Decoupled-Then-Joint Training

• Decoupled-then-joint training is a staged optimization strategy that first decouples learning phases to mitigate interference and improve tractability before recombining them.
• The approach employs various re-coupling mechanisms—such as joint global objectives, coordinated multi-objective training, parameter merging, or physical calibration—to align separated modules.
• Empirical evidence across deep autoencoders, reasoning models, diffusion finetuning, and optical learning demonstrates its practical benefits in performance, efficiency, and robustness.

Decoupled-then-joint training denotes a family of staged optimization strategies in which learning is first separated across layers, modules, tasks, data subsets, views, timesteps, or physics surrogates, and is later re-coupled through end-to-end fine-tuning, coordinated multi-objective optimization, parameter-space merging, or system-level calibration. In the literature represented here, the term is not uniform: some works instantiate a strict sequential recipe, such as greedy layer-wise pretraining followed by joint optimization, whereas others realize weaker forms of re-coupling, including auxiliary coordination, model merging, or system-level calibration rather than a single later end-to-end objective (Zhou et al., 2014).

1. Conceptual scope and defining structure

At its most general, decoupled-then-joint training has two components. The first is a decoupled stage, in which some aspect of the learning problem is separated to reduce interference, improve tractability, or exploit specialized supervision. The second is a re-coupling stage, in which those separated parts are made to interact again so that the final model benefits from shared optimization pressure or unified deployment.

The literature shows that the object of decoupling varies substantially. In general reasoning, DeReason partitions training data by estimated reasoning intensity, assigns broad-coverage non-reasoning-intensive problems to supervised fine-tuning, and reserves difficult reasoning-intensive problems for RL (Hu et al., 11 Mar 2026). In video reward modeling, DeScore decouples Chain-of-Thought reasoning from scalar reward prediction by using an MLLM to generate explicit CoT and a dedicated discriminative scoring module with a learnable $[Reward]$ token and regression head (Wang et al., 7 May 2026). In diffusion-model finetuning, DeMe partitions timesteps into disjoint ranges and finetunes separate timestep specialists before merging them back into one model in parameter space (Ma et al., 2024). In PDE-constrained optical training, SP$^2$RINT relaxes metasurfaces into trainable transfer matrices and separately solves local inverse-design patches before progressively reimposing physical constraints and system-level calibration (Ma et al., 23 May 2025). In point-cloud self-supervision, Point-PQAE first generates two decoupled views and then jointly couples them through bidirectional cross-reconstruction with view-relative positional embedding (Zhang et al., 1 Sep 2025).

The corresponding re-coupling operators are equally heterogeneous. Some methods return to a single objective over all parameters; some use coordinated but non-identical losses on different subsystems; some perform weighted task-vector merging; and some reintroduce consistency only at the level of physical feasibility or posterior sampling. This suggests that “joint” is a family resemblance term rather than a single optimization primitive.

Decoupled object	Re-coupling mechanism	Representative paper
Layers in deep autoencoders	Joint global reconstruction objective with local regularizers	(Zhou et al., 2014)
Easy vs hard reasoning data	Sequential SFT then RL on complementary subsets	(Hu et al., 11 Mar 2026)
CoT reasoning vs scalar scoring	Dual-objective RL with auxiliary BT reward calibration	(Wang et al., 7 May 2026)
Timestep-specific diffusion tasks	Parameter-space task-vector merge	(Ma et al., 2024)
Relaxed optical transfer matrices and local patches	Progressive physical projection and system-level calibration	(Ma et al., 23 May 2025)
Two cropped point-cloud views	Bidirectional cross-reconstruction with VRPE	(Zhang et al., 1 Sep 2025)

2. Historical roots and canonical formulations

A canonical early formulation appears in the comparison between greedy layer-wise training and joint training for deep autoencoders. “Is Joint Training Better for Deep Auto-Encoders?” formalizes greedy layer-wise training as a decoupled procedure in which one shallow autoencoder is trained at a time and lower-layer parameters are fixed while higher layers are learned (Zhou et al., 2014). The paper argues that this makes upper layers optimize reconstructions of intermediate representations rather than the original input and introduces a joint objective

$$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$

where the first term is a global reconstruction loss in input space and the second term is a sum of layer-level regularizers (Zhou et al., 2014). The same study also includes explicit pretrain-then-joint variants, denoted $U$ and $UJ$, but its central conclusion is not that pretraining is always essential; rather, joint optimization with strong denoising or contractive regularization is the decisive factor (Zhou et al., 2014).

