Paper deep dive
RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models
Ahmed Aboulfotouh, Akram Bin Sediq, Koosha Pourtahmasi Roshandeh, Omar Mashaal, Ahmad M. Nagib, Jale Sadreddini, Hatem Abou-Zeid
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/6/2026, 11:13:48 AM
Summary
The paper introduces RRM-GPT, an autoregressive framework for Radio Resource Management (RRM) foundation models that generate RRM decisions (such as scheduling grants, handovers, and interference coordination) as sequences of tokens, similar to how language models generate text. The framework uses an encoder-decoder architecture where an encoder maps heterogeneous network observations into a common token representation, and a decoder generates decisions field-by-field conditioned on the network state and previously generated fields. Pretraining on unannotated network logs teaches the model validity rules and controller behavior, while post-training adapts the model to specific operator objectives via imitation or reinforcement learning. A case study in 5G New Radio demonstrates that a single model can generate complete scheduling grants, capturing dependencies among fields and transferring learned behavior to unseen scenarios.
Entities (8)
Relation Signals (7)
RRM-GPT → uses → Autoregressive Decoder
confidence 95% · We propose RRM-GPT, an autoregressive framework... a decoder emits the decision one field at a time
RRM-GPT → uses → State Encoder
confidence 95% · An encoder maps heterogeneous network observations into a common token representation
RRM-GPT → generates → Scheduling Grant
confidence 92% · The model captures dependencies among grant fields... generates complete scheduling grants
RRM-GPT → appliesto → 5G New Radio
confidence 90% · In a 5G New Radio (NR) case study, we demonstrate that a single model generates complete scheduling grants
RRM-GPT → trainedwith → Imitation Learning
confidence 85% · post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning.
RRM-GPT → trainedwith → Reinforcement Learning
confidence 85% · post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning.
State Encoder → processes → Radio Resource Management
confidence 80% · An encoder maps heterogeneous network observations into a common token representation... for radio resource management
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Learning-based models for radio resource management (RRM) are typically built for a single function and deployment, so each new setting repeats the development pipeline. RRM decisions, however, share a common structure: each is assembled from interdependent fields, defined by the standard, whose values are selected in view of the network state. We propose RRM-GPT, an autoregressive framework for RRM foundation models that generate these decisions as a language model generates text. An encoder maps heterogeneous network observations into a common token representation, and a decoder emits the decision one field at a time, each conditioned on the network state and the fields already committed. Pretraining on unannotated network logs teaches the model what makes a decision valid and how controllers choose among valid decisions; post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning. The framework targets two forms of reuse: a function-specific model reused across deployments, and a model shared across RRM functions that generates their interdependent decisions as one sequence. In a 5G New Radio (NR) case study, we demonstrate that a single model generates complete scheduling grants spanning user selection, timing, link adaptation, resource allocation, and control signaling. The model captures dependencies among grant fields and transfers learned behavior to an unseen scenario without adaptation.
Tags
Links
- Source: https://arxiv.org/abs/2610.04296v1
- Canonical: https://arxiv.org/abs/2610.04296v1
Trouble viewing inline? Open PDF directly →
Full Text
41,078 characters extracted from source content.
Expand or collapse full text
1 RRM-GPT: A Framework and Vision for Radio Resource Management Foundation Models Ahmed Aboulfotouh ∗ , Akram Bin Sediq † , Koosha Pourtahmasi Roshandeh † , Omar Mashaal ∗ , Ahmad M. Nagib † , Jale Sadreddini † , and Hatem Abou-Zeid ∗ ∗ Department of Electrical and Software Engineering, University of Calgary, Canada † Ericsson Canada Inc., Canada Abstract—Learning-based models for radio resource manage- ment (RRM) are typically built for a single function and deploy- ment, so each new setting repeats the development pipeline. RRM decisions, however, share a common structure: each is assembled from interdependent fields, defined by the standard, whose values are selected in view of the network state. We propose RRM- GPT, an autoregressive framework for RRM foundation models that generate these decisions as a language model generates text. An encoder maps heterogeneous network observations into a common token representation, and a decoder emits the decision one field at a time, each conditioned on the network state and the fields already committed. Pretraining on unannotated network logs teaches the model what makes a decision valid and how con- trollers choose among valid decisions; post-training then adapts it to deployment-specific operator objectives through imitation or reinforcement learning. The framework targets two forms of reuse: a function-specific model reused across deployments, and a model shared across RRM functions that generates their interdependent decisions as one sequence. In a 5G New Radio (NR) case study, we demonstrate that a single model generates complete scheduling