BRAIN: Bayesian Reasoning via Active Inference

• The paper introduces BRAIN—a deep active inference agent that unifies Bayesian reasoning for real-time RAN slicing and radio resource management in O-RAN networks.
• It employs a latent-state generative model with variational free energy minimization to update network conditions and guide action selection.
• Empirical evaluations on a GPU-accelerated 5G AI-RAN testbed show faster reward convergence and up to 28.3% enhanced robustness versus DRL baselines.

Bayesian Reasoning via Active Inference (BRAIN) is a deep active-inference agent proposed for real-time control in mobile networks, specifically for closed-loop RAN slicing and radio resource management in an O-RAN deployment. It maintains an explicit probabilistic generative model of the network, performs Bayesian belief updating over latent network conditions, and chooses actions by minimizing variational free energy and expected free energy, thereby unifying perception and control in a single closed-loop paradigm. In the reported implementation, BRAIN operates as a Near-RT RIC xApp on a GPU-accelerated private 5G AI-RAN testbed and is evaluated against tuned heuristics and several deep reinforcement learning baselines (Basaran et al., 15 Feb 2026).

1. Conceptual lineage and scope

BRAIN belongs to the active-inference tradition in which an agent is modeled as maintaining beliefs about hidden causes of observations and selecting actions through inference under a generative model rather than through reward maximization alone. In that literature, active inference is presented as a normative Bayesian framework in which action, perception, exploration, and exploitation emerge from inference under a generative world model, with policy selection governed by expected free energy (Costa et al., 2024).

The framework also inherits the standard active-inference interpretation of the brain as an approximate Bayesian inference engine. In deep active-inference robotics, this interpretation was operationalized through latent-state generative models learned directly from raw sensory inputs, with perception, learning, and action all expressed in terms of probability distributions, priors, approximate posteriors, and KL divergences (Çatal et al., 2020). More broadly, active inference has been surveyed as a variational Bayesian framework that unifies state estimation, action, planning, and learning in robotics and artificial agents (Lanillos et al., 2021).

Within that lineage, BRAIN is specialized to mobile networks. Its stated target is future sixth-generation mobile networking, where artificial intelligence agents are required to be autonomous, real-time adaptive in dynamic environments, and transparent in their decisionmaking. The paper frames prevailing DRL-based network agents as limited by lack of explainability, brittle adaptation under non-stationary conditions, and catastrophic forgetting. BRAIN is proposed as an alternative in which service intent is expressed through prior preferences over outcomes and control is implemented by active inference rather than by reward-maximizing policy learning alone (Basaran et al., 15 Feb 2026).

2. Probabilistic formulation

BRAIN treats network control as Bayesian inference in a latent dynamical system. The hidden state $s_t$ represents underlying network or slice conditions, the observation $o_t$ consists of O-RAN telemetry, and the action $a_t$ is a radio resource management decision. The paper gives the active-inference state-space model as

$$P(o_{1:T},\,s_{0:T}) \;=\; P(s_0)\,\prod_{t=1}^{T} P(s_t \mid s_{t-1}, a_{t-1}) \; P(o_t \mid s_t),$$

and also writes the expanded factorization

$$P(s_{0:T}, o_{0:T}, a_{0:T}) = P(s_0)\prod_{t=0}^{T-1} P(o_t \mid s_t)\,P(s_{t+1} \mid s_t, a_t)\,P(a_t).$$

These expressions specify an initial prior over hidden network state, an observation model, controlled latent dynamics, and an action term (Basaran et al., 15 Feb 2026).

Perception is cast as variational Bayesian inference. The paper defines a generic variational free energy

$$\mathcal{F}(q, o_t) = \mathbb{E}_{q(s_t)} \Big[\ln q(s_t) - \ln P(o_t, s_t)\Big],$$

and then uses the time-indexed form

$$F_t(Q(s_t)) = \mathbb{E}_{Q(s_t)} \big[ -\ln P(o_t \mid s_t) -\ln P(s_t \mid s_{t-1}, a_{t-1}) \big] + \mathbb{H}[Q(s_t)].$$

The posterior belief state is written as

$Q(s_t) = P(s_t \mid o_{1:t}, a_{1:t-1}).$

This is the inferential state made available to the controller and to the interpretability layer (Basaran et al., 15 Feb 2026).

