Utility-Aware Adaptive Downlink DM-RS Allocation from Uplink CSI for Lightweight O-DU dApps

arXiv:2609.22270 · eess.SP, cs.NI · Submitted 2026-09-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Utility-Aware Adaptive Downlink DM-RS Allocation from Uplink CSI for Lightweight O-DU dApps".

Dev: Here's the gist of the paper. Adaptive demodulation reference signal (DM-RS) placement in 5G New Radio (NR) trades pilot overhead against channel-tracking robustness, making a single static density suboptimal across heterogeneous mobility and propagation regimes<ref:2609.22270#pg2> This paper studies a lightweight per-UE controller intended for O-DUside distributed application (dApp) execution<ref:2609.22270#pg2>.

Rosa: The Core Problem and Approach The central engineering problem involves selecting a reference signal pattern whose tracking benefit justifies its overhead under the current channel state, rather than just identifying the most frequent oracle class<ref:2609.22270#pg2>.

Dev: The authors address this by developing a lightweight per-UE controller that consumes eleven deployable uplink-CSI statistics to select among six standards-derived PDSCH DM-RS configurations<ref:2609.22270#pg2>.

Rosa: This controller is an MLP with dimensions 11 → 128 → 64 → 32 → 6, containing 12,070 parameters<ref:2609.22270#pg4>.

Title and authors: Dev: The core idea here is that instead of just trying to guess the 'best' pattern based on what we used to call oracle class accuracy, they are optimizing for the actual communication utility of the system under those eleven uplink CSI statistics.

Rosa: That makes sense because if we only care about getting the perfect classification, we might pick a pattern that looks good in a test but is actually terrible for how much data we can actually send reliably in the air.

Taro: So, they are trying to bridge that gap between theoretical accuracy and actual performance under real channel conditions.

Dev: They achieve this by using an MLP, which is a multilayer perceptron with layers like eleven to one hundred twenty-eight to sixty-four back down to thirty-two and finally six outputs. This model takes those eleven uplink CSI features and maps them onto a utility target.

Rosa: And what they're optimizing for isn't just the classification itself, but this objective function called U SoftCE, which is their way of valuing the communication utility over just getting the right class label.

Taro: So, it’s not about being perfectly correct on a label; it’s about maximizing how much data flows through the system efficiently given the current channel state.

The paper's summary: Rosa: They summarize this by saying they developed this per-UE controller that uses those eleven deployable uplink CSI statistics to pick one of six PDSCH DM-RS configurations derived from standards, rather than just picking the most common one.

Dev: The key improvement they introduce is moving away from exact class accuracy by optimizing for communication utility instead, which they call the U SoftCE objective. They minimize the mean square error and then add this soft classification term that weighs how useful each action is in terms of throughput.

Taro: It sounds like they’re saying that the cost of making a mistake in controlling how we send signals is measured by how much communication utility we lose, not just whether we picked the right pattern for a theoretical test.

Rosa: And the numbers show this has real results. Over two hundred ninety-four large-scale operating conditions with seven thousand five hundred six observations, this U SoftCE policy improves throughput over their best fixed training-selected pattern by about one point thirty, zero point eighty-six, and one point thirteen percentage points on three different scenarios <ref:2609.22270#pg1>.

Dev: That’s a significant jump in performance when you look at those specific splits: indistribution, CDL channel out of distribution, and that three hundred kilometers per hour mobility out of distribution split.

The paper's improvements: Taro: So what’s the actual benefit here for someone who needs to deploy this? It suggests we can adapt our downlink reference signal allocation per user based on those eleven uplink CSI measurements to select one of six standards-derived PDSCH DM-RS configurations.

Rosa: That means we get real-time adjustment of how we balance pilot overhead against how robustly the channel is being tracked, which is exactly what you need when mobility and propagation are changing constantly.

Dev: The controller itself is designed to be lightweight, it’s an MLP with twelve thousand seventy parameters, but the important part for deployment is that it only consumes those eleven deployable uplink CSI statistics for online inference.

