acceptodds
Under review as a conference paper at ICLR 2027

Beyond Forecasting: World Models for Coupled Radio Network Control

Abstract

Model-based RAN control requires predicting how future resource allocations reshape network dynamics, including effects that persist after an intervention has passed. We propose TRACTOR+, an action-conditioned world model that separates recurrent network-state evolution from persistent action-response dynamics. TRACTOR+ combines multi-scale recurrent states, an instantaneous action-conditioned predictor, and adaptive fast/medium/slow response memories for free-running multi-horizon rollout. Experiments on three-slice RAN traces show that TRACTOR+ achieves the lowest standardized RMSE across all evaluated natural and allocation-event horizons with only 0.193M parameters. Removing future-action inputs degrades event prediction, while mechanism ablations show that persistent-response memory strengthens action-path discrimination. The learned dynamics also support short-horizon offline planning within the evaluated Near-RT budget. All alternative-action results are interpreted as predictive diagnostics under observational support rather than causal or closed-loop effects.

Primary area:
learning on time series and dynamical systems
Keywords:
World Models, Radio Access Networks, Network Dynamics, Resource Allocation, Model Based Control, Multi Slice Networks, Counterfactual Reasoning, Sequential Decision Making

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LFpVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRACTOR+, an action-conditioned world model designed for resource allocation in radio access networks (RANs). The core idea is to decouple the modeling of recurrent network-state evolution from the persistent responses triggered by resource adjustments. By employing multi-scale recurrent states, an instantaneous action-conditioned branch, and adaptive persistent-response memories, the model enables multi-step autoregressive prediction and supports offline model predictive control (MPC). Experimental results show that TRACTOR+ achieves lower prediction errors on both natural dynamics and strict allocation-event windows than the evaluated baselines, while using 0.193M parameters and maintaining planning latency within the Near-RT budget.

Critical strengths
  1. The persistent-response memory provides a targeted inductive bias. By explicitly maintaining the residual effects of past actions, the model avoids relying solely on standard recurrent states to implicitly encode this information. This design is computationally efficient and well-suited for short-horizon predictive planning.
  2. The empirical results are compelling. TRACTOR+ achieves the lowest RMSE across all evaluated horizons, performing particularly well during resource-adjustment events. Ablation studies—specifically those involving the removal of future-action inputs or the persistent-response memory—demonstrate that the model’s performance is not merely a result of standard time-series extrapolation.
Critical weaknesses
  1. The separation between instantaneous and persistent responses is not explicitly enforced.. Since the recurrent state already incorporates action history, while persistent memory is updated by the current action and contributes to the same-step prediction, the two pathways may encode overlapping effects.. Progressive training may encourage this separation, but it does not ensure that the two components learn distinct responses. Consequently, the learned decomposition may partly reflect overlapping representations rather than distinct response mechanisms.

  2. The planning algorithm does not explicitly account for model uncertainty. Although the planner requires feasible, TRAIN-supported transitions, it selects actions based on the world model's predictions without explicitly penalizing model uncertainty. At H = 4, the reported exact sequence-level training support is only 9.1% on strict-event anchors, compared with 26.6% on natural anchors. In such scenarios, longer planning horizons may amplify model errors, potentially leading the model to favor action sequences with overestimated predicted gains. Consequently, prediction accuracy and action sensitivity alone do not establish the reliability of the selected plans.

Questions for authors
  1. Since the instantaneous branch uses an action-conditioned recurrent state and the persistent branch also depends on the current action, how do the authors ensure that they do not encode overlapping responses? Is there an explicit constraint or evidence that the two pathways learn distinct effects?
  2. Do the three memory components possess distinct, effective decay time scales? Without explicit constraints, how can the authors separate the benefits of multiple timescales from those of additional model capacity?
  3. How does the planner limit the exploitation of world model prediction errors when full action sequences have limited exact training support? Would the planning results still hold if an uncertainty penalty were introduced, or if planning were restricted to full action sequences observed in the training data?
Reviewer_edi3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a specialize method namely TRACTOR+ for PRB-budget allocation, using world models (transformer-based) and by integrating three level times steps (slow, medium, fast) to predict the effect of action-dependent dynamic on the future time steps in a heterogeneous, challenging RAN environment. These types of environments can have two types of dynamics, those related to the action and those independent of actions. The paper suggests an embedding structure-based approach by modifying a given world model (that needs to be explained in the current paper better) to predict the state/environment dynamic within future time steps. These features contain action representation, previous state predictions and the current state. In this way, the predicted state is dependent on the current available allocation as well, and they help the system to better allocate resources by predicting the future situation of the RAN environment.

