Two-Stage Decoupling Framework (TSDF)

• Two-Stage Decoupling Framework is a structural design principle that separates complex coupled problems into two sequential stages with an explicit inter-stage transfer.
• It is applied across diverse domains—from speech enhancement to forecasting—yielding measurable gains such as improved PESQ, AUC, and execution accuracy.
• Successful TSDF relies on carefully designed inter-stage transfer methods, whether through residual refinement, reward curriculum, or schema re-coupling, to stabilize training and enhance performance.

Searching arXiv for recent and representative papers on two-stage decoupling frameworks and closely related formulations. Two-Stage Decoupling Framework (TSDF) denotes a recurrent methodological pattern in which a difficult coupled problem is split into two sequential stages, with the first stage solving a simpler, lower-variance, or structurally cleaner subproblem and the second stage refining, reconstructing, or re-coupling the result under the full objective. The term is used explicitly in some papers, notably for variable-length glaucoma prognosis and federated learning, and appears in closely related forms such as “two-stage approach,” “two-stage reward curriculum,” “view-based two-stage framework,” and “semantic decoupling-based two-stage” across speech enhancement, robotics, text-to-SQL, forecasting, multimodal robustness, and circuit analysis (Song et al., 15 Sep 2025, Zhu et al., 2023, Li et al., 2020, Freitag et al., 5 Mar 2026, You et al., 2024, Xu et al., 2021, Hu et al., 19 Jan 2026, Garcia et al., 22 Apr 2026).

1. Definitional scope

In the literature, TSDF is not a single algorithmic family but a structural design principle. The common element is the separation of a coupled objective into two ordered subproblems with an explicit transfer mechanism. In monaural speech enhancement, the coupled target is complex-spectrum recovery, which is split into magnitude-first estimation and later complex refinement (Li et al., 2020). In robotics reinforcement learning, the coupled target is a reward containing both task and behavioral terms, which is split into task acquisition under a base reward and behavior shaping under the full reward (Freitag et al., 5 Mar 2026). In text-to-SQL, the coupled target is SQL generation over tightly linked schemas, which is split into dummy SQL generation over low-coupling views and reconstruction into executable SQL over original tables (You et al., 2024). In seasonal forecasting, the coupled target is near-horizon forecasting under long-range seasonal dependence, which is split into farther-future structure prediction and final-horizon forecasting conditioned on that proxy (Xu et al., 2021).

The scope is broader than machine learning. A topological decoupling of modified nodal analysis transforms circuit equations into a semi-explicit index-one differential-algebraic equation through graph-derived basis matrices (Garcia et al., 22 Apr 2026). Coupled transmission-distribution analysis likewise separates transmission and distribution solves and iteratively exchanges boundary variables at the point of common coupling (Krishnamoorthy et al., 2019). This suggests that TSDF is best understood as a general decomposition strategy for coupled systems rather than as a domain-specific architecture.

A crucial boundary of the concept is terminological. Some papers use the name directly, such as “Two-Stage Decoupling Framework for Variable-Length Glaucoma Prognosis” (Song et al., 15 Sep 2025) and “Federated Two Stage Decoupling With Adaptive Personalization Layers” (Zhu et al., 2023). Others instantiate the same pattern without the name. Conversely, “TSDF” in 3D vision very often means Truncated Signed Distance Function, not a two-stage framework (Splietker et al., 2019, Ding et al., 2024).

2. Canonical architecture and mathematical forms

A canonical TSDF has three ingredients: a simplified Stage 1 objective, an intermediate object passed to Stage 2, and a Stage 2 objective that either refines or reconstructs under the full task. In speech enhancement, CTS-Net makes this explicit. Stage 1 predicts only clean magnitude,

$|\tilde S^{cm}|=\mathcal{G}_{cm}(|X|;\phi_{cm}),$

then forms a coarse complex spectrum by reusing noisy phase. Stage 2 performs residual complex refinement,

$$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$

so the second stage does not restart from scratch but corrects a structured intermediate estimate (Li et al., 2020).

A different mathematical form appears in forecasting. The first stage predicts a farther future segment, and the second stage consumes that segment as an intermediate representation: $$\hat{\mathbf{x}}_f = f_2^*(\mathbf{x}_{his}, f_1^*(\mathbf{x}_{his})).$$ Here the transfer object is not a latent vector but an explicit future block intended to encode long-range seasonal structure (Xu et al., 2021).

Representation-space TSDFs use another form. In V-SQL, the original schema $D$ is replaced by a view-based schema $D'$, and generation proceeds in the simplified space: $y = f(x, I, S, D' \mid \theta).$ The stage interface is the dummy SQL produced over $D'$, which is then reconstructed into executable SQL over the original schema (You et al., 2024).

Reward-space TSDFs split optimization by curriculum rather than by representational space. In robotics, the stage-conditioned reward is

$$r_w = (1-w)\cdot r_{\text{base}} + w\cdot r_{\text{aux}},$$

with $w=0$ in Stage 1 and $w$ increased toward a target weight in Stage 2. The intermediate transfer is the policy, critic, and replay memory acquired under the task-only regime (Freitag et al., 5 Mar 2026).

