RAN World Model

RAN World Model

2 November 2026

Introduction

Physical AI will depend on the RAN for low-latency connectivity and sensing. That same RAN must allocate radio, compute, and transmit power among traffic classes with different deadlines, a sequential decision problem resembling robot control. An observed state is followed by a control action, and the world, here mixed user demand and inter-cell interference, responds.

A world model predicts the next state from the current state and an action, and can be rolled forward under a planned sequence, so a planner can score candidate actions in imagination rather than by spending live traffic on each one. The research question behind this demonstration is whether a learned world model of that kind can forecast a small RAN under a candidate resource policy well enough for receding-horizon control.

RAN control is a useful test of the idea because capable reactive heuristics already exist, commands take some time to take effect, measurements arrive often, and the objective is an explicit cost on late work, drops, energy, and cloud use. These properties can all be explored in the interactive demonstration player.

The control task

A RAN is the segment of a mobile network between user devices and the wired core. This study uses three simulated cells and three slices, or traffic classes. XR video frames have a 30 ms deadline, voice packets 50 ms, and AI compute jobs 400 ms. After uplink, an AI job may run on a graphics processor at the cell or on a shared cloud processor. Every tenth of a second the controller may change the radio split among slices, the fraction of AI jobs sent to the cloud, and a transmit-power offset down to 10 dB below full power, and those commands take three seconds to take effect.

The world-model action is a seven-dimensional policy parameter θ. Its coordinates are slice weights for XR, AI, and voice, a radio headroom relative to measured arrivals, an offload bias, a low-power floor, and the utilization knee of a power ramp. A fixed feedback rule maps θ and the 100 ms measurements to radio, offload, and power commands, so planning over θ at a one-second step still expresses reactive behaviour within each second. Each slice first receives a guaranteed radio share equal to the headroom coordinate times its measured arrival rate divided by the cell's spectral capacity. If those guarantees fill the carrier they are renormalized. Otherwise leftover resource is split in proportion to exp(w_s) times remaining queue, where w_s is the slice weight in θ. The AI offload fraction is a sigmoid of the offload-bias coordinate plus the gap between edge wait and cloud wait after two cloud round-trip times. Transmit power ramps from the low-power floor to full power as cell utilization approaches the knee. The tuned baseline uses this same rule with θ held fixed.

Let c index cells and s index slices. Slice weights in the cost are w_XR = 0.6, w_AI = 0.25, and w_voice = 0.15. Nominal offered load ν_s is 160, 7, and 150 jobs per second per cell for XR, AI, and voice. There are N = 3 cells. In each tenth of a second, ℓ_c,s and d_c,s are the late and dropped job counts, E_c is radiated energy, and n_c^cloud is the number of AI jobs completed in the cloud. The energy and cloud weights are λ_E = 0.1 and λ_cloud = 0.002, and the inner step is Δt = 0.1 s. The stage cost is

C_stage = ∑_c,s (w_s / (ν_s N)) (ℓ_c,s + d_c,s) + λ_E (Δt / N) ∑_c E_c + λ_cloud ∑_c n_c^cloud.

Jobs still queued at the end of the five-second horizon are charged in proportion to remaining drain time divided by that drain time plus the slice deadline, so a growing backlog affects the cost even before it becomes late. Predicted observations are inserted into this same cost.

The three-second delay of commands is matched to a three-second notice of scheduled XR traffic. A rule that only reacts to queues therefore starts moving after the traffic has arrived, whereas a planner that utilizes a forecast can change θ pre-emptively.

Models

Both models consume the last four seconds of observations and the policy parameters θ that produced them, and output expected observations for the next five seconds. That observation vector is the model interface. A one-second step is 41 numbers per cell and two cloud numbers, 125 in all. Ten of the cell fields are the effective and pending resource commands. They condition the model and are never prediction targets. The remaining 31 cell fields and the two cloud fields are the targets, 95 numbers per second.