grants spanning user selection, timing, link adaptation, resource allocation, and control signaling. The model captures dependencies among grant fields and transfers learned behavior to an unseen scenario without adaptation. Index Terms—Radio resource management, autoregressive modeling, wireless foundation models, AI-native RAN, 5G NR scheduling. INTRODUCTION Every fraction of a millisecond, a cell decides which users to serve, on which spectrum resources, and at what rate. At slower timescales, neighboring cells coordinate spectrum use, users move between cells or carriers, and traffic is redistributed to balance load. Together, these decisions constitute radio resource management (RRM) [1]. Cellular systems organize them into specialized functions for scheduling, link adaptation, interference coordination, mobility, and load management, each with its inputs, constraints, algorithms, and timescale [2]. With growing interest in AI-native resource management, learning-based methods have been applied successfully to individual RRM functions, such as scheduling [3] and link adaptation [4]. Such models are typically specialized to a par- ticular RRM function and deployment scenario, with inputs, outputs, and training objectives tailored to that setting [5], [6]. Developing a model for a different function or deployment requires revisiting the engineering pipeline: assembling repre- sentative data, designing the objective, training, and validating the model. This specialization limits model reuse and increases the engineering effort required to accommodate changes in network configuration, traffic conditions, and operator ob- jectives. Developing models independently for each function also makes it harder to capture dependencies among RRM decisions within a shared context. To address these limitations, we propose an autoregres- sive framework, illustrated in Fig. 1, for building RRM foundation models: models pretrained on network logs for reuse across network deployments and RRM functions. The figure distinguishes two scopes of reuse. In the function- specific setting, a model supports one RRM function across deployments; a scheduling foundation model, for example, could be adapted to different cell configurations, environments, and traffic conditions. The multi-function setting represents a broader ambition: sharing a model across RRM functions so that their decisions can be generated within a common context. Autoregressive Transformer models have been explored for generating scheduling decisions, with a dedicated model for each scheduling subtask [7]. Prior work on wireless foundation models has primarily explored representation learning from physical-layer signals and channels [8]–[10]. The framework presented here provides a common formulation for structured RRM decision generation, with model reuse across deploy- ments and functions as a central objective. The framework builds on a structural observation: despite their different roles, RRM functions construct decisions in a common form. Each decision is assembled from fields whose values are selected in view of network observations, subject to standards, configuration, and available resources. As illustrated on the right of Fig. 1, a scheduling grant identifies a user, transmission parameters, and radio resources; a handover iden- tifies a user and a target cell; and an interference-coordination decision specifies participating cells, affected time-frequency resources, and restrictions such as muting. The fields within a decision are interdependent. In a scheduling grant, channel conditions constrain the rate, the selected rate and queued data determine the resources required, and existing allocations constrain their placement. Dependen- cies also cross functions: a handover determines which cell subsequently schedules a user, while interference coordination restricts the resources available to affected schedulers. RRM decisions thus exhibit sequential structure at two levels: among fields within a decision, and among decisions across functions. Our autoregressive framework, RRM-GPT, generates a de- cision one field at a time, each choice conditioned on the network state and on the fields already committed. Training proceeds in two stages: reusable pretraining and objective- arXiv:2610.04296v1 [eess.SP] 3 Oct 2026 2 EVOLVING NETWORK Cell 1Cell 2Cell 3 TRAFFIC LOAD C1 C2 C3 SPECTRUM USE CONTESTED REGION C1 C2 C3 RRM FOUNDATION MODEL Autoregressive framework LATEST NETWORK STATE ConfigurationMeasurements Resource mapsActive controls STATE ENCODER AUTOREGRESSIVE DECODER FIELD 1FIELD 2...End FEASIBLE, STANDARD-COMPLIANT DECISION PRETRAINING Learn reusable patterns from network logs POST-TRAINING Adapt to target deployment and objectives RRM DECISION GENERATION FUNCTION-SPECIFIC MODELS One model per function, reused across deployments SCHEDULING UserMCSResourcesEnd HANDOVER UserTarget cellEnd INTERFERENCE COORDINATION CellsResourcesRestrictionEnd LOAD MANAGEMENT SourceTargetTraffic shareEnd MULTI-FUNCTION MODEL Shared weights across functions and deployments EXAMPLE DECISION SEQUENCE Handover Interference coordination Scheduling ... End APPLY DECISIONS Execution updates the network; new observations form the next state. Fig. 1: Overview of the proposed autoregressive framework for RRM. Network state is encoded, and feasible, standard- compliant decisions are generated field by field. The formulation supports function-specific models with separate weights and, at its broadest scope, a shared model across functions. Models are pretrained on network logs and post-trained for