Taro: And the paper shows that this isn't just a theoretical exercise; they tested it using an emulated O-DU prototype paired with a native Linux dApp prototype, and they even have a separate C++ inference kernel that measures dense network performance alone.

Conclusion: Rosa: So, to wrap up, this paper on "Utility-Aware Adaptive Downlink DM-RS Allocation from Uplink CSI for Lightweight O-DU dApps" shows that using communication utility as the learning target instead of exact oracle class recovery is a much better way to approach this problem.

Dev: It achieves higher throughput than standard cross-entropy methods, but they are careful to point out that there’s still a gap between accuracy and throughput—for instance, cross-entropy gets top one accuracies around forty five percent on some splits, while the U SoftCE gets them down to around twenty eight percent.

Taro: It seems the main limitation they flag is that those six actions are standards-derived under a specific PDSCH Type A configuration, and they're testing it on an emulated O-DU prototype instead of a full production system.

Rosa: But the results in the large-scale trajectory disjoint study are pretty compelling, showing gains like one point thirteen percentage points on those key splits.

Dev: It supports using communication utility optimization as the learning target for adaptive downlink DM-RS selection rather than just trying to recover an oracle class label. We also see that other loss variants, like U MSE, show that the utility-aware formulation is more robustly supported than any single loss variant.

Taro: And the fact that they have a native C++ implementation and a flexric based emulated O-DU prototype means this controller is computationally compatible with running in lower-layer distributed applications.

Rosa: So, this work fundamentally revisits adaptive downlink DM-RS selection as a communication utility optimization problem instead of just an oracle class classification problem. It’s a practical tool for making those systems more robust to changing conditions out there.

Seungwon Min

Republic of Korea

eess.SP, cs.NI

Submitted: 2026-09-10

Updated: 2026-10-31

Comments: 4 pages, 3 figures, 2 tables. Includes large-scale trajectory-disjoint evaluation, independent bit-accurate NR LDPC validation, and a FlexRIC-based emulated O-DU/native C++ dApp prototype. Extended follow-up "Receiver-Aware DM-RS Selection With Uplink CSI Statistics for O-RAN xApps" submitted to IEEE WCL, 31 Oct 2026. v2 adds notice only; technical content unchanged

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 78/100

The gist: Adaptive demodulation reference signal (DM-RS) placement in 5G New Radio (NR) trades pilot overhead against channel-tracking robustness…

Key concepts

DM-RS Placement
This refers to where and what type of reference signals are placed in the downlink. The paper investigates how to adapt these placements dynamically based on real-time channel conditions rather than using a single, fixed strategy. This adaptation is crucial for maintaining good signal quality when users move or the environment changes.
U-SoftCE Objective
This is a custom loss function used to train the AI controller. Instead of just trying to correctly identify the reference signal type (exact accuracy), this objective focuses on maximizing how well the communication link performs in reality (communication utility). It balances minimizing signal error with a soft classification term, prioritizing practical performance over theoretical labeling.
Uplink CSI Statistics
These are specific measurements of the radio channel obtained from the user's uplink transmission. The controller uses eleven of these statistics as input features to make its decision. By using these deployable statistics instead of complex or unavailable data, the system remains lightweight and practical for real-time use in distributed applications.

Terminology

Summary

Adaptive demodulation reference signal (DM-RS) placement in 5G New Radio (NR) trades pilot overhead against channel-tracking robustness, making a single static density suboptimal across heterogeneous mobility and propagation regimes<ref:2609. p. 022270#pg2> This paper studies a lightweight per-UE controller intended for O-DUside distributed application (dApp) execution<ref:2609. p. 022270#pg2>.

The Core Problem and Approach

The central engineering problem involves selecting a reference signal pattern whose tracking benefit justifies its overhead under the current channel state, rather than just identifying the most frequent oracle class<ref:2609.22270#pg2>. The authors address this by developing a lightweight per-UE controller that consumes eleven deployable uplink-CSI statistics to select among six standards-derived PDSCH DM-RS configurations<ref:2609. p. 022270#pg2>. This controller is an MLP with dimensions 11 → 128 → 64 → 32 → 6, containing 12,070 parameters<ref:2609. p. 122270#pg4>.

