Decoupling then Perceive (DTP) Paradigm

• Decoupling then Perceive (DTP) is a research paradigm that first isolates entangled factors into coherent components before applying task-specific perception.
• It is applied across domains such as hyperspectral image fusion, low-light image super-resolution, night-time segmentation, tactile sensing, and vision-language reasoning.
• Empirical studies show that explicit decoupling boosts performance metrics while necessitating careful calibration and domain-specific regularization to address inherent challenges.

Decoupling then Perceive (DTP) denotes a recurrent research paradigm in which a system first separates entangled factors into components that are more semantically, physically, or functionally coherent, and only afterward performs the downstream perceptual, inferential, reconstructive, or control step on those decoupled variables. In the literature, the label spans unsupervised hyperspectral–multispectral image fusion, night-time semantic segmentation, soft tactile sensing, low-light image super-resolution, and vision-language reasoning. Taken together, these works present DTP not as a single canonical architecture, but as a family of methods organized around the same ordering principle: decouple first, then let task-specific perception operate on cleaner evidence (Du et al., 2024, Wei et al., 2023, Yang et al., 2024, He et al., 28 Mar 2026, Wu et al., 19 May 2026, Li et al., 1 Jul 2026).

1. Terminology and recurring structure

The acronym is used with closely related meanings across domains. In night-time segmentation, the queried phrase “Decoupling then Perceive” is explicitly identified with the paper’s “Disentangle then Parse”; there, “decoupling” and “disentangling” are treated as synonymous, and “Perceive” matches “Parse” (Wei et al., 2023). In other works, the same structural idea appears even when the paper title does not use the exact phrase, as in capability-staged vision-language post-training and Perceive-to-Reason (Wu et al., 19 May 2026, Li et al., 1 Jul 2026).

Paper	Decoupled factors	Downstream stage
MossFuse	modality-shared latent LR-MSI vs. modality-complementary spatial and spectral features	spatial–spectral aggregation and HR-HSI decoding
Night-time DTP	reflectance vs. illumination	illumination-aware semantic parsing
F3T	temperature, normal force, tangential forces	contact perception and control
LLISR DTP	low-frequency luminance vs. high-frequency texture	dual-path enhancement and semantic recomposition
P2R	Perceiver localization vs. Reasoner answering	final answer generation from annotated image and crops
Staged VLM post-training	visual perception, textual reasoning, visual reasoning	sequential capability refinement

A common pattern is visible across these instances. First, a confounded observation or capability set is partitioned into components that are intended to be more stable or more interpretable. Second, the model applies branch-specific processing, calibration, or role-specific optimization. Third, a downstream module fuses, reasons over, or controls from the resulting representation. This suggests that DTP is less a domain-specific trick than a design schema for reducing interference between factors that would otherwise be learned jointly.

2. Remote sensing and image reconstruction

In hyperspectral and multispectral image fusion, MossFuse instantiates DTP as a modality-decoupled, self-supervised HMIF pipeline. The key claim is that the shared component between HR-MSI and LR-HSI is a latent LR-MSI, whereas the complementary components are spatial details present primarily in HR-MSI and spectral details present primarily in LR-HSI. Accordingly, the decouple stage separates modality-shared features $F_Y^S, F_x^S$ from modality-complementary features $F_Y^C, F_x^C$, summarized by

$$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$

The perceive stage then aggregates the shared latent LR-MSI with complementary spatial and spectral cues through spatial- and spectral-modality aggregation blocks before decoding the HR-HSI estimate $\hat{X}$. The system is constrained by the HMIF forward models

$Y = X R,\qquad x = C X,$

and by self-supervised consistency terms requiring $Y \approx \hat{X}R$ and $x \approx C\hat{X}$. Its training objective combines modality clustering, modality reconstruction, modality constraint, and degradation estimation losses:

$$\mathcal{L}_{\text{total}}=\mathcal{L}_{MC}+\mathcal{L}_{MR}+\mathcal{L}_{MCT}+\mathcal{L}_{DE}.$$

On synthetic datasets under six degradations, MossFuse reports CAVE averages of PSNR 42.05, SSIM 0.990, SAM 6.57, and ERGAS 0.42; Harvard averages of 44.52, 0.992, 3.89, and 0.39; and NTIRE2018 averages of 48.53, 0.999, 0.79, and 0.04. On CAVE it is also reported at 0.54M parameters and 0.15 s test time with single-shot inference (Du et al., 2024).