Later work clarifies that decoupling by itself can be a complete training philosophy rather than a prelude to re-coupling. Decoupled Greedy Learning decomposes CNN optimization into local supervised module-wise objectives and shows that strong CNN and ImageNet performance can be achieved without a later end-to-end fine-tuning phase (Belilovsky et al., 2019). DeInfoReg similarly blocks gradient flow between modules, assigns local objectives to each module, and remains decoupled throughout training, even though the modules are still forward-composed and trained for the same supervised task (Huang et al., 22 Jun 2025). These papers are important because they define the baseline alternative against which decoupled-then-joint strategies are typically justified.

A second canonical lineage arises in contexts where the decoupled stage is not layer-wise but objective- or task-specific. DeReason treats supervised fine-tuning and RLVR as complementary stages for general STEM reasoning and argues that the key design choice is the partition of training data across those stages (Hu et al., 11 Mar 2026). DeScore separates natural-language reasoning from scalar reward calibration in video reward modeling and then coordinates them through a later dual-objective RL stage (Wang et al., 7 May 2026). DeMe decomposes diffusion denoising by timestep range and then recovers a single deployable model through task-vector arithmetic (Ma et al., 2024). In all three cases, the central concern is not vanishing gradients, but interference between heterogeneous objectives or task slices.

3. Re-coupling mechanisms in mathematical form

The most literal form of re-coupling is a later joint objective over the full model, as in deep autoencoders. There, the decoupled layer-wise view is replaced by a global reconstruction objective measured in input space rather than at intermediate representations (Zhou et al., 2014). This is the closest match to the classical intuition behind decoupled-then-joint training.

A second form is coordinated multi-objective training with architecturally decoupled subsystems. DeScore keeps reasoning and scoring separate: an MLLM first generates CoT $\boldsymbol{o}$, a $[Reward]$ query token is appended, and a regression head predicts the scalar score $s=\Phi(\Theta(\mathcal{X})_{[Reward]})$ from $$\mathcal{X}=(\boldsymbol{v},\boldsymbol{c},\boldsymbol{o},[Reward])$$ (Wang et al., 7 May 2026). Stage 1 is a discriminative cold start with Bradley–Terry loss,

$$\mathcal{L}_{\mathrm{BT}}=-\mathbb{E}\left[\log \sigma(s^w-s^l)\right],$$

while Stage 2 uses a dual-objective RL formulation,

$^2$0

so that reasoning quality and final reward calibration are optimized by distinct but coordinated signals (Wang et al., 7 May 2026). This is not a monolithic end-to-end loss over a single homogeneous prediction head; it is a staged re-coupling of modules with different credit-assignment routes.

A third form is parameter-space consolidation. DeMe begins with a pretrained diffusion model, partitions the timestep interval $^2$1 into $^2$2 contiguous ranges, finetunes $^2$3 specialist copies $^2$4, and then defines task vectors $^2$5 relative to the common initialization $^2$6 (Ma et al., 2024). The unified model is obtained by

$^2$7

The paper interprets this merge as a weighted combination of timestep-specific objectives, rather than as a later joint SGD phase (Ma et al., 2024). In this variant, “joint” means consolidation into one parameter set for practical inference.

A fourth form is projection back into a constrained feasible set. SP$^2$8RINT first optimizes relaxed transfer matrices $^2$9 for a diffractive optical neural network, then periodically solves inverse design problems over physical permittivities $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$0 so that the realizable transfer matrices $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$1 approach the learned surrogates (Ma et al., 23 May 2025). The layerwise projection objective is

$$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$2

subject to Maxwell constraints, and the later system-level calibration step aligns the total stacked transfer matrix rather than each layer in isolation (Ma et al., 23 May 2025). Here the recoupling is not primarily statistical; it is physical and system-level.