The model does not emit a point estimate per field. It emits the parameters of a probability distribution, and training maximizes the probability of the true next observation under that distribution, equivalently minimizing negative log-likelihood. Integer event counts such as arrivals, on-time jobs, late jobs, drops, and cloud completions are nonnegative and overdispersed, so those channels use a negative binomial. Queue depth, head-of-line delay, spectral efficiency, utilization, and energy are continuous stocks, so those channels use a Gaussian on a log or linear transform of the field. This is the model's observation likelihood. It is not the generative mechanism of the simulator.

Each family is trained five times from different random initializations. The five members form an ensemble. When a candidate plan is scored, the planner uses the mean predicted cost across members plus one standard deviation of that cost, so disagreement raises the score of an uncertain plan. Width is 128, with four attention heads and three layers. The rest of this text refers to each family as a model.

The observation-space model is an entity transformer. Each cell is one token, built from its four-second history of all 41 fields, and a fourth token holds the cloud history. Cells share weights and carry no index embedding, so a permutation of the cells permutes the predictions. The last executed values of θ and the five-second plan are encoded into a condition vector. Each transformer layer first layer-normalizes its token features. An affine map of the condition then produces a scale, a shift, and a residual gate for that layer, and the normalized features are scaled and shifted before attention and the MLP. In this way the imagined policy can change the computation at every depth rather than being added once as an extra token. This is adaptive layer normalization. All five future seconds are predicted in one shot, so predictions are never fed back as inputs.

The latent model follows JEPA as used in LeWorld Model, with a decoder added so that the RAN cost can be evaluated. An encoder sees only the 31 target fields of each cell and the two cloud fields. Four seconds of those channels are 380 numbers, which the encoder pools through a CLS token and cell-index embeddings into a single vector z ∈ R¹²⁸. A predictor then takes residual steps modifying that vector. If z_k is the latent k seconds into the horizon, k ∈ {0, 1, 2, 3, 4}, and f is a residual network conditioned on the planned θ, the step is z_k+1 = z_k + f(z_k, θ). A decoder maps each z_k back to the 95 target channels.

Training of this model has two likelihood terms in addition to the JEPA losses. The predictive term is the likelihood of the five future seconds after the latent has been rolled forward under the plan and decoded. The reconstructive term encodes the current four-second window to z₀, decodes z₀ immediately, and scores that reconstruction against the current observation, so the latent stays aligned with the named metrics rather than becoming an arbitrary code. A JEPA mean-squared error between predicted and encoded latents, and a SIGReg penalty that discourages latent collapse, complete the objective.

That latent path is the architecture that is documented to work for pixel input in robotics tasks. A camera frame of a few hundred pixels on a side contains on the order of 10⁵ intensities, most of them appearance, so a JEPA encoder can ignore the detail and keep what is relevant for planning. The RAN observation is already a compact, named summary. Late counts, queue depths, energy, and cloud occupancy are a large fraction of the 95 targets, and they are exactly the coordinates of the cost. Compressing 380 history numbers into 128 dimensions, then asking a decoder to restore those coordinates, is therefore a bottleneck on the quantities the planner must rank. The pooled latent also has to remember which cell is which, so the encoder uses index embeddings and is no longer equivariant to a permutation of the sites. Error in z is stepped five times before decoding, which compounds for example a miss on a burst.

Training used two thousand simulated episodes. Each episode is five unscored warmup seconds under a joint heuristic, then 120 scored seconds. Controllers during collection mixed noisy reactive behaviour with random θ. After that first fit, the observation-space CEM controller was run on new seeds, the recorded trajectories were added to the training set, and the model was trained again. That extra collect-and-retrain cycle is one round of Dataset Aggregation (DAgger), which reduces the mismatch between training states and the states a learned controller later visits. The observation-space checkpoint in the player includes that round. The latent checkpoint is the first training run on the original mix.

In every scene a CEM planner samples θ in the feasible box, scores each sample with the chosen model, and applies the cheapest one for the next second. A plan consists of two values, one for the coming second and one for the remaining four. The score is the mean predicted cost across the five members plus one standard deviation. The previous plan is kept as a sample so the controller can reuse it.