The same ordering appears in low-light image super-resolution, but the variables are frequency-structural rather than cross-modal. The LLISR DTP framework first applies Frequency-aware Structural Decoupling (FSD), which uses a learnable wavelet transform

$\{x^{LL},x^{LH},x^{HL},x^{HH}\}=W_\theta(x)$

followed by simplex-constrained subband reweighting

$$\hat{f}=\alpha^\top\!\cdot f,\qquad \|\alpha\|_1=1,$$

and an explicit separation into a low-frequency, structure-preserving luminance component $F_Y^C, F_x^C$0 and a high-frequency texture component $F_Y^C, F_x^C$1. Semantics-specific Dual-path Representation (SDR) then processes luminance and texture independently: the luminance branch uses a Naka–Rushton-like monotonic nonlinearity, while the texture branch performs residual hierarchical denoising. Cross-frequency Semantic Recomposition (CSR) subsequently fuses both paths through channel attention, spatial attention, and a learned gate:

$F_Y^C, F_x^C$2

On RELLISUR, the full DTP model reports PSNR 23.15, SSIM 0.816, and LPIPS 0.124 at $F_Y^C, F_x^C$3, and 21.86, 0.814, and 0.253 at $F_Y^C, F_x^C$4. The main text explicitly states that the complete multi-term objective and detailed implementation settings are deferred to supplementary material, so the published description is intentionally incomplete on those points (He et al., 28 Mar 2026).

These two reconstruction-oriented variants share a substantive feature: the decoupled variables are aligned with a forward model or with a frequency structure that is already meaningful for the task. In MossFuse, this is the observation model of HR-MSI and LR-HSI. In LLISR, it is the separation between low-frequency luminance and high-frequency texture. This suggests that DTP is especially effective when the decomposition corresponds to an identifiable measurement process or signal subspace.

3. Night-time semantic segmentation

In night-time semantic segmentation, DTP is formulated explicitly as “Disentangle then Parse.” The central claim is that night scenes confound semantic content with insufficient, spatially nonuniform, and artificial illumination, so the network should not parse semantics from lighting-entangled features. The method therefore factorizes the input RGB image with a Retinex-style model,

$F_Y^C, F_x^C$5

where $F_Y^C, F_x^C$6 is light-invariant reflectance and $F_Y^C, F_x^C$7 is light-specific illumination. Reflectance is treated as the stable semantic carrier, while illumination is treated as both a nuisance factor and, in some categories, a useful cue (Wei et al., 2023).

The disentangling stage is implemented by the Semantic-Oriented Disentanglement (SOD) framework. A light-weight encoder–decoder $F_Y^C, F_x^C$8 predicts $F_Y^C, F_x^C$9 and $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$0; illumination outputs are disturbed during training by guidance noise,

$$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$1

and paired images are re-entangled into synthetic night scenes before a second weight-shared disentanglement pass. The training objective combines disentanglement consistency, Retinex reconstruction, structure-aware illumination smoothness,

$$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$2

and a semantic disentanglement loss $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$3 in which segmentation supervision acts as a disentanglement discriminator. The downstream parse stage is the Illumination-Aware Parser (IAParser), which fuses reflectance features $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$4 from a day-time segmentation backbone with illumination features $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$5 from a lightweight PPM using a learned attention mask

$$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$6

Predictions are then formed from the adaptive mixture of reflectance and illumination features.

Reported results show consistent gains across backbones and datasets. On NightCity validation, UPer-Swin improves from 58.4% mIoU to 61.2%, and on NightCity-fine from 61.1% to 64.2%. On BDD100K-night, UPer-Swin improves from 31.7% to 36.9% with night-only training, and from 48.0% to 53.1% with joint BDD100K-day+night training. The paper also reports that DTP can improve both night and day performance in full-time joint training, and describes the method as plug-and-play with negligible additional parameters (Wei et al., 2023).

A persistent misconception is that illumination should simply be removed. The segmentation paper does not treat illumination that way. Instead, it isolates illumination and then reintroduces it selectively through IAParser, on the premise that some semantically confused regions benefit from light-specific cues. DTP here therefore does not mean “discard the nuisance”; it means “control its influence after disentanglement.”

4. Tactile sensing and robotic contact perception

In the tactile sensing literature, DTP takes a strongly physical form. The F3T sensor realizes the paradigm by first isolating transduction pathways for temperature, normal pressure, and omnidirectional tangential force, and then mathematically inverting a calibrated sensitivity model to recover the underlying physical state. The sensor stack consists of a 1 mm ion-gel temperature sensing film, a 0.5 mm circular coaxial Halbach magnetic film, and a 5 mm floating-mountain multi-layer capacitor. Each layer is designed to suppress cross-talk: the ion-gel channel is experimentally near-insensitive to force, the capacitor is designed so tangential loads leave effective area and separation essentially constant, and the Halbach layer converts tangential shear primarily into $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$7 and $$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$8 modulation while normal force changes spacing (Yang et al., 2024).

The measured signals are defined as

$$F_Y^B = F_Y^S \oplus F_Y^C,\qquad F_x^B = F_x^S \oplus F_x^C.$$9

and the physical quantities as

$\hat{X}$0

Around an operating point, the model is linearized as

$\hat{X}$1

with an approximately block-diagonal sensitivity induced by physical decoupling. Temperature is estimated from an ion-gel current calibration $\hat{X}$2, inverted to obtain $\hat{X}$3. Normal force is estimated from the NE555 frequency with a piecewise linear law:

• $\hat{X}$4 for 0–1 N with $\hat{X}$5 N/Hz,