A fifth form is cross-view coupling after view-level decoupling. Point-PQAE creates two independently cropped, normalized, and rotated point-cloud views, encodes them separately, and then reconstructs one from the other using view-relative positional embedding. The coupling is realized by bidirectional cross-reconstruction,

$$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$3

with each directional loss implemented as Chamfer distance between predicted target patches and target-view patches (Zhang et al., 1 Sep 2025). This is decoupled-then-joint at the data/view level rather than the optimization-schedule level.

4. Representative application domains and empirical signatures

In unsupervised representation learning, the deep-autoencoder results establish the basic tension. Pure greedy layer-wise training is a decoupled baseline, while joint training with appropriate denoising or contractive regularization generally improves generative modeling, reconstruction, and higher-layer representation quality (Zhou et al., 2014). The paper’s $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$4 and $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$5 variants show that pretrain-then-joint pipelines exist, but also that the presence of strong regularization during the joint stage matters more than the existence of pretraining alone (Zhou et al., 2014).

In post-training for reasoning models, DeReason gives a data-centric formulation of staged complementarity. On WebInstruct-Verified with 4B models, RL-only reaches AVG $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$6, SFT-only reaches AVG $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$7, sequential SFT then random-split RL reaches AVG $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$8, and the difficulty-aware split “SFT easy + RL hard” reaches AVG $$\mathcal{J}_{Joint}(\Lambda) = \sum_{x \in \mathcal{D}} \mathbb{E}_{Q(x, h_c^0,\ldots,h_c^N)}[\mathcal{L}(x,x_r)] + \sum_{i=1}^N \lambda^i \mathcal{R}^i(\Theta^i),$$9 (Hu et al., 11 Mar 2026). On Webscale-RL, RL-only reaches AVG $U$0, SFT-only reaches AVG $U$1, and the same difficulty-aware sequential pipeline reaches AVG $U$2 (Hu et al., 11 Mar 2026). These numbers support the paper’s claim that SFT and RL are complementary but should not share the same data indiscriminately.

In reward modeling for video generation, DeScore provides a clearer module-level example. It uses a 22K-pair training set, achieves in-domain preference accuracy $U$3, reaches $U$4 Acc w/o Tie on GenAI, and reaches $U$5 Acc w/o Tie on VideoGen-Bench, outperforming both discriminative and generative baselines reported in the paper (Wang et al., 7 May 2026). The ablations are especially diagnostic: a generative coupled SFT+GRPO baseline is weaker than the full think-then-score formulation, and Stage 2 works best when cold-start scoring and auxiliary BT calibration are both present (Wang et al., 7 May 2026). This suggests that later coordination can matter as much as the initial separation.

In diffusion finetuning, DeMe reports that decoupling timesteps into specialists and then merging them can improve generation quality without requiring multi-model deployment. On CIFAR10, the baseline FID $U$6 improves to $U$7 before merge and $U$8 after merge; on LSUN-Church, $U$9 improves to $UJ$0 before merge and $UJ$1 after merge (Ma et al., 2024). The merge stage is therefore not merely a compression device; in the reported DDPM settings it can improve upon the specialist ensemble.

In physics-constrained optical learning, SP$UJ$2RINT frames decoupled-then-joint training as a computational necessity. By optimizing relaxed transfer matrices between periodic physical projections, and by spatially decoupling inverse design into local patches, the method reports digital-comparable accuracy while being $UJ$3 times faster than simulation-in-the-loop approaches (Ma et al., 23 May 2025). This is a case where the later joint calibration is required for physical realizability, not only for predictive accuracy.

5. Contrasts, misconceptions, and adjacent but non-equivalent paradigms

A common misconception is that any staged or hybrid training procedure qualifies as decoupled-then-joint. Several prominent examples in the surrounding literature do not. D-Train in multi-domain learning is explicitly joint-then-decoupled: it first pre-trains a shared root model on all domains, then post-trains with shared-bottom multi-heads, and finally fine-tunes the heads with the backbone frozen to achieve “domain independence” (Wang et al., 2023). It therefore instantiates the opposite stage order.

A second misconception is that decoupling is always followed by a later joint optimization stage. Decoupled Greedy Learning and DeInfoReg remain decoupled throughout training, even though their modules are forward-composed and trained on the same supervised task (Belilovsky et al., 2019, Huang et al., 22 Jun 2025). In those frameworks, there is no later phase in which end-to-end gradients are restored.