These examples indicate that TSDF is not tied to one kind of variable. The inter-stage object may be a coarse estimate, a reward schedule, a dummy representation, a future-horizon proxy, a pretrained embedding, or a graph-based variable partition.

3. Major decoupling axes

The most common decoupling axis is objective decoupling. In speech denoising, the hard joint optimization over magnitude and phase is split so that magnitude is optimized first and both magnitude and phase are refined afterward (Li et al., 2020, Li et al., 2021). In personalized acoustic echo cancellation, the optimal stage order separates echo cancellation from later noise and interfering-speaker suppression; the best variant first models echo and then performs personalized speech enhancement (Zhang et al., 2023). In deepfake defense, interruption and poisoning are separated by an intensity-allocation mechanism: the first stage generates an interruption perturbation, and the second injects poisoning functionality into low-intensity regions so that the same perturbation both distorts forged outputs and disrupts attacker retraining (Zheng et al., 11 Aug 2025).

A second axis is representation decoupling. Glaucoma prognosis separates single-image representation learning from temporal modeling. Stage 1 learns per-image embeddings with masked autoencoding,

$$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$0

and Stage 2 aggregates variable-length visit sequences with a compact temporal module (Song et al., 15 Sep 2025). V-SQL likewise separates semantic parsing over low-coupling views from structural realization over the original normalized schema (You et al., 2024). In rainy-day attacks on vision-LLMs, semantic weakening in embedding space is separated from physically grounded weather synthesis; Stage 1 weakens decision boundaries with a low-dimensional global rain modulation, and Stage 2 searches a structured weather parameter space with multi-scale raindrops and illumination changes (Hu et al., 19 Jan 2026).

A third axis is heterogeneity decoupling. FedTSDP performs two rounds of clustering: first by inference-output similarity and then by model-weight distance, while also separating shared from personalized layers (Zhu et al., 2023). A fourth axis is topological or structural decoupling. Modified nodal analysis is split into differential, algebraic, and output blocks through basis matrices derived from graph contractions, spanning trees, and loop spaces (Garcia et al., 22 Apr 2026). Coupled T&D analysis separates transmission and distribution subproblems and enforces consistency by repeated exchange of PCC voltages and powers (Krishnamoorthy et al., 2019).

Domain	Stage 1	Stage 2
Speech enhancement (Li et al., 2020)	Magnitude estimation	Complex residual refinement
Robotics RL (Freitag et al., 5 Mar 2026)	Task-only reward	Full reward with behavior terms
Text-to-SQL (You et al., 2024)	Dummy SQL over views	SQL reconstruction
Forecasting (Xu et al., 2021)	Farther-future prediction	Target-horizon prediction
Glaucoma prognosis (Song et al., 15 Sep 2025)	Self-supervised image representation	Variable-length temporal aggregation
Federated learning (Zhu et al., 2023)	Inference-output clustering	Weight-space refinement

This diversity is important because it excludes a narrow reading of TSDF as only “coarse-to-fine.” In some cases the second stage is a residual corrector, in others a reward refinement, a schema re-coupling step, a temporal aggregator, or an algebraic reconstruction.

4. Optimization, training, and inter-stage transfer

TSDF does not imply a single training protocol. Several works use separate-then-joint optimization. CTS-Net first pretrains CME-Net on magnitude loss and then jointly trains both stages with RI-domain and magnitude-domain losses; the first stage is fine-tuned rather than frozen (Li et al., 2020). The DNS-Challenge system based on the same decomposition trains a magnitude stage and then jointly optimizes magnitude and complex refinement, again preserving a stage-specific magnitude term (Li et al., 2021). In personalized acoustic echo cancellation, the strongest strategy is to pretrain both stages separately and then fine-tune the full cascade without freezing, which outperforms both joint-from-scratch and frozen-stage alternatives (Zhang et al., 2023).

Other TSDFs remain fully sequential. The seasonal forecasting framework trains its two stages independently, and the second stage receives the true farther-future block during training but a predicted block at inference time (Xu et al., 2021). Glaucoma prognosis is also explicitly decoupled by stage, with self-supervised MAE pretraining followed by temporal modeling; the supplied text states that the two stages are trained separately, though it does not fully specify whether the encoder is frozen during Stage 2 (Song et al., 15 Sep 2025). V-SQL is not optimized end-to-end at all; it is a prompt-driven pipeline with a table-mapping module, dummy SQL generation, and SQL reconstruction under frozen LLM parameters (You et al., 2024).

A different pattern is curriculum transfer with sample reuse. In robotics RL, the reward curriculum is embedded inside standard TD3 and SAC updates, but the replay buffer stores $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$1 and $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$2 separately so that past samples can be reinterpreted under the current reward weight. The paper argues that this replay reuse is critical for stability and distinguishes the method from a naïve pretrain-then-finetune recipe (Freitag et al., 5 Mar 2026).

Inter-stage transfer is therefore central. It can occur through residual initialization, explicit feature concatenation, reward relabeling, schema-mapping rules, clustered model inheritance, or graph-derived coordinate transforms. A plausible implication is that TSDF succeeds when the first stage produces an intermediate object that is both easier to learn and genuinely useful to constrain the second stage, rather than merely delaying the full optimization.