The models are evaluated against a tuned baseline, which is the same feedback rule described above, with θ held fixed at values tuned on the training mix. Every tenth of a second it reads the current radio and queue measurements and produces slice, offload, and power commands by that rule. It has no receding-horizon search and produces no forecast.

Demonstration

The simulator has three sites 120 degrees apart on a 600 m plane, sharing a 20 MHz carrier at 3.5 GHz. Users move at walking speed and attach to the strongest site. A typical moment has approximately sixty simultaneous users, mostly voice and AI, and load in each cell rises and falls over tens of seconds. On top of this, each episode schedules a few XR groups on one cell, announced three seconds before they start. Neighbouring cells interfere in proportion to how busy they are, and cloud GPU jobs share a processor with other tenants.

The player holds six recorded episodes, the same three loads driven by each model. The first, nominal, is ordinary mixed load. The second, extra XR, adds five unannounced XR sessions on one cell at 53 s, lasting 25 s, on that same seed. The third, heavy cloud, doubles background tenants on the shared processor. After four seconds of history, dashed overlays show the five-second forecast of queued XR, AI, and voice traffic and of late jobs, while solid lines are the simulator under the θ that CEM selected.

Results

Quiet windows appear to be easy for the models. Voice queues stay below about 2 Mbit in every tape, and both models track them. XR and AI peaks are not tracked that accurately. Around t = 30 s every scene has a scheduled XR group of eight sessions. In nominal and extra XR the simulator XR queue reaches about 7.5 Mbit, yet the observation-space forecast peaks near 1.7 Mbit and the latent forecast near 1.6 Mbit. Late jobs in that window climb to about 1 200 per second.

Extra XR after the unannounced sessions at 53 s is the most difficult window in the demo. True AI queues reach about 38 Mbit, compared with an observation-space mean near 16 Mbit and a latent mean near 8 Mbit. CEM in both families increases the XR weight and raises headroom, causing AI traffic to be underserved.

Heavy cloud shows a thick outer ring of other tenants at the cloud GPU. Both models keep edge-to-cloud compute offload strongly negative and, when edge AI piles up, raise the AI slice weight. True edge AI queues again exceed 20 Mbit, and neither forecast follows that rise. The latent-space model also cut transmit power more aggressively, with a mean floor of about −3.8 dB against −1.0 dB for observation-space on this load.

Cost on these three recorded CEM scenes, each one seed, is compared with the tuned baseline.

On nominal, observation-space CEM is slightly better than the baseline, with similar XR on-time (0.74) and slightly better AI and voice, while the latent controller is slightly worse with XR on-time falling to 0.71. On extra XR both world models lose, mainly on XR, with on-time 0.75 and 0.72 against 0.80 for the baseline. The XR burst is not announced, so every controller sees it only through queues, and the three-second delay is long relative to the burst. The gap is largest on heavy cloud. The baseline keeps a negative AI weight and protects XR (on-time 0.89). Both world models, seeing a crowded cloud, raise the AI weight and withhold offload, which steals radio from XR (on-time 0.79 and 0.76). The latent model is the worse of the two on every scene.

Discussion

The demonstration shows that a world model planner can control a small simulated RAN. The models forecast the metrics the cost uses, and a planner can act on those forecasts. On nominal, observation-space CEM is 2.2% cheaper than the reactive baseline, while the latent controller is 1.2% more expensive. Both models are more expensive on extra XR and heavy cloud, and the latent model is the worse of the two on every scene. The models here forecast the magnitudes of XR and AI load peaks poorly, so the horizon is spent on a policy whose predicted cost does not match the simulation. A latent of 128 dimensions is a natural compression of a camera image, but causes an information bottleneck on a RAN state vector whose named coordinates already are the cost.

A trustworthy forecast would have an advantage over a reactive heuristic by allowing to change θ when the load mix shifts, while keeping the same 100 ms rule. The baseline holds one θ after a single tuning pass on the training data. A planner that scored candidates as the simulator scores them could pick a new θ without another offline search and without label information of the scene. On heavy cloud both current model searches do pick a new θ, but they raise the AI weight and steal radio from XR. The correct ranking of candidate policies remains a viable target for future research.