LLM Agents for 6G Networks Mitigate Anchoring Bias for Energy Efficiency

Researchers have developed an autonomous agentic resource negotiation framework specifically designed for zero-touch network slicing in future 6G architectures, leveraging Large Language Model (LLM) agents. While LLMs offer powerful reasoning capabilities for such complex tasks, the study identifies a critical flaw: these agents inherently exhibit anchoring bias. This bias causes them to rigidly adhere to initial proposals, often leading to significant network over-provisioning, which is inefficient. To counteract this cognitive bias, the paper proposes a novel randomized anchoring strategy. This strategy is mathematically modeled using a Truncated 3-Parameter Weibull distribution, providing a bounded approach that seamlessly integrates with burst-aware Digital Twins. These Digital Twins, in turn, employ Conditional Value at Risk (CVaR) to rigorously guarantee strict Service Level Agreement (SLA) tail-latencies, ensuring network performance remains robust. The methodology is supported by the "Bimodal Constraint-Avoidance Utility Theorem," which demonstrates distinct negotiation behaviors under different constraint levels. Empirical results, obtained using a lightweight 1B-parameter LLM, confirm these theoretical bounds. The cognitive de-biasing technique successfully breaks rigid negotiation patterns, prompting agents to actively explore optimal solutions that ride close to SLA boundaries. This approach not only ensures compliance but also achieves substantial energy savings of up to 25%, all while maintaining sub-second inference latencies compatible with O-RAN non-Real-Time RIC operational timescales.