Mitigating Anchoring Bias in LLM-Based Agents for Energy-Efficient 6G Autonomous Networks

Researchers present an LLM-based autonomous framework for 6G network resource negotiation that addresses anchoring bias—a cognitive limitation causing agents to over-provision resources. Using a Weibull distribution-based randomization strategy combined with Digital Twins and CVaR constraints, the system achieves up to 25% energy savings while maintaining SLA compliance, with a 1B-parameter model delivering sub-second inference latencies suitable for O-RAN deployment.

This research tackles a fundamental challenge in deploying LLM agents for infrastructure automation: cognitive biases that degrade system performance. The authors identify that language models exhibit anchoring bias when negotiating network resources, causing them to rigidly accept initial proposals and over-provision capacity. Rather than treating this as an insurmountable limitation, they develop a mathematically rigorous mitigation strategy using truncated Weibull distributions to inject controlled randomness into negotiation anchors, preventing agents from fixating on suboptimal solutions.

The broader context reflects growing interest in applying AI agents to telecommunications infrastructure, particularly in emerging 6G standards where autonomous zero-touch provisioning offers operational efficiency gains. However, deploying LLMs in critical infrastructure requires guarantees on both performance and resource consumption—constraints that traditional LLM applications ignore. The integration with Digital Twins and Conditional Value at Risk demonstrates how safety-critical systems can leverage LLM reasoning while maintaining formal reliability bounds.

For infrastructure operators and telecom vendors, this work suggests that energy-efficient network operation is achievable through de-biased AI agents, offering competitive advantages in operational expenditure reduction. The 25% energy savings directly translate to cost reduction and environmental benefits. The lightweight 1B model achieving sub-second latencies proves that enterprise-grade AI autonomy doesn't require expensive large-scale models, reducing deployment costs.

The practical implications extend beyond 6G: cognitive de-biasing techniques could improve LLM agent performance across distributed systems requiring resource negotiation. Future research should explore how these randomization strategies scale to multi-agent scenarios with competing objectives and whether the phase transition behavior observed generalizes to other infrastructure domains.

• →LLM agents exhibit anchoring bias that causes network over-provisioning; randomized Weibull-based strategies effectively mitigate this cognitive limitation.
• →The framework achieves 25% energy savings while maintaining strict SLA tail-latency guarantees through CVaR-constrained Digital Twins.
• →A 1B-parameter model delivers sub-second inference (0.95s mean), proving lightweight models can serve real-time infrastructure automation.
• →Bimodal constraint-avoidance behavior emerges at phase transition points, with feasible scenarios following convex bounds and constrained scenarios showing inverse rational decay.
• →Integration with O-RAN non-RT RIC demonstrates compatibility with existing 6G standards, enabling practical deployment in production networks.