FlowSep-RF: Rectified Flow Matching for Single-Channel Radio-Frequency Signal Separation

The clean-endpoint formulation is theoretically sound and well suited to separation because it predicts a direct transport from mixture to target rather than an interference residual. The dual-domain architecture, which combines time-domain IQ features with frequency-domain structure and merges them through a learned scalar gate, is sensible for communication waveforms. The authors write that "FlowSep-RF treats the mixture as the initial endpoint and the SOI as the target endpoint. This gives a clean-endpoint rectified flow formulation aligned with the RF separation objective." This design choice is backed by the broader rectified-flow literature, which notes that "flows with straight paths bridge the gap between one-step and continuous-time models" and that nearly straight flows can work well with a single simulation step.

The experimental scope is narrow. The paper evaluates only a QPSK signal of interest with $N=2560$ complex samples under a simple additive model $y = s + \kappa b$ and a small set of interferer types. The blind receiver still uses a power-based SIR estimate that requires prior knowledge of nominal source and interference powers, and the training jitter is mild: "During training, the conditioning inputs are jittered as $\tilde{\kappa}=\kappa \xi$, with $\xi \sim \mathcal{U}(0.9,1.1)$, to improve robustness to estimation error." Baseline comparisons are also under-specified; all models are trained for only "20 epochs, and the checkpoint with the best validation performance is used for testing," leaving open how much training variance or early-stopping tuning affects the rankings. The paper does not describe exact data splits, baseline architecture depths, or compute budgets, which limits confidence in the head-to-head comparisons.

The internal evidence supports FlowSep-RF outperforming the tested baselines on the reported test grid, with Tables 1 and 2 and Figure 2 showing consistent BER and MSE gains across the SIR sweep. The authors note that "FlowSep-RF achieves the lowest average BER andMSE in all three settings using the same configuration." However, the comparisons are selective: iterative generative baselines are run with many steps but no step-ablation or compute-matched comparison is shown, and the discriminative baselines are in-house models with undisclosed parameter counts and training recipes. The paper does not evaluate against recent Transformer-based RF separators under identical conditions, so it remains unclear whether the gains come from rectified flow or from stronger architecture and better conditioning. The evidence is internally consistent but not conclusive about why the proposed method wins.

The paper supplies several implementation details but falls short of full reproducibility. The architecture is described as using "$d_{\mathrm{model}}=192$ hidden channels and eight residual blocks," and training hyperparameters are listed: "We train with AdamW using a learning rate of $2\times 10^{-4}$, batch size 96, and loss weights $\lambda_{\mathrm{rf}}=1.0$, $\lambda_{\mathrm{end}}=1.0$, $\lambda_{\mathrm{wf}}=0.10$, $\lambda_{\mathrm{sym}}=0.25$, and $\lambda_{\mathrm{spec}}=0.05$." Still, no code or data is released, random seeds are not reported, and baseline configurations are missing. The external RFChallenge CommSignal2 dataset is mentioned but its preprocessing is not described. Independent reproduction is therefore possible in principle but would require substantial guesswork, especially for the custom baselines and the exact power-estimation pipeline.