The Utility-Aware Learning Objective

The paper introduces the U-SoftCE objective to move beyond exact class accuracy by optimizing for communication utility rather than class identity. This objective minimizes LMSE, which is then augmented with a soft classification term: LU-SoftCE = LMSE + α"−X p qp log softmax(uˆ)p, where α = 0.1. This approach is motivated by the fact that the cost of an error in PHY control is determined by link utility rather than class identity<ref:2609. p. 022270#pg2>.

Evaluation and Performance Gains

The controller was evaluated over 294 large-scale operating conditions with 7,056 observations, and the U-SoftCE policy significantly improves throughput over a training-selected best fixed pattern by 1.30, 0.86, and 1.13 percentage points on indistribution, CDL channel-OOD, and 300-km/h mobility-OOD splits respectively. p. 0 The independent bit-accurate Sionna NR LDPC validation retained a +1. p. 021-point gain with a 95% cluster-bootstrap interval of [0. p. 027, 2.13].

Key Findings and Limitations

The results support the use of communication utility rather than exact oracle-class recovery as the learning target. While U-SoftCE achieves higher throughput than standard Cross-Entropy (CE), the accuracy-throughput mismatch is pronounced, with CE-MLP obtaining top-1 accuracies of 45. p. 20%, 44.3%, and 42.4% compared to U-SoftCE's 28.0%, 34.0%, and 31.0%. Limitations include the fact that the six actions are standards-derived under a specific PDSCH Type-A configuration, and the evaluation uses an emulated O-DU prototype rather than a production OAI/srsRAN system. p. 2

Conclusion

A 12.5k-parameter controller using only deployable UL-CSI statistics selects among six standards-derived PDSCH DM-RS configurations<ref:2609. p. 322270#pg2>. In a large-scale trajectory-disjoint study, U-SoftCE significantly outperforms the training-selected best fixed pattern on ID, CDL channel-OOD, and 300-km/h mobility-OOD splits. p. 3 A separate bit-accurate NR LDPC validation retains a +1. p. 321-percentage-point gain with a 95% cluster-bootstrap interval excluding zero. p. 3 The accuracy-throughput mismatch shows why exact pattern labels are a poor surrogate for PHY utility, while the U-MSE ablation indicates that the broader utility-aware formulation is more robustly supported than any single loss variant. p. 3 A native C++ implementation and FlexRICbased emulated O-DU prototype demonstrate that the controller is computationally compatible with lower-layer dApp execution. p. 3 The paper revisits adaptive downlink DM-RS selection as a communication-utility optimization problem rather than an oracle-class classification problem<ref:2609. p. 322270#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can adapt downlink DM-RS allocation per user based on eleven deployable uplink CSI statistics to select one of six standards-derived PDSCH DM-RS configurations. This allows for real-time adjustment of pilot overhead against channel tracking robustness under varying mobility and propagation regimes.

  2. The controller utilizes a multilayer perceptron with dimensions 11 to 128 to map the eleven deployable uplink CSI features to a utility-aware objective function, specifically the USoftCE loss. This enables the system to select an action that maximizes communication utility rather than optimizing for exact oracle class accuracy.

  3. The system can achieve a significant throughput improvement over a training-selected best fixed pattern by 1.30, 0.86, and 1.13 percentage points on ID, CDL channel-OOD, and 300 km/h mobility-OOD splits in large-scale trajectory-disjoint evaluations. This demonstrates that utility-aware learning substantially outperforms hard oracle class recovery for PHY control tasks.

  4. The implementation can be deployed as a native C++ controller requiring only 5.59 microsecond mean inference time, making it computationally compatible with lower-layer distributed application execution in an O-DU environment. This feasibility supports the integration of this adaptive logic into compact AI-RAN controllers via FlexRIC prototypes.

Sources

Related papers