🤖 AI

LLM-Based Agentic Negotiation for 6G: Addressing Uncertainty Neglect and Tail-Event Risk

This paper proposes a risk-aware, LLM-based agentic negotiation framework for 6G networks that leverages Digital Twins and Conditional Value-at-Risk (CVaR) to overcome uncertainty neglect bias, thereby eliminating SLA violations and tail-event risks in resource allocation while maintaining feasible inference speeds.

Original authors: Hatim Chergui, Farhad Rezazadeh, Mehdi Bennis, Merouane Debbah, Christos Verikoukis

Published 2026-05-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hatim Chergui, Farhad Rezazadeh, Mehdi Bennis, Merouane Debbah, Christos Verikoukis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the future internet (called 6G) as a massive, busy highway system where different types of vehicles need to travel at the same time. Some are heavy trucks carrying huge amounts of data (like streaming movies), and others are tiny, super-fast race cars that must arrive instantly (like self-driving cars or remote surgery).

To keep traffic flowing, the highway needs a smart traffic controller. In this paper, the authors propose using AI agents (smart computer programs) to negotiate who gets how much road space and computing power.

However, there's a big problem: these AI agents are prone to a specific kind of "cognitive blind spot."

The Problem: The "Average" Trap

The paper argues that current AI agents suffer from Uncertainty Neglect. Think of it like a weather forecaster who only looks at the average temperature for the next week.

The Biased Agent: "The average temperature is 70°F, so it's a perfect day for a picnic!"
The Reality: The average hides the fact that on Tuesday, there will be a massive blizzard that cancels the picnic.

In the 6G world, the "blizzard" is a sudden, extreme spike in network traffic (a "tail event"). If the AI only plans for the average traffic, it will allocate just enough resources to get by. But when that rare, massive spike happens, the network crashes, and the "race cars" (critical services) miss their deadlines.

The Solution: The "Paranoid" Planner

The authors propose a new, "unbiased" framework that forces the AI to stop looking at averages and start looking at the worst-case scenarios. They use three main tools to fix this:

1. The Digital Twin (The Crystal Ball)

Instead of guessing, the AI uses a Digital Twin. Imagine a perfect, virtual copy of the entire highway system running in a simulator. Before making a decision, the AI runs thousands of simulations in this twin to see every possible outcome, including the rare, chaotic ones.

2. CVaR (The "Worst-Case" Safety Net)

The paper introduces a mathematical concept called Conditional Value-at-Risk (CVaR).

Old Way: "What is the average delay?"
New Way: "If we are in the worst 0.001% of traffic jams, how bad will the delay be?"
The AI is now programmed to ensure that even in these worst-case scenarios, the "race cars" still make it on time. It builds a massive safety buffer, just like a pilot who plans for a storm even if the forecast says it's sunny.

3. Confidence Checking (The "I'm Not Sure" Button)

Sometimes, the AI's crystal ball (the Digital Twin) is blurry because it doesn't have enough data. The new framework forces the AI to admit, "I'm not 100% sure about this prediction."

If the AI is unsure, it automatically assumes the situation is worse than it looks and asks for more resources to be safe.
This prevents the AI from making risky bets based on shaky data.

The Experiment: Trucks vs. Race Cars

The authors tested this in a simulation with two agents:

The eMBB Agent: Represents the "Trucks" (streaming video). They can handle some delay.
The URLLC Agent: Represents the "Race Cars" (critical data). They have a strict 10-millisecond deadline.

The Results:

The Biased Agent (Old Way): It tried to save energy by allocating the bare minimum resources. It looked good on paper, but in 200 trials, it failed 11 times, causing the "race cars" to miss their deadlines. It was a false economy: saving a little energy but risking a crash.
The Unbiased Agent (New Way): It allocated slightly more resources to build a safety buffer.
- Result: Zero failures. The "race cars" always made it on time.
- The Cost: It saved slightly less energy (about 2–15% less than the biased agent), but the authors argue this is a rational trade-off. You pay a little extra for safety, rather than risking a catastrophic failure.

Speed and Feasibility

One might worry that doing all these complex "what-if" simulations would be too slow. The authors proved otherwise. They ran their system on a single, standard graphics card (an NVIDIA RTX A4000) using a small, efficient AI model.

The system made decisions in under 1.5 seconds.
This proves that this "paranoid," ultra-safe planning can happen fast enough to be used in real-time 6G networks without needing slow, expensive cloud computers.

The Takeaway

The paper concludes that for 6G to be truly trustworthy, AI agents must stop being "optimistic optimizers" who only look at averages. Instead, they must become "risk-aware guardians" that plan for the worst-case storms, admit when they are unsure, and prioritize safety over squeezing out every last drop of energy efficiency.

1. Problem Statement

The paper addresses a critical barrier to the deployment of trustworthy 6G autonomous networks (specifically TM Forum Levels 4 and 5): Uncertainty Neglect Bias in Large Language Model (LLM) agents.

The Core Issue: LLM-powered agents managing 6G network slicing often rely on simple statistical averages (mean values) to make high-stakes decisions. This cognitive bias causes them to ignore tail risks (extreme events) and epistemic uncertainty (confidence in their own predictions).
The Consequence: In ultra-reliable low-latency communication (URLLC) scenarios, relying on the mean leads to catastrophic Service Level Agreement (SLA) violations when rare, high-latency spikes occur. The agents fail to account for the "worst-case" scenarios, leading to network instability and service failures.
The Gap: Existing work addresses cognitive biases or risk optimization separately but fails to integrate uncertainty quantification and communication into the agentic reasoning loop, leaving agents "blind" to the reliability of their own data.