a deployment objective. specific post-training. Pretraining uses state-action records collected from operational networks and simulators without manual annotation, from which the model learns how a deci- sion is assembled and how it follows from network conditions. Post-training adapts the model to a target deployment, either by imitating a reference controller or through reinforcement learning to optimize an operator-defined objective expressed in terms of network key performance indicators (KPIs), such as throughput and latency. Our contributions in this article are: • We propose an autoregressive framework for generating feasible RRM decisions, with pretraining on network logs and post-training for deployment-specific objectives. • We introduce a representation of network observations and RRM decisions that accommodates heterogeneous inputs and function-specific fields and constraints. • We outline how the framework can be applied to mobility, interference coordination, and load management, and discuss the prospects for model reuse across deployments and functions. • We demonstrate complete scheduling-grant generation in a 5G New Radio (NR) case study and evaluate general- ization to an unseen scenario. SCHEDULING AS SEQUENCE GENERATION Consider downlink (DL) scheduling in a 5G NR cell, where time is divided into slots. In each slot, the base station decides which users to serve, how much spectrum each receives, and how its data is encoded. A user cannot know in advance what it has been assigned, so the base station must also send a control message carrying this information and place it where that user is configured to look. A scheduling decision therefore allocates two related kinds of resources: those that carry the data and those that carry the message describing how to receive it [11]. The base station makes its choices based on what it ob- serves: how much data each user has waiting and for how long, the channel quality each user reports, and the outcomes of recent transmissions. Consider the example in Fig. 2, where the scheduler selects User 7, whose data queue is long. The reported channel quality indicator (CQI) guides the choice of modulation and coding scheme (MCS) 18, which sets how many bits each unit of spectrum carries. That rate and the waiting data determine how much spectrum is needed, twelve physical resource blocks (PRBs), the units in which frequency is assigned. In this example, we assume resource allocation Type 1, which requires allocating consecutive PRBs. The scheduler assigns twelve PRBs one slot ahead, starting at PRB 24. The same channel report shapes the control message: a weaker channel needs more control resources, so it is spread over four control channel elements (CCEs), an aggregation level of 4, beginning at CCE 8, on one of the candidate positions User 7 monitors according to the 3GPP specifications. Together, these choices form one DL grant, and a slot may contain none, one, or several, so how many to issue is itself part of the decision. Because each choice is shaped by those before it, scheduling can be viewed as a sequence of interdependent decisions. FROM SENTENCES TO GRANTS Although most fields within a grant have only a limited set of possible values, the number of possible grants grows quickly when they are combined. For a single user, choosing the MCS, the slot, the frequency allocation, the control- 3 (a) LANGUAGE MODELING The growing text prefix is the decoder's only conditioning input (b) RRM DECISION GENERATION A growing field prefix and a fixed encoded state condition the decoder TEXT PREFIX NEXT WORD AUTOREGRESSIVE DECODER DECISION PREFIX NEXT FIELD AUTOREGRESSIVE DECODER Thelinkisreliable linkisreliableEnd Start User 7 MCS 18 PRB count 12 Start PRB 24 User 7 MCS 18 PRB count 12 Start PRB 24 End NETWORK STATE UE BUFFER UE CHANNEL OCCUPANCY UE HARQ HISTORY STATE ENCODER CONTEXT TOKENS Fixed for this decision CELLOCCUPANCY UE CHANNELUE BUFFER cross- attention AppendAppend Fig. 2: Autoregressive generation in language and in RRM. A language model emits one word at a time, each conditioned on the words already written (a). A scheduling decision is generated the same way (b), except that the decoder also conditions on the network state, encoded once per decision, as it emits the user, MCS, PRB count, and starting PRB in turn. channel settings, and the remaining fields already yields sev- eral million possible grants. A slot may contain several grants, multiplying the possibilities further. Not every combination is valid, either: each grant consumes resources and changes what can be issued next. Representing the entire scheduling decision as one output would therefore require choosing from an enormous set, whose size is not known in advance and whose valid options change as the slot fills. How a language model writes. Language generation also involves an effectively unbounded set of possible outputs. Rather than selecting a complete sentence from a catalog, an autoregressive model generates it one word at a time, each drawn from a manageable vocabulary and conditioned on the words already written, as shown in Fig. 2a. Pretraining on large text corpora teaches the model grammatical and contextual patterns that guide its choices: after “the base station allocates,” for example, words describing resources become plausible continuations. Neither grammatical rules nor likely continuations are explicitly programmed; the model learns both by repeatedly predicting the next word. Generating the grant. A scheduling grant can likewise be generated as a sequence, as illustrated in