xDiff for O-RAN ICIM via Diffusion RL

• The paper introduces xDiff as a diffusion-based reinforcement learning framework that formulates collaborative inter-cell interference management as a resource allocation optimization problem in O-RAN.
• xDiff integrates a conditional denoising diffusion model with double-Q critics to learn multi-modal, preference-value actions that bias DU scheduling towards interference-aware allocations.
• The framework achieves up to 120% improvement in aggregate rewards over baselines while converging in approximately 20 iterations with low latency and enhanced scheduling flexibility.

xDiff is a diffusion-based reinforcement learning framework for collaborative inter-cell interference management (ICIM) in Open Radio Access Network (O-RAN). It formulates ICIM as a resource allocation optimization problem aimed at maximizing a user-defined reward function, and it develops an online learning solution by integrating a diffusion model into an RL framework for near-real-time policy generation. A central design choice is the introduction of “preference values” as the policy representation, enabling efficient policy-guided resource allocation within O-RAN distributed units (DUs) (Yan et al., 19 Aug 2025).

1. ICIM formulation in O-RAN

xDiff considers a set of small cells $\mathcal{K}$, each with $DU_k$ and served UEs $\mathcal{U}_k$. Time is slotted at the Near-RT RIC granularity $t=0,1,2,\dots$. For UE $i\in\mathcal U$, the formulation uses throughput demand $P_i$, achieved throughput $\rho_i(t)$ in slot $t$, delay demand $D_i$, and measured average queueing delay $\tau_i(t)$. Per-UE “regret” is defined as $DU_k$0 for throughput and $DU_k$1 for delay. Converting regret into a reward and aggregating by cell-level weights $DU_k$2 gives

$DU_k$3

$DU_k$4

The control problem is posed as an MDP in which the Near-RT RIC chooses a policy $DU_k$5 mapping observed DU/KPM/MAC data, i.e. state $DU_k$6, to a “preference-value” action $DU_k$7, with objective

$DU_k$8

There are no hard coupling constraints among DUs at each millisecond, but the RIC’s policies must steer each DU’s resource allocation to mitigate inter-cell interference.

This formulation places xDiff at the interface between near-real-time policy generation and real-time MAC scheduling. A plausible implication is that the method is designed less as a replacement for DU scheduling than as a mechanism for biasing DU decisions toward interference-aware allocations.

2. Diffusion policy learning and online update loop

The core of xDiff is a conditional denoising diffusion probabilistic model acting as a policy generator $DU_k$9, combined with double-Q critics $\mathcal{U}_k$0 (Yan et al., 19 Aug 2025). At iteration $\mathcal{U}_k$1, the framework collects a batch of transitions $\mathcal{U}_k$2 where $\mathcal{U}_k$3 is sampled via the diffusion network, stores them in replay buffer $\mathcal{U}_k$4, samples mini-batches, and updates critics by double Q-learning:

$\mathcal{U}_k$5

$\mathcal{U}_k$6

where prime denotes target networks.

The diffusion policy parameters $\mathcal{U}_k$7 are updated by minimizing

$\mathcal{U}_k$8

which combines the standard DDPM denoising loss $\mathcal{U}_k$9 with a Q-value “guidance” term $t=0,1,2,\dots$0. Here $t=0,1,2,\dots$1 trades off generative fidelity vs. reward maximization, and $t=0,1,2,\dots$2 indexes diffusion steps. Target networks are soft-updated as

$t=0,1,2,\dots$3

At inference, the RIC runs $t=0,1,2,\dots$4 denoising steps, starting from Gaussian noise and conditioning on $t=0,1,2,\dots$5, to produce a new action $t=0,1,2,\dots$6. In xDiff, this action is the set of preference values. The paper attributes the utility of diffusion to multi-modal policy capture: in interference management, multiple distinct resource-allocation patterns can be near-optimal, whereas Gaussian policies average these into suboptimal “mean” actions. It also reports convergence in approximately 20 Near-RT iterations versus 50–100 for DDPG/DDQN.

3. Preference values and DU-side scheduling

To bridge the RIC’s near-RT timescale of 10 ms–1 s and each DU’s real-time 1 ms scheduling, xDiff outputs a continuous “preference value”

$t=0,1,2,\dots$7

The semantics are explicit: $t=0,1,2,\dots$8 is a strong recommendation to assign resource block $t=0,1,2,\dots$9 to UE $i\in\mathcal U$0, $i\in\mathcal U$1 is strong avoidance due to expected interference, and $i\in\mathcal U$2 means no RIC preference and the DU scheduler decides. The representation is motivated by two observations: hard $i\in\mathcal U$3 policies were too rigid and starved some UEs, while continuous weights preserve DU scheduling flexibility, for example proportional fairness, while biasing away from heavily interfered RBs. The reported ablation showed up to 40–50% higher reward and much lower delay when using continuous vs. discrete policies (Yan et al., 19 Aug 2025).