Action selection is based on expected free energy. At policy level, the paper gives

$$G(\pi) \;=\; \mathbb{E}_{q(o_{>t},\,s_{>t} \mid \pi)} \Big[ \ln q(s_{>t} \mid o_{>t}) - \ln P(o_{>t}, s_{>t}) \Big].$$

For one-step action selection it writes

$G(a_t) = \mathbb{E}_{Q(s_{t+1}\mid s_t, a_t)} \Big[ -\ln P(o_{t+1} \mid C) + \mathrm{D_{KL}\big(Q(s_{t+1}\mid o_{t+1}) \,\|\, Q(s_{t+1})\big) \Big].$

The first term is the preference-alignment or extrinsic term, while the second is an information-gain term. This reproduces the standard active-inference view that good actions are those expected to realize preferred observations while reducing uncertainty about hidden states. The same logic is explicit in the broader active-inference literature, where expected free energy decomposes into pragmatic and epistemic terms and, in limiting cases, reduces to Bayesian decision theory or optimal Bayesian design (Sajid et al., 2021).

3. Embodiment in O-RAN mobile networks

BRAIN is implemented as a Near-RT RIC xApp for closed-loop RAN slicing control. The xApp receives KPM telemetry, infers latent slice and network conditions, evaluates candidate control actions by expected free energy, and sends control updates to the gNB. In this setting, the agent is embodied in the control loop rather than acting as an offline predictor (Basaran et al., 15 Feb 2026).

The application domain is a multi-slice RAN with three slice types: eMBB, URLLC, and mMTC. The controller aims to maintain slice-specific QoS targets associated with throughput, latency, and reliability. Observations are slice-level metrics such as downlink throughput, buffer occupancy or queue length, and downlink transport-block count. Actions comprise two coupled decisions: PRB allocation fractions among slices and scheduler choice per slice, including PF, RR, and WFQ (Basaran et al., 15 Feb 2026).

The reported deployment uses a GPU-accelerated private 5G AI-RAN testbed based on NVIDIA Aerial Research Cloud, NVIDIA Aerial SDK, OpenAirInterface higher layers, a Foxconn O-RU in n78 TDD, and 100 MHz bandwidth with 273 PRBs at 30 kHz SCS. The architecture places O-DU Low or PHY on NVIDIA GPU and O-DU High or CU on x86 CPU, with O-RAN 7.2 fronthaul to the radio unit. The evaluation uses one cell, two UEs, and three slices, and all agents are trained through direct interaction with the RAN testbed in an online learning fashion over about $o_t$0 environment interactions at a 20 ms control interval (Basaran et al., 15 Feb 2026).

The baselines are a tuned heuristic and five DRL agents: DQN, A2C, REINFORCE, PPO, and SAC. For comparison, the RL reward is defined as

$o_t$1

with the usual discounted return

$o_t$2

This baseline makes explicit that BRAIN is not evaluated in isolation but against conventional reward-driven controllers (Basaran et al., 15 Feb 2026).

4. Belief-state diagnostics and interpretability

A defining feature of BRAIN is that its internal control variables are exposed as probabilistic beliefs rather than only as network outputs. The paper emphasizes posterior beliefs over latent slice conditions,

$o_t$3

together with action-level free-energy terms. This is presented as intrinsic interpretability rather than post hoc explanation (Basaran et al., 15 Feb 2026).

The interpretability visualizations discretize slice demand into latent states such as Low, Medium, and High, and display belief trajectories for eMBB, URLLC, and mMTC. The reported heatmaps show posterior mass shifting between these latent demand states as traffic evolves and after information-gathering actions. The paper describes these diagnostics as human-interpretable belief state diagnostics and uses them to support the claim that BRAIN performs robust causal reasoning for dynamic radio resource allocation (Basaran et al., 15 Feb 2026).

The same section reports expected-free-energy decompositions in which epistemic value is dominant early in operation and extrinsic value becomes stronger later. The slice-level pattern differs by service type: eMBB shows a clearer epistemic-then-extrinsic shift, URLLC remains more balanced because latency constraints are stricter, and mMTC exhibits a longer epistemic phase under sparse or bursty demand. This suggests a control regime in which uncertainty reduction is emphasized before more exploitative QoS maintenance. A related interpretability motif already appeared in deep active-inference robotics, where imagined trajectories and decoded preferred observations were used to make latent beliefs and predictions inspectable during action selection (Çatal et al., 2020).