Fig. 2b. The standard defines the grant’s fields and their allowed values, much as grammar defines sentence structure. The scheduler’s policy selects among valid grants, only a small fraction of which a well-designed scheduler would issue. We generate each grant one field at a time, following a chosen order: the user, the slot, the MCS, the PRB count, the starting PRB, the aggregation level, and the starting CCE. Later fields are conditioned on earlier choices. For example, choosing the MCS before the PRB count allows the model to account for the modulation and coding when determining how many PRBs are needed to serve the queued data. The model generates successive grants until the slot’s decision is complete. At each step, it selects from a manageable set of values, conditioned on the network state and the fields already generated. TOKENS AND VOCABULARIES In language modeling, a token is the unit the model emits, such as a word or part of one, and the vocabulary is the set of tokens it may choose from, constructed from a text corpus. Both have direct counterparts in RRM-GPT. Each output token specifies the value of one decision field: for example, when generating the MCS field, the model selects an index from a configured MCS table. That set of indices is the field’s vocabulary. Unlike a text vocabulary, these choices come from the 3GPP standards and network configuration. We set out this correspondence for relevant RRM functions in Table I. A scheduling sequence specifies the grants issued within a slot, an interference-coordination sequence specifies a muting pattern, and a handover sequence specifies which users move and to which target cells. A sequence may contain zero or more such actions, each spanning multiple tokens, with an end-of-sequence token marking completion of the decision step. The same generation principle therefore carries across different RRM functions and decisions. In contrast to text, the input to RRM-GPT has no shared, predefined vocabulary and varies with the use case. The inputs include the cell’s configuration, per-user reports such as the CQI and queued data, and the cell’s resource occupancy as a two-dimensional map. These observations are heterogeneous, and none arrives as a token. We propose encoding each input according to its structure: categorical values through a learned dictionary, continuous values by their position within learned intervals to retain magnitude information, and the occupancy map in patches to preserve its layout. Related features are then grouped and projected into a common embedding space to form input tokens. For example, a user’s CQI and SINR estimate could form a channel token, while queue lengths and waiting times could form a demand token. The model thus receives a uniform sequence of tokens describing each user, the cell, and its resources. 4 TABLE I: Correspondence between autoregressive language generation and RRM decision generation. RRM functions differ in their observations, output fields, constraints, and decision scope, while sharing a token-by-token generation process. Language modelSchedulingInterference coordinationHandover InputThe prompt and text already written Per-user queues and channel reports, resource occupancy, cell configuration, and previously generated fields Per-cell load, interference between neighbors, and previously generated fields User reports of signal strength for serving and neighboring cells, cell loads, and previously generated fields TokenA word or part of a word A value for one grant fieldA value for one muting-decision field A value for one handover field VocabularyWords and sub-words of a text corpus Users, MCS indices, PRB counts and positions, control resources Cell identifiers, resource ranges, transmit or mute indicators User, cell and carrier identifiers Complete sequence A generated text sequence All grants issued in one slotA muting pattern for one coordination period Handovers issued in one decision step ExampleThe afternoon sun warmed the garden User 7· MCS 18· 12 PRBs· from PRB 24 Cell 2· PRBs 30–45· muteUser 7· to Cell 3 RulesGrammar and semantics of the language 3GPP compliance: allocations must not conflict, and control must fall in the user’s search space Muting must agree between neighbors and fit each cell’s resource budget The target must be a neighbor the user can measure and is configured to use End-of- sequence An end-of-text symbol No further grant to issueEvery cell in the cluster assigned No further handover to issue ENCODER–DECODER ARCHITECTURE RRM decisions depend on relationships among users and shared resources. A user’s queue indicates its demand, but deciding whether and how to serve it also requires considering its channel conditions, competing users, and resource avail- ability. We choose an encoder–decoder architecture to separate joint processing of the network state from sequential decision generation, as illustrated in Fig. 2b. The encoder processes the input tokens jointly through self-attention, allowing each token to incorporate information from the rest of the state. For example, a user’s demand token can draw on its channel information and the resource- occupancy tokens, providing the decoder with a representation of that demand in its network context. The encoder preserves a sequence of such representations, which the decoder can attend to selectively as it generates different fields. The autoregressive decoder combines this encoded state with the previously generated tokens to select each new field value. The encoded state provides the network context, while the growing output sequence