LLM Anchoring Bias Wastes 25% Energy in 6G Network Slicing

Up to 25% of energy in LLM-driven 6G network slicing goes to waste because agents anchor on their first heuristic estimate. That's not a simulation artifact - the paper's authors prove it with a concrete 1B-parameter model running on real O-RAN timescales.

Anchoring Bias: The Hidden Tax on LLM-Based 6G Slicing

Large Language Models bring powerful reasoning to autonomous network slicing, but they inherit a cognitive flaw: anchoring bias. Once an agent proposes an initial resource allocation, it rigidly sticks to that heuristic, even when the network conditions shift. The result is severe over-provisioning - more bandwidth and compute allocated than any service actually needs.

That wasted capacity burns energy. The paper's empirical results, generated using a locally hosted otel-llm-1b-it model, confirm that unmitigated agents settle into negotiation patterns that leave energy efficiency on the table. Standard convex optimization theory says negotiations should converge to efficient points, but anchoring creates a systematic deviation.

A Randomized Weibull Cure

The solution is mathematically precise: a randomized anchoring strategy modeled with a Truncated 3-Parameter Weibull distribution. This isn't a hand-wavy temperature tweak. The distribution is mathematically bounded, and it integrates with burst-aware Digital Twins using Conditional Value at Risk (CVaR) to guarantee strict SLA tail-latencies.

The authors introduce and prove the Bimodal Constraint-Avoidance Utility Theorem, which shows that feasible negotiations follow classical convex bounds, but highly constrained scenarios undergo a phase transition governed by an inverse rational decay envelope. In plain terms: when things get tight, the usual optimization fails - and the Weibull randomization forces agents to actively explore instead of locking into a bad initial guess.

Sub-Second Inference with 1B Parameters

A cognitive fix is useless if it can't run in real time. The otel-llm-1b-it model achieves a mean inference latency of 0.95 seconds. That's fast enough to slot into the O-RAN non-Real-Time RAN Intelligent Controller (non-RT RIC) operational cycle without custom hardware.

Energy savings hit 25% across the tested scenarios. The agents safely ride the boundary of SLA compliance rather than padding every allocation. This is the difference between a network that optimizes and one that just reacts.

The lightweight 1B model proves that you don't need a massive parameter count to drive autonomous 6G operations - you need the right cognitive architecture. Expect this line of work to show up in real O-RAN deployments before the 6G standard freezes.

Source: Mitigating Anchoring Bias in LLM-Based Agents for Energy-Efficient 6G Autonomous Networks
Domain: arxiv.org