UE‑Based AI in 3GPP: The Next Frontier for Network Performance Management

AI/ML has been creeping into mobile networks for years, but 3GPP’s recent work marks a decisive shift: intelligence is no longer confined to the network. User Equipment (UE) is becoming an active participant in performance optimisation, prediction, and context‑aware decision‑making. With Release 18 and the ongoing Release 19 work, 3GPP is formalising how UE‑based AI models are managed, monitored, and integrated into end‑to‑end network performance management (initial focus is in RAN work groups).

Why UE‑Based AI Matters

UEs have always been the richest source of real‑world performance data. Embedding AI models directly on devices unlocks new capabilities:

• Local inference for mobility prediction, beam selection, and channel estimation

• Context‑aware optimisation using sensor fusion and application‑level insights

• Reduced signalling overhead by enabling smarter reporting rather than raw measurement dumping

• Personalised QoE optimisation, especially for XR, gaming, and high‑mobility scenarios

But these benefits only scale if operators can trust, manage, and govern UE‑side models — which is exactly where 3GPP standardisation is now focused.

What 3GPP Is Standardising (and What It Isn’t)

3GPP has been explicit: AI models themselves are not standardised. Instead, the standards define the ecosystem around them — the control points, data flows, and lifecycle hooks that allow operators to deploy and manage UE‑based intelligence safely and consistently.

Key areas include:

1. Model Lifecycle Management (SA5)

SA5 has developed a domain‑agnostic AI/ML lifecycle management framework covering training, validation, deployment, activation, deactivation, and monitoring 3GPP.

For UE‑based AI, this means:

• Operators can activate or deactivate UE‑resident models

• Models can be swapped or updated via standardised mechanisms

• Behaviour can be monitored for consistency and testability

This is essential for ensuring predictable performance across diverse device vendors.

2. Data Collection and Model Delivery

3GPP is defining mechanisms for:

• Standardised data collection to support offline training

• Model delivery from network to UE (e.g., improved models, fine‑tuning packages)

• New UE measurements that enable AI‑driven inference

This ensures that UE‑based AI is not a black box but a managed component of the network.

3. Air Interface Enhancements (RAN)

RAN groups are introducing:

• New measurement types

• Extensions to CSI reporting

• Support for AI‑assisted beamforming, mobility, and scheduling

These enhancements allow UE‑side models to operate with richer, more precise inputs.

4. Performance Management Integration

For operators, the real value comes when UE‑based AI feeds into network performance management systems:

• Predictive mobility and handover optimisation

• Early detection of QoE degradation

• Context‑aware anomaly detection

• Adaptive configuration based on UE‑side inference

SA5’s management framework ensures these insights can be consumed consistently across vendors and domains.

Challenges 3GPP Is Addressing

UE‑based AI introduces new risks and complexities:

• Testability and consistent behaviour across devices

• Energy consumption constraints on UE

• Data ownership and privacy boundaries

• Feasibility of online training (still considered future‑looking)

3GPP’s work aims to create a controlled environment where UE intelligence enhances — rather than destabilises — network performance.

What This Means for Operators

UE‑based AI is moving from vendor‑specific innovation to a standards‑aligned, operator‑governed capability.

This unlocks:

• Cross‑vendor consistency

• Safer deployment of on‑device intelligence

• Better integration with OAM and performance management systems

• A path toward distributed, collaborative AI across RAN, Core, and UE

As Release 19 progresses, UE‑based AI will become a foundational element of 5G‑Advanced performance optimisation — and a stepping stone to 6G’s fully distributed intelligence.