5. Empirical effects and operating conditions

Across domains, TSDF-style decomposition is usually defended empirically against one-stage or non-decoupled baselines. In speech enhancement, CTS-Net reports average gains over the one-stage complex baseline GCRN of 0.18 PESQ, 4.29% ESTOI, and 1.38 dB SDR, and Stage 2 refinement over CME-Net alone yields 0.25 PESQ and 6.94% ESTOI on average (Li et al., 2020). The related DNS-Challenge system attains 3.38 overall MOS versus 3.21 for the NSnet2 baseline and ranks top-1 in the real-time track, which supports the practical value of a magnitude-first, complex-refinement-second design under real-time constraints (Li et al., 2021).

In robotics, the two-stage reward curriculum improves average $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$3 on DM Control from 0.637 to 0.690, MobileRobot success from 52.4% to 65.8%, and ManiSkill3 success at $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$4 from 62.1% to 97.6% (Freitag et al., 5 Mar 2026). In text-to-SQL, V-SQL improves execution accuracy over TA-SQL on Bird from 50.16 to 52.8 with GPT-4-turbo and from 43.8 to 45.4 with GPT-3.5-turbo, with the gains concentrated on harder join-heavy queries (You et al., 2024). In seasonal forecasting, the two-stage framework improves the MLP+MAR baseline at $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$5 from 1.454 to 1.399 MAPE, from 0.331 to 0.305 RMSE, and from 0.237 to 0.214 MAE (Xu et al., 2021).

In medical prognosis, explicit TSDF design yields strong gains while retaining variable-length data. On OHTS, the glaucoma TSDF reports ACC = 0.907 and AUC = 0.931, exceeding fixed-length LSTM baselines while retaining the full patient count; on GRAPE it reaches AUC 0.956 for PLR3 (Song et al., 15 Sep 2025). In federated learning, FedTSDP consistently improves over FedAvg, FedProx, Ditto, FedPer, and CFL under Non-IID settings, including 84.75% on CNN/CIFAR10 and 54.96% on CNN/CIFAR100 at Dirichlet $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$6 (Zhu et al., 2023).

The benefits are conditional. Stage ordering matters in personalized acoustic echo cancellation, where the best results come from echo cancellation first and personalized enhancement second, not the reverse (Zhang et al., 2023). Switch timing matters in reward curricula: premature switching harms performance, even if decoupling still beats direct full-reward training (Freitag et al., 5 Mar 2026). Auxiliary-horizon choice matters in forecasting, where $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$7 is best and larger values degrade MAPE (Xu et al., 2021). View design matters in V-SQL, where too many views make retrieval harder (You et al., 2024). In deepfake defense, the threshold $$(\tilde S_r^{cs},\tilde S_i^{cs}) = (\tilde S_r^{cm},\tilde S_i^{cm}) + \mathcal{G}_{cs}(\tilde S_r^{cm},\tilde S_i^{cm},X_r,X_i;\phi_{cs}),$$8 governing intensity separation controls a real trade-off between interruption and poisoning (Zheng et al., 11 Aug 2025). These results show that TSDF is not a free gain from staging alone; the stage interface must be well matched to the underlying coupling structure.

6. Terminology, misconceptions, and conceptual limits

A persistent misconception is that a two-stage framework is merely a network with two output heads. CTS-Net explicitly rejects this reading: its contribution is a two-stage cascaded architecture, not simply two heads on a shared trunk (Li et al., 2020). The same point appears in personalized acoustic echo cancellation, where a physically two-stage system without explicit task separation underperforms a properly decoupled stage assignment (Zhang et al., 2023). “Two-stage” therefore denotes ordered subproblems, not just parallel outputs.

A second misconception is that decoupling always means separate training. The evidence is mixed. Some frameworks are separately trained, some are separate-then-joint, and some are nonparametric prompt pipelines (Xu et al., 2021, Li et al., 2020, You et al., 2024). TSDF is architectural before it is procedural.

A third issue is acronym ambiguity. In 3D reconstruction and semantic scene completion, TSDF overwhelmingly denotes Truncated Signed Distance Function, as in “Directional TSDF” and “RGB-TSDF fusion” (Splietker et al., 2019, Ding et al., 2024). LiftNav, for example, is TSDF-guided in the geometric sense, not a method named “Two-Stage Decoupling Framework” (Schieber et al., 29 May 2026). Any encyclopedic use of TSDF therefore requires disambiguation between the framework concept and the volumetric geometry representation.

Finally, not every apparent two-stage method can be assigned to TSDF with equal confidence. One 2.5D IC PDN paper describes a “novel two-phase optimization flow” that first optimizes impedance in the frequency domain and then refines the PDN in the time domain using VVI (Duan et al., 2024). However, the supplied material also states that the full paper text is unavailable, so the broader TSDF characterization cannot be established beyond the abstract-level two-phase description. This illustrates a broader limit: TSDF is a useful analytical category, but its application should remain evidence-sensitive and should distinguish explicit framework definitions from looser analogies.