2. Methodology

The authors propose an unbiased, risk-aware negotiation framework that shifts agent reasoning from mean-based logic to tail-risk-aware logic. The system involves two competing agents (eMBB and URLLC slices) negotiating for shared bandwidth and edge CPU resources.

A. System Architecture

Digital Twins (DTs): Each agent utilizes a private DT based on queuing theory to simulate network states. The DT runs vectorized Monte Carlo simulations to predict full latency distributions ( $L_{edge} + L_{RAN}$ ) rather than single-point estimates.
Hybrid Decision Engine:
- LLM Component: A locally deployed, 1-billion-parameter model (otel-llm-1b-it) handles strategic goal-setting (generating a single optimal counter-proposal).
- Deterministic Component: A 2D Proportional-Integral-Derivative (PID) algorithm performs fine-grained local search to mathematically guarantee resource feasibility (bandwidth and CPU constraints) and convergence.

B. Risk-Aware Mechanisms

The framework introduces two key mathematical concepts to mitigate bias:

Aleatoric Uncertainty (Statistical Risk):
- Instead of using the mean latency ( $\mu_L$ ), the agent calculates the Conditional Value-at-Risk (CVaR) at a strict confidence level ( $\alpha = 0.99999$ ).
- CVaR represents the expected latency in the worst $(1-\alpha)\%$ of cases, effectively focusing the agent on the "tail" of the distribution.
- A Greedy Guardband is applied, setting a safety target ( $L_{gb}$ ) at 80% of the actual SLA to absorb extreme fluctuations.
Epistemic Uncertainty (Prediction Confidence):
- The agent quantifies its confidence in the DT prediction using the Coefficient of Variation (CV) of the latency distribution.
- Uncertainty-Adjusted Risk ( $\widehat{CVaR}_\alpha$ ): If the DT's confidence is low (high variance), the measured risk is artificially inflated. This forces the agent to treat low-confidence predictions as high-risk, preventing decisions based on unreliable data.

C. Negotiation Protocol

Multi-round Negotiation: Agents engage in a constrained, multi-round dialogue to partition resources.
Validation Loop: Before proposing, an agent validates its proposal against the $\widehat{CVaR}_\alpha$ and the Guardband. If the risk is too high or confidence too low, the agent refuses to optimize resources further, prioritizing safety over energy savings.

3. Key Contributions

Risk-Aware Negotiation Framework: A novel approach that replaces mean-based reasoning with Uncertainty-Adjusted CVaR, enforcing a strict safety buffer against tail events.
Epistemic Uncertainty Propagation: A mechanism where agents explicitly quantify and propagate their own prediction confidence, dynamically penalizing low-confidence data to prevent rigid, unsafe decisions.
Hybrid Edge-Deployable Architecture: A design combining a fast-inference 1B-parameter LLM for strategy with a deterministic PID controller for feasibility, ensuring sub-second execution suitable for non-real-time RIC (RAN Intelligent Controller) use cases.
Empirical Validation: Extensive simulation demonstrating that the proposed method eliminates SLA violations that plague mean-based approaches, while transparently exposing the trade-off between reliability and energy efficiency.

4. Experimental Results

The framework was validated over 200 independent negotiation trials between an eMBB agent (50ms SLA) and a URLLC agent (10ms SLA).

SLA Compliance:
- Biased (Mean-based) Agent: Systematically failed, causing 11 critical SLA violations for the URLLC slice. Its latency distribution extended beyond the 10ms limit.
- Unbiased (Risk-Aware) Agent: Achieved 0 SLA violations. It successfully kept the entire latency distribution below the 10ms threshold.
Tail Latency Reduction:
- The risk-aware agent reduced the 99.999th-percentile (p99.999) latency by up to 51.7% for eMBB and significantly for URLLC (reducing from 7.4ms to 4.4ms).
Energy vs. Reliability Trade-off:
- The biased agent achieved higher energy savings (26–30%) but at the cost of safety (operating dangerously close to failure boundaries).
- The risk-aware agent achieved lower energy savings (2–15%), which the authors argue is the rational cost of reliability, avoiding the "false economy" of the biased approach.
Inference Performance:
- Executing the framework on a single NVIDIA RTX A4000 GPU with the 1B-parameter model resulted in sub-1.5-second inference times (mean ~1.29s for the risk-aware agent), validating its feasibility for edge deployment.

5. Significance

This paper provides a foundational step toward trustworthy 6G autonomy. It demonstrates that:

Cognitive biases in AI are not just theoretical: They have quantifiable, catastrophic impacts on network performance (SLA violations).
Mean-based optimization is insufficient for 6G: Critical services like URLLC require reasoning over the tail of distributions, not the average.
Uncertainty must be communicated: Agents must not only predict outcomes but also quantify their confidence in those predictions to prevent cascading failures in multi-agent systems.
Feasibility: Advanced risk reasoning can be performed locally on edge hardware without relying on slow, privacy-compromising cloud APIs, making robust autonomous networks a practical reality.