Critical strengths
  • One interesting motivation of this paper is its application, that is resource allocation for Radio Access Network (RAN). Such studies are necessary for problems with many uncertainties and unpredictable environments such as optimizing different objectives in RAN.
  • The paper compares its method against an exhaustive list of available baselines, mostly available in the domain literature.
  • The analytical part of the paper on effect of roll-out horizon on the model performance in motivating too.
Critical weaknesses
  • The methodological novelty of TRACTOR+ can be better presented. Currently the proposed architecture merges the three time-scales in an embedding space with two types of actions (persistent and instantaneous). In order to better present this contribution and not to consider it just as a modification in the architecture, a better discussion of the methodology in compare to the available literature is required. This also necessitates a better discussion of the used approach/architecture behind it.
  • The paper would benefit from a more explicit mathematical formulation of the underlying world-model learning problem, including the assumptions governing action-conditioned network dynamics. Although the authors provide several training details, the description of the architecture and experimental protocol could be made more comprehensive in order to simplify the reproduction of the work.
  • The performance improvements of TRACTOR+ over the strongest competing methods are relatively modest, particularly on natural dynamics, where the reported RMSE reductions range from approximately 1.4% to 3.9%. Although the evaluation includes 16093 natural test windows, the strict allocation-event analysis relies on only 91 windows from 20 trials. Given that persistent action responses constitute the main motivation for the proposed architecture, a broader evaluation of these events would strengthen the experimental evidence. In addition, reporting variability across independent training runs would help establish the robustness of the improvements.
  • The authors highlight the relatively small size of TRACTOR+ (0.193M parameters). However, parameter count alone does not establish computational efficiency, particularly for applications with stringent latency requirements such as Near-RT RAN control. A more systematic comparison of inference latency, computational cost, and memory requirements against competing world models would strengthen this aspect of the evaluation.
Questions for authors

see above

Reviewer_mQHUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed TRACTOR+, which is an action-conditioned world model that separates recurrent network-state evolution from persistent action-response dynamics. The proposed method, TRACTOR+, is evaluated through multi-horizon prediction, future-action sensitivity, offline predictive planning, and mechanism ablations, demonstrating improved predictive accuracy and action-path discrimination around allocation events.

Critical strengths
  • The paper studies the timing problem in next generation radio access networks. It addresses the resource-allocation changes beyond ordinary traffic forecasting. +Some of the performance results seem interesting, such as achieving the lowest RMSE across all evaluated horizons, as authors claim. This can improve the resource allocation by 3.6-6.8% compared to based line methos.
  • The evaluation uses different protocols, such as disjoint trials, training-only normalization, action-conditioning diagnostics.
Critical weaknesses
  • The introduction does not well present the O-RAN architecture. For example, the term “Near-RT controller” is introduced without a definition or citation.
  • The evaluation in dataset is limited. The authors evaluate on an offline three-slice RAN dataset sampled every 250 ms. The strict allocation event contains only 91 windows from 20 trials and isolates the transient and persistent dynamics following a clearly localized resource-allocation event.
Questions for authors
  1. Some parameters and abbreviations are not defined. Such as in such as in the parameters “delta u, PRB, WM by 3.6%, H = 2, 4, and 8”.
  2. In section 3, how are the fast, medium, and slow recurrent states updated, and what loss function is used during each tanning stage?
  3. What is the mathematical model for the throughput-buffer planning objective?
  4. How are “TRAIN-supported” paths defined?
  5. The computational complexity of the proposed method is not analyzed. How do the time require with number of slices and planning horizons?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.