records the choices already made within the decision step. The state is encoded once per decision step, and its representations are reused throughout generation. Let t index decision steps, s t denote the network state, and a t = (a 1 t ,...,a L t t ) denote the generated decision sequence, where L t is its length and k indexes its tokens. Autoregressive decoding factorizes the conditional distribution as p(a t | s t ) = L t Y k=1 p a k t | a 1:k−1 t ,s t ,(1) where a 1:k−1 t denotes the preceding tokens. This chain-rule factorization is exact and imposes no independence assumption among the tokens. For example, when generating a second grant, the decoder has access to the resources assigned by the first; when selecting a PRB count, it has access to the user’s previously selected MCS. This conditioning allows the model to learn dependencies within and across grants. LEARNING FROM NETWORK LOGS Pretraining uses recorded network observations paired with the RRM decisions taken in response. For scheduling, each record contains the cell and user states, and the grants issued in a slot. Logs of this kind already exist in volume: vendors and operators record scheduling decisions continuously during normal operation, and simulators can generate them at scale. As illustrated in Fig. 3, the model is given the network state and the preceding fields of the recorded decision and trained to predict the next field, following the same next-token prediction objective used in language modeling. Logged RRM decisions provide two kinds of information: the rules that determine whether a decision is valid, and the controller’s behavior when choosing among valid decisions. For example, valid scheduling records avoid conflicting re- source allocations and place control information on resources the scheduled user is configured to monitor under 3GPP spec- ifications. The choice of which users receive grants and which MCS values they use reflects the controller’s policy. Predicting logged fields allows the model to learn both validity rules and decision-making behavior without explicitly implementing the controller’s algorithms, much as predicting text teaches a language model both grammar and semantics. The policy learned through imitation is limited by the behavior represented in the logs. Objective-driven post-training can subsequently refine that policy, with the aim of improving network performance beyond that of the reference controller. 5 (a) LEARNING FROM LOGS Every field of every logged grant is a training target LOGGED SLOTMany logged slots · no manual annotation or rewards NETWORK STATE UE BUFFERUE CHANNELOCCUPANCYUE HARQ HISTORY ACKNACKACK GRANT User 7 MCS 18 PRB count 12 Start PRB 24 Aggregation level 4 Start CCE 8 LOGGED FIELDS SUPPLIED AS CONTEXT AUTOREGRESSIVE MODEL PREDICTED START PRB DISTRIBUTION 202224262830 Logged target Start PRB 24 CROSS-ENTROPY (b) GENERATING FEASIBLY Infeasible values are removed before each choice COMMITTED SO FARRESOURCE OCCUPANCY SLOTS t+1 t+2 t+3 GRANT PLACED PRBs 20–23 202224262830 PRBs AUTOREGRESSIVE MODEL PREDICTED START PRB DISTRIBUTION Low predicted probability; masked regardless 202224262830 ALREADY ASSIGNED MASK For the selected slot and PRB count 24 SELECTED Fig. 3: Training and generation at the same step, the starting PRB of User 7’s grant. (a) Every field of a logged grant is a supervised target, predicted from the state and the fields before it and scored against the logged value. (b) At generation, values conflicting with what the slot has committed are removed first; the model already puts almost no weight on them. GUARANTEEING FEASIBILITY The rules learned from logs capture statistical regularities but do not provide hard guarantees. While in language this may produce unconventional content, in RRM an invalid grant or control information placed outside a user’s monitored search space can directly disrupt network operation. Moreover, feasibility is conditional: the validity of each value depends on previously selected fields and resources already committed within the slot. Earlier allocations restrict the resources available to subsequent users, and the remaining control-channel capacity decreases as grants are issued. Be- cause each grant is generated field by field, these constraints can be enforced during decoding, as shown in Fig. 3b. At each step, values that would violate the current constraints are masked before selection, ensuring that the model chooses only from the remaining feasible values. A lightweight state tracker records the selected field values and resource commitments, updating them after each generated field so that the mask can determine which values remain feasible at the next step. Masking is applied only during inference to ensure feasibil- ity. A model trained on sufficient logs largely respects these rules on its own, so masking rarely changes the outcome. Even rare violations, however, can disrupt network operation, which makes lightweight guarantees of this kind essential for practical deployment. BEYOND SCHEDULING: AUTOREGRESSIVE MODELING OF JOINT RRM The scheduling formulation above generates all grants within a slot as one decision sequence. It takes the user’s serving cell and any resource restrictions imposed by inter-cell coordination as given. These conditions are themselves shaped by other RRM functions, including mobility, interference coordination, and load management. Extending to functions beyond scheduling does not require a new generation mechanism. The framework accommodates these functions naturally by moving the boundary of what the model treats