Each DU’s MAC scheduler remains conventional proportional fairness (PF). For UE $i\in\mathcal U$4 and subframe $i\in\mathcal U$5, PF computes

$i\in\mathcal U$6

where $i\in\mathcal U$7 is an instantaneous rate estimate from CQI/MCS and $i\in\mathcal U$8 is UE $i\in\mathcal U$9’s average throughput history. xDiff modifies PF in two steps. First, it computes the per-UE weight

$P_i$0

the fraction of RBs flagged as “bad.” Second, it multiplies the PF metric by $P_i$1:

$P_i$2

The DU then ranks UEs by $P_i$3 and, for each in turn, allocates the highest-$P_i$4 RBs until its data demand is met or no RB remain. If a chosen RB is underfilled, any residual portion may be reclaimed by lower-priority UEs.

A common misconception is that xDiff produces hard per-RB assignments. The described deployment instead uses continuous preference values to steer, rather than replace, DU scheduling logic.

4. Implementation on a 5G O-RAN testbed

The reported implementation uses three radio units: USRPs N310 and X310 for 2×2 MIMO, and B210 for 1×1, all synchronized for TDD on the n78 band with center frequency 3319.68 MHz and 30 kHz SCS (Yan et al., 19 Aug 2025). Distributed Units and CUs run OpenAirInterface RAN, patched for E2AP/KPM and custom MAC scheduling. The Near-RT RIC is Intel FlexRIC with a Python/C xApp for xDiff, using SWIG for binding. The UE set consists of ten commercial smartphones from Pixel, OnePlus, Motorola, Xiaomi, and Samsung.

The data plane uses iPerf to generate per-UE UDP/TCP downlink loads up to 165 Mbps, while KPM/MAC reports including PRBs, PHR, SNR, MCS, BLER, throughput, and delay stream at 100–1000 samples/s via E2. To meet the timing target of at most 1 s, the 106 RBs are clustered into 10 groups for the diffusion network output and then expanded back at the DU.

The neural architecture uses a 4-layer MLP with 256 units and Mish activations for both the diffusion denoiser $P_i$5 and the two Q-networks. The deployed configuration uses 5 denoising steps and $P_i$6, selected by hyperparameter sweep. This suggests that the practical instantiation is intentionally compact, with policy quality obtained from the diffusion-RL coupling rather than from very large network capacity.

5. Empirical behavior, baselines, and latency-quality trade-offs

The evaluation uses two representative scenarios: a lab setting with strong interference, 3 cells, and 10 UEs with moving obstacles; and a building setting with light interference, cells approximately 18 m apart through walls, and 200 ms E2 delay over Wi-Fi (Yan et al., 19 Aug 2025). Baseline ICIM methods are CIRA, described as cell-independent PF over all RBs; OTFR, described as one-third static RB partition; CSRS, a cluster-based spectrum sharing method; and IAIS, an ML-based interference prediction plus scheduling method.

Across 5 h traffic variations, the key aggregates are explicit. For throughput, xDiff and CSRS both met more than 95% of UE demands in the lab; xDiff outperformed IAIS by 12%, CIRA by 28%, and OTFR by 35%. In the building scenario, xDiff was marginally 3–5% above CSRS. For delay, xDiff achieved median queueing delay at most 2 ms in the lab, versus 5–20 ms for the alternatives, and at most 1 ms in the building, versus 3–8 ms. BLER remained moderate at 10–15%, higher than OTFR but far below CIRA’s 30–40%. For aggregate reward $P_i$7, xDiff improved by 40% versus CSRS, 70% versus IAIS, and 120% versus CIRA.

The latency-quality trade-off is also quantified. On a 14th-gen i9 CPU, OTFR and CIRA require at most 1 ms, CSRS 5.4 ms, IAIS 38.3 ms, and xDiff 21.8 ms. All meet O-RAN Near-RT at at most 1 s, but xDiff trades extra latency for quality. The paper also characterizes the method as robust because continuous policy distributions let DUs gracefully degrade or shift resource blocks when channels or loads vary, reducing UE disconnections.

6. Limitations, future directions, and terminological ambiguity

The reported limitations are operational rather than purely algorithmic (Yan et al., 19 Aug 2025). Inference latency of approximately 20 ms still blocks sub-10 ms adaptation and therefore requires model compression or fewer denoising steps. KPM/MAC reporting over E2 may become a bottleneck if $P_i$8 Hz or if more cells join. Clustering RBs trades granularity versus timing. Future work identified in the paper includes hierarchical diffusion with coarse and fine RB nets for sub-5 ms inference, multi-agent diffusion across multiple RICs for ultra-dense networks, learning sparse or structured noise schedules to further reduce steps, and joint uplink/downlink diffusion policies for full duplex or TDD cross-link interference.

The term “xDiff” is not unique in recent arXiv usage. In materials informatics, the XRDiff paper uses “xDiff” for the PXRD-conditioned diffusion framework introduced for solving the inverse problem of powder X-ray diffraction to recover full 3D crystal structures (Segal et al., 12 Jun 2026). In mechanistic interpretability, the Delta-Crosscoder paper describes the standard crosscoder formulation for model diffing as “xDiff,” and then extends it with a delta-based objective for narrow fine-tuning regimes (Kassem et al., 16 Feb 2026). This suggests that “xDiff” functions as a label shared across distinct subfields rather than as a uniquely identifying method name. Within wireless systems and O-RAN, however, xDiff refers specifically to the online diffusion model for collaborative inter-cell interference management.