A third misconception is that later interaction must be a training-time phenomenon. DDIS for inverse PDEs deliberately rejects joint coefficient-solution training, learns the coefficient prior and forward operator separately, and couples them only during posterior sampling at inference time (Lin et al., 30 Jan 2026). This is “decouple first, combine later,” but the combination is Bayesian sampling rather than joint parameter learning.

A fourth misconception is that decoupled-then-joint is always preferable to direct joint training. JointTuner is presented precisely as a critique of naive two-stage appearance-motion customization in video generation. Its claim is that sequentially optimizing appearance and motion in separate feature domains creates feature-domain mismatch and concept interference, so it replaces that schedule with simultaneous adaptive joint training using shared Adaptive LoRA and modality-specific losses (Chen et al., 31 Mar 2025). Likewise, E2E-GRec argues that the standard industrial GNN-precompute pipeline is a suboptimal decoupled baseline and that the recommendation objective should directly influence graph learning through end-to-end coupling (Xue et al., 25 Nov 2025).

These counterexamples matter because they locate decoupled-then-joint training within a wider design space. The real axes are not simply “staged” versus “end-to-end,” but which variables are separated, what kind of interference is being controlled, and what operator later restores consistency.

6. Diagnostics, limitations, and open directions

The literature does not yet provide a single general theory of when decoupled-then-joint training should outperform either pure decoupling or pure end-to-end optimization. One of the clearest diagnostic foundations comes from the Regime Change Hypothesis for ReLU-based networks. That work proves a local stability property of activation patterns outside measure-zero parameter and input sets, implying locally affine behavior within stable activation regions, and then measures per-iteration activation-pattern change and weight-update magnitude across MLPs, CNNs, and Transformer FFN blocks (Pérez-Corral et al., 9 Feb 2026). Its headline empirical finding is that activation-pattern changes decay 3 times earlier than weight-update magnitudes, suggesting an early structural/gating phase followed by a later quantitative refinement phase (Pérez-Corral et al., 9 Feb 2026). A plausible implication is that stage switches in future decoupled-then-joint algorithms could be triggered by measured regime stabilization rather than fixed epoch counts.

Another open direction concerns whether the “joint” stage should be true end-to-end optimization, iterative coordination, or only consolidation. Search-and-regress training in representation space provides a sharp example of this unresolved boundary. “Decoupling Search and Learning in Neural Net Training” separates evolutionary search over intermediate activations from later parameter learning via layerwise regression and shows that the resulting models can approach SGD on MNIST, CIFAR-10, and CIFAR-100, but it explicitly does not include a later joint fine-tuning stage or alternating search/learning loop (Vegesna et al., 13 Sep 2025). This suggests that decoupling alone can already be competitive, while also leaving open whether a later joint stage would close the remaining gap.

Across the surveyed applications, the major limitations recur in different forms. DeReason relies on heuristic difficulty scoring and an underexplored threshold between SFT and RL subsets (Hu et al., 11 Mar 2026). DeScore still uses a discrete CoT interface, so the scoring loss does not become a fully differentiable credit path through reasoning generation (Wang et al., 7 May 2026). DeMe’s merge stage is not a later joint SGD phase, which leaves open whether post-merge calibration could improve the unified model further (Ma et al., 2024). SP$UJ$4RINT depends on locality assumptions in metasurface interactions and on the quality of the periodic projection schedule (Ma et al., 23 May 2025). Point-PQAE, although strongly decoupled-then-joint at the view level, is not a literal two-stage optimizer, which highlights that the same term is still used for both schedule-level and objective-level designs (Zhang et al., 1 Sep 2025).

The cumulative record therefore supports a restrained conclusion. Decoupled-then-joint training is best understood not as a single algorithmic recipe, but as a recurring design principle: separate components when joint optimization is hindered by interference, data scarcity, physics constraints, or mismatched supervision; then restore consistency only through the form of coupling that the application can support. In some domains that form is a true joint loss, in others it is coordinated multi-objective optimization, task-vector merging, system-level calibration, or even inference-time posterior composition (Zhou et al., 2014, Wang et al., 7 May 2026, Ma et al., 2024, Ma et al., 23 May 2025).