as a decision. As shown in Fig. 1, a shared model can generate load-management, handover, interference- coordination, and scheduling decisions autoregressively as an interdependent sequence. Decisions made by other RRM functions, previously treated as inputs to scheduling, can now be generated by the shared model, allowing decisions across functions to condition on one another. A UNIFIED GENERATIVE ARCHITECTURE Sharing a model does not require all RRM functions to act at every step. Each function retains its own decision structure and control timescale and is invoked when its control opportunity arises. In a slot-based implementation, most steps would contain scheduling grants alone, while some would also include handover, interference-coordination, or load- management decisions. Generation uses only the observations and prior decisions available at the relevant time and network scope. Joint modeling therefore does not assume access to a complete, synchronized network state. When several functions are invoked together, the generation order follows their dependencies. An interference-coordination decision precedes the grants whose resources it constrains, while a handover decision conditions scheduling for the affected user. Slower decisions need not be regenerated at every slot: their continuing effects remain part of the network state and condition faster decisions until they are revised. The functions share the model and its decision context while retaining their own decision structures and control timescales. 6 The value of this architecture lies in modeling interactions among RRM functions. An overloaded cell may be relieved by moving its users to a neighbor, by coordinating resources with neighboring cells, or by changing which users are scheduled. Handover, load management, and interference coordination each change the problem presented to the others. Their de- cisions can complement or substitute for one another, making it useful to model them jointly. A unified model captures these dependencies within one policy. Autoregressive generation breaks this broader joint decision into a sequence of smaller choices, each conditioned on those already committed. POST-TRAINING AND SPECIALIZATION Pretraining provides a common initialization, but the be- havior needed in a particular deployment may differ from that represented in the original logs. During post-training, the pretrained model retains the same state representation and autoregressive output structure, while its weights are adapted using deployment-specific data or feedback. This may update the full model or only a small set of specialization parameters attached to the shared backbone. Behavioral cloning. The pretrained model can be fine-tuned to imitate a reference controller’s logged decisions, using the same prediction objective as pretraining without an explicit reward. This offers a way to approximate a computationally expensive controller: high-quality decisions generated offline can train a model intended for real-time operation. Pretrain- ing on logs from different controllers exposes the model to multiple scheduling approaches; behavioral cloning then specializes it to a target deployment, potentially requiring fewer demonstrations than training from scratch. Reinforcement learning. Alternatively, the model can be optimized directly for desired network outcomes. The model generates RRM decisions, and their effects on network per- formance are evaluated using an operator-defined objective. Reinforcement learning uses this feedback to adjust the prob- abilities of the generated decisions to increase the expected long-term return, similar to RL-based post-training of language models [12]. Throughput, latency, fairness, and energy effi- ciency can inform the reward, with their relative importance reflecting deployment priorities. Multi-objective reinforcement learning offers a further extension: conditioning the policy on these priorities and training across different trade-offs could allow one model to support changing operator preferences without retraining for each setting [13]. In each case, post-training builds on the relationships between network observations and decisions learned during pretraining. Slow convergence and the training required to learn effective RRM policies motivate methods that accelerate learning through knowledge transfer [14]. Reusing a pretrained model offers a way to address this challenge, potentially reducing the deployment-specific data or interaction needed to adapt to a target controller or objective. The following case study instantiates the framework for 5G NR scheduling. It evaluates whether the proposed au- toregressive model can learn and generate coupled scheduling decisions from network logs. CASE STUDY: 5G NR SCHEDULING EXPERIMENTAL SETUP We generate scheduling logs using a system-level 5G NR simulator with a reference scheduler. The simulated urban macro cell operates at 3.5 GHz over 40 MHz, with 106 PRBs, numerology μ = 1, and a time-division duplexing pattern of three downlink (DL) slots followed by one uplink (UL) slot. Users request back-to-back 100 KB FTP files. The training set contains approximately 6.5 million grants from runs lasting up to 60 s, with 5, 10, or 20 active users moving at 50 km/h within a cell of radius 1 km. We evaluate on 637,000 grants from held-out runs of the same scenario. To assess generalization, we also test on 158,000 grants from a separate scenario with static users in a cell of radius 167 m and the same active-user counts. This scenario is excluded entirely from training. At each