LLM Agents for 6G: Mitigating Anchoring Bias, 25% Energy Savings

Hatim Chergui, Claudia Carballo González, Farhad Rezazadeh and Merouane Debbah present an autonomous agentic resource negotiation framework to enable zero-touch network slicing in 6G using LLM agents. The authors identify anchoring bias in such agents and introduce a randomized anchoring strategy, combined with burst-aware digital twins using Conditional Value at Risk, that the paper says can raise system energy savings by up to 25%.

What did the authors propose and why

The paper proposes a negotiation framework that injects mathematically bounded randomness into initial proposals to mitigate anchoring bias, while using digital twins with CVaR to guarantee strict SLA tail-latencies. The randomized anchoring is modeled via a Truncated 3-Parameter Weibull distribution and integrates with burst-aware digital twins that employ Conditional Value at Risk to bound tail latency violations. The authors also state a formal result they call the Bimodal Constraint-Avoidance Utility Theorem, which shows feasible negotiations obey classical convex bounds, while highly constrained scenarios undergo a phase transition governed by an inverse rational decay envelope.

How was the approach validated and what were the key numbers

The authors evaluated the method using a locally hosted 1B-parameter model named otel-llm-1b-it and report empirical confirmation of the paper's dual-regime bounds. The lightweight 1B LLM achieves sub-second inference latencies with a 0.95s mean, which the authors highlight as compatible with the operational timescales of the O-RAN non-Real-Time RAN Intelligent Controller non-RT RIC. The evaluation produced the claim that cognitive de-biasing forced agents into active exploration and yielded system energy savings up to 25%.

How does the randomized anchoring work in practice

Randomized anchoring perturbs initial heuristic proposals so agents do not rigidly adhere to an anchor and over-provision resources. The perturbation is explicitly modeled with a Truncated 3-Parameter Weibull distribution, a bounded randomization that the framework uses during multi-agent negotiation. Digital twins that are burst-aware evaluate candidate allocations under CVaR constraints to ensure service-level agreement tail-latencies remain within strict bounds before any change is applied to the network slice.

Why it matters

If LLM agents can operate with sub-second inference and avoid anchoring-driven over-provisioning, they become practical for non-RT RIC timescales while lowering energy use. The combination of a bounded randomization model, formal utility analysis via the Bimodal Constraint-Avoidance Utility Theorem, and CVaR-backed digital twins addresses both the cognitive and the safety sides of automated negotiation. Operators seeking energy efficiency in 6G slicing could adopt such mechanisms to reduce waste from rigid heuristics while preserving tail-latency guarantees.

What to watch

Watch whether the method scales beyond the locally hosted 1B-parameter otel-llm-1b-it and whether independent evaluations reproduce the reported up-to-25% energy savings and the 0.95s mean inference latency. The authors note their source code is available for non-commercial use at this https URL, which will be a place to follow for replication and extensions.

Additional notes: the paper was submitted 5 Jun 2026 and revised 18 Jun 2026. It runs seven pages with four figures and situates its contributions at the intersection of agentic LLM reasoning, formal utility bounds, and risk-aware digital twin evaluation for zero-touch 6G network slicing.