The action posterior is written in softmax form,

$o_t$4

and the associated policy entropy is logged over time. The reported training dynamics show policy entropy starting high and then declining gradually without collapsing to zero, which the paper interprets as a controlled exploration-exploitation balance (Basaran et al., 15 Feb 2026).

5. Relation to reinforcement learning

BRAIN is explicitly positioned against DRL, but it is not orthogonal to the planning-as-inference literature. Active inference and RL are strongly related whenever utilities are encoded as prior preferences over outcomes or states. In the active-inference literature, expected free energy can be decomposed into an extrinsic value term and an intrinsic value term, and when information-gain and ambiguity-related terms are removed the objective reduces to expected utility maximization (Sajid et al., 2021).

That relationship is also visible in other machine-learning treatments of active inference. Reinforcement-learning formulations inspired by active inference replace primitive rewards with preferences over observations and unify exploration with exploitation through information gain, although they may use objectives such as the free energy of the expected future rather than canonical expected free energy (Tschantz et al., 2020). Likewise, work on prior preference learning from expert demonstrations treats expected free energy as a negative value function and uses learned prior preferences as a route from active inference to inverse reinforcement learning (Shin et al., 2021).

BRAIN departs from standard DRL most clearly in two places. First, its control objective is not a reward regressor but a preference-sensitive expected free energy under a generative model. Second, adaptation proceeds through belief updating rather than by retraining a policy network after distribution shift. The paper therefore frames DRL brittleness and catastrophic forgetting as consequences of opaque reward-driven policies, while BRAIN’s latent-state and free-energy formulation is presented as supporting real-time adaptation without environment-specific retraining (Basaran et al., 15 Feb 2026).

A recurrent misconception is that active inference is simply reward maximization stated in Bayesian language. The broader literature does not support that reduction. Active inference canonically integrates exploration and exploitation by minimizing risk and ambiguity, and it evaluates policies by posterior probability under a generative-preference model rather than by return alone (Costa et al., 2024).

6. Empirical findings, limitations, and open directions

In the reported experiments, BRAIN achieves faster reward convergence, higher asymptotic reward, and more stable training loss than the listed baselines. Slice-specific empirical CDFs are reported as most favorable for BRAIN: eMBB throughput is most right-shifted, the URLLC PRB-ratio distribution is most right-shifted, and mMTC downlink transport-block delivery is improved with fewer near-starvation intervals. The abstract summarizes the main result as robust causal reasoning for dynamic radio resource allocation while maintaining slice-specific QoS targets under varying traffic loads (Basaran et al., 15 Feb 2026).

The most prominent robustness claim concerns non-stationarity. After a controlled traffic-distribution shift, all agents show a QoS satisfaction drop, but BRAIN is reported to exhibit the smallest degradation and the fastest recovery. The abstract quantifies this as superior adaptability with up to 28.3% higher robustness to sudden traffic shifts versus benchmarks, achieved without any retraining (Basaran et al., 15 Feb 2026).

The paper also states several practical and conceptual limitations. The generative-model mathematics are given, but the exact deep architecture and parameter-learning procedure are not specified in comparable detail. The robustness figure is reported, but the exact algebraic definition of the robustness score is not explicitly presented. The deployment is realistic yet limited to a single-cell setting with two UEs and three slices. These points leave open how the method scales under denser mobility, stronger interference coupling, and larger action spaces (Basaran et al., 15 Feb 2026).

The proposed future directions are hierarchical, multi-timescale active inference in O-RAN, multi-cell and multi-agent scaling, and distributed belief sharing under realistic telemetry and fronthaul constraints. A plausible implication is that BRAIN is best understood not as a finished universal controller, but as a concrete mobile-network instantiation of a broader shift from passive estimation and reward-driven optimization toward active Bayesian control. That broader shift is also visible in adjacent domains, including robotics and active digital twins, where active inference is used to turn latent-state models into closed-loop agents that combine probabilistic prediction, uncertainty reduction, and goal-directed action (Torzoni et al., 17 Jun 2025).