slot, the model generates zero or more grants. Each grant specifies the selected user, data-slot offset, MCS, PRB count, starting PRB, control-channel aggregation level, and starting CCE, in that order. These fields span user selec- tion, timing, link adaptation, resource allocation, and control signaling. An end-of-sequence token marks completion of the slot decision. Grants are issued for both DL and UL transmissions, and each schedules either an initial transmission or a retransmission. The model receives the cell configuration, a time-frequency occupancy grid, and per-user information on channel condi- tions, traffic demand, and control-channel search space. The encoder and decoder each comprise two Transformer layers with an embedding dimension of 256 and four attention heads per layer, yielding approximately 4.3 million parameters. We train for 50 epochs using AdamW with a batch size of 128. NUMERICAL RESULTS We compare autoregressive rollout with teacher-forced pre- diction (Fig. 4). During rollout, the model conditions each pre- diction on its own previously generated fields. Under teacher forcing, the model receives the preceding ground-truth fields from the logged decision and predicts the next field, which is scored against its logged value. The logged value, rather than the prediction, is then used when predicting subsequent fields. This comparison helps distinguish difficulty in predicting a field from the effects of errors in earlier choices. Generated grants are matched to logged grants by user, so differences in grant ordering do not affect the scores. Missing grants count as errors for all fields in exact-match accuracy but are excluded from mean absolute error (MAE). The results show that the model learns dependencies among grant fields, although reproducing the full decision sequence remains more difficult. Timing, aggregation level, and retrans- mission MCS are reproduced with relatively high accuracy during rollout. Resource placement depends more strongly on the preceding choices: uplink starting-PRB accuracy rises from 48% to 99% under teacher forcing, with MAE falling from 2.8 to 0.1 PRB indices. The model thus reproduces resource placement accurately when given the preceding logged fields, suggesting that error propagation accounts for much of the mismatch during rollout. 7 Selection Timing Link adaptation Resource placement Control channel Grant issued Data-slot offset Initial transmission MCS Retransmission MCS PRB count Starting PRB Aggregation level Starting CCE 0%20%40%60%80%100% Exact-match accuracy (%) Autoregressive rollout 93% 83% 93% 83% 40% 47% 90% 89% 78% 55% 89% 48% 93% 83% 71% 57% 0%20%40%60%80%100% Exact-match accuracy (%) Teacher-forced prediction 100% 100% 100% 100% 44% 64% 96% 99% 89% 77% 100% 99% 100% 100% 77% 89% Link adaptation Resource placement Control channel Initial transmission MCS PRB count Starting PRB Starting CCE 0246810 Mean absolute error 1.0 1.2 3.2 2.2 0.7 2.8 1.6 2.0 0246810 Mean absolute error 1.0 1.1 2.3 1.3 0.0 0.1 1.5 0.6 DownlinkUplink Fig. 4: Scheduling-field prediction on the mobile test set. Exact-match accuracy (top) and mean absolute error (bottom) are compared under autoregressive rollout and teacher forcing. MAE is expressed in the units of each field. Selection Timing Link adaptation Resource placement Control channel Grant issued Data-slot offset Initial transmission MCS Retransmission MCS PRB count Starting PRB Aggregation level Starting CCE 0%20%40%60%80%100% Exact-match accuracy (%) 93%98% 83%93% 93%98% 83%93% 40%68% 47%91% 90%98% 87%89% 78%94% 55%92% 89%92% 37%48% 93%98% 83%93% 71%94% 57%86% Mobile testStatic testDownlinkUplink Fig. 5: Generalization of the mobile-trained model to the held-out static scenario without adaptation. Exact-match accuracy is evaluated under autoregressive rollout; arrows indicate the change from mobile to static testing. For initial transmissions, the model predicts MCS values close to those selected by the reference scheduler, with rollout MAEs of 1.0 and 1.2 MCS indices for downlink and uplink, respectively. The small MAEs in both evaluation modes show that the model approximates the reference scheduler’s MCS choices more closely than exact-match accuracy alone sug- gests. We next evaluate the model on the held-out static scenario without adaptation (Fig. 5). Accuracy improves for most reported metrics under the change in mobility and cell radius. 8 Static users and a smaller cell yield more stable channel conditions, which may ease prediction; the gains nonetheless show that the model has not memorized the training scenario. Initial-transmission MCS accuracy, for example, rises from 40% to 68% for downlink and from 47% to 91% for uplink. Transfer is not uniform: uplink starting-PRB accuracy falls from 48% to 37%. The same model thus reproduces substantial aspects of the reference scheduler’s behavior in a scenario excluded from training, supporting reuse across deployments. Together, these results support the function-specific scope of the proposed framework: one autoregressive model generates the fields of scheduling grants and transfers learned behavior to an unseen deployment scenario. They provide an initial basis for model reuse, while broader generalization, the benefits of post-training, and the impact on network performance remain open for future evaluation. CONCLUSION AND OPEN PROBLEMS RRM decisions retain a common structure across deploy- ments, even as cell configurations, propagation conditions, traffic demands, and operator objectives change. We proposed an autoregressive framework that exploits this