% at 0 dB). In Case 2, compression is essentially lossless for task purposes—LC+SC maintains 100% SR at 10 and 5 dB and 87% at 0 dB, where the conventional baseline fails entirely (0% SR at 5 and 0 dB because transmission errors prevent task completion within the 5-round limit). This asymmetry is a notable finding: the benefit of aggressive compression depends strongly on message structure, with instruction-following tasks tolerant of compression and detail-rich inspection tasks not.
The importance-aware and KB-enhanced variants address the Case 1 deficiency. LC+SC(Im) stabilizes SR at 82% across all SNRs but cannot reach the best performance because the extractor, lacking task knowledge, consistently omits certain details. Adding the KB raises SR to 100% at 10 and 5 dB and 91% at 0 dB, at the cost of modestly higher bandwidth and an upfront KB transmission of several hundred bits. The ablation study quantifies the complementary roles of the components: roughly 40% of agent-generated content can be removed by the LC without losing key items; LC+SC achieves the best SR once the bandwidth ratio exceeds 0.3; and the ablation also reveals a trade-off the authors acknowledge plainly—LC compression strips redundancy from messages, which amplifies the impact of residual semantic errors, so LC+SC underperforms pure SC at low SNR.
Limitations and Open Questions
The paper's evaluation is deliberately narrow: only two scenarios generated from existing datasets, with a maximum of five communication rounds, and no claim of generalization across tasks. The importance-aware mechanism's effectiveness depends on correct key-item selection by the LLM extractor; the authors concede that incorrect choices lead to loss of fine-grained details in task-specific cases, which is precisely what the KB mechanism patches—but only for recurring tasks and only after a first execution with detailed user feedback. The framework also assumes an asymmetric architecture with a computationally powerful BS, and the KB update requires user participation for unseen tasks. Interoperability remains an acknowledged open issue: without standardized semantic representations, different agents may interpret the same information inconsistently. Whether the reported bandwidth savings hold under multi-user settings, dynamic channel conditions beyond the tested SNR points, or tasks requiring more than five rounds is left unexamined.
Conclusion
AgentComm demonstrates that LLM-based explicit semantic processing, combined with importance-aware physical-layer protection and task-specific knowledge accumulation, can halve the bandwidth consumed by embodied multi-agent communication while preserving or improving task success rates relative to conventional coding—particularly in low-SNR regimes where conventional schemes fail outright. The results also delineate the boundary conditions of this approach: compression is safe for structured, low-detail messages but requires unequal protection and learned task knowledge for detail-rich content.