structure to learn RRM decisions from network logs, with post-training as the route to deployment-specific adaptation. Our 5G NR case study demonstrates complete scheduling-grant generation and partial generalization to an unseen scenario. Establishing reliable adaptation across a wider range of deployments and objectives is the next step toward RRM foundation models. Sharing models across RRM functions represents a broader ambition, extending reuse to interdependent decisions taken at different scopes and timescales. Pursuing these directions requires addressing several open research problems. Representation and transfer across systems. The frame- work expresses each decision as an ordered set of fields, but which fields to use, their granularity, and order remain design questions. A representation suited to one standard release may not transfer to the next, and the vocabularies describing 5G NR grants differ from those used in Wi-Fi or satellite systems. An open question is whether a shared representation can be defined that is general enough to transfer across systems yet specific enough to generate valid decisions, allowing a model pretrained on one network to provide a useful starting point for another. Network logs and data coverage. Pretraining depends on the coverage and detail of network logs. Models intended for reuse across deployments require records spanning dif- ferent configurations, channel conditions, traffic demands, and controller policies. Extending learning across RRM functions also requires traces that preserve the timing and dependencies between their decisions, together with the relevant network observations. Establishing what to record, at what granularity, and how to combine logs from different vendors and deploy- ments is therefore a central research challenge. Complexity, latency, and deployment. Decisions taken every slot impose tight inference budgets, while decisions and inputs processed every few minutes allow greater com- putational complexity. Architectures should therefore avoid repeating expensive computation at fast timescales. Where models run and how their instances coordinate are also open questions. A single pretrained model may be deployed at multiple points in the network, each operating on information available to its local control loop. Maintaining consistency among decisions made from these different local observations remains an important challenge. Expressing and pursuing operator intent. Post-training requires a well-defined objective, yet operators may not always be able to specify one explicitly. Some goals are directly quantifiable, while others are more naturally expressed through relative preferences, and many vary with time, deployment conditions, and commercial priorities. This raises two key questions: how operational intent should be expressed and how to adapt a model to that intent without unsafe exploration in a live network. Designing a single model that supports diverse objectives through lightweight specialization is another related open challenge. REFERENCES [1] B. Agarwal, M. A. Togou, M. Marco, and G.-M. Muntean, “A compre- hensive survey on radio resource management in 5G HetNets: Current solutions, future trends and open issues,” IEEE Commun. Surveys Tuts., vol. 24, no. 4, p. 2495–2534, 2022. [2] 3GPP, “NR and NG-RAN overall description; stage 2,” Technical Specification (TS) 38.300, 3rd Generation Partnership Project (3GPP), 2024. Version 18.4.0. [3] F. Al-Tam, N. Correia, and J. Rodriguez, “Learn to schedule (LEASCH): A deep reinforcement learning approach for radio resource scheduling in the 5G MAC layer,” IEEE Access, vol. 8, p. 108088–108101, 2020. [4] X. Ye, Y. Yu, and L. Fu, “Deep reinforcement learning based link adaptation technique for LTE/NR systems,” IEEE Trans. Veh. Technol., vol. 72, no. 6, p. 7364–7379, 2023. [5] Y. Sun, M. Peng, Y. Zhou, Y. Huang, and S. Mao, “Application of machine learning in wireless networks: Key techniques and open issues,” IEEE Commun. Surveys Tuts., vol. 21, no. 4, p. 3072–3108, 2019. [6] A. Valcarce, P. Kela, S. Mandelli, and H. Viswanathan, “The role of AI in 6G MAC,” in Proc. Joint Eur. Conf. Netw. Commun. 6G Summit (EuCNC/6G Summit), p. 723–728, 2024. [7] K. Pourtahmasi Roshandeh, A. Bin Sediq, and J. Sadreddini, “Artificial intelligence native radio resource management inspired by large lan- guage models.” International Patent Application WO 2026/093945 A1, May 2026. [8] T. Yang, P. Zhang, M. Zheng, Y. Shi, L. Jing, J. Huang, and N. Li, “WirelessGPT: A generative pre-trained multi-task learning framework for wireless communication,” IEEE Netw., vol. 39, no. 5, p. 58–65, 2025. [9] A. Aboulfotouh, E. Mohammed, and H. Abou-Zeid, “6G WavesFM: A foundation model for sensing, communication, and localization,” IEEE Open J. Commun. Soc., vol. 6, 2025. [10] O. Mashaal and H. Abou-Zeid, “IQFM—A wireless foundation model for I/Q streams in AI-Native 6G,” IEEE Open J. Commun. Soc., vol. 7, p. 1426–1441, 2026. [11] E. Dahlman, S. Parkvall, and J. Sk ̈ old, 5G NR: The Next Generation Wireless Access Technology. Academic Press, 2nd ed., 2020. [12] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. K. Li, Y. Wu, and D. Guo, “DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300, 2024. [13] B. Demirel, P. Soldati, and Y. Wang, “From intents to actions: Agentic AI in autonomous networks,” arXiv preprint arXiv:2602.01271, 2026. [14] A. M. Nagib, H. Abou-Zeid, and H. S. Hassanein, “Toward safe and accelerated deep reinforcement learning for next-generation wireless networks,” IEEE Netw., vol. 37, no. 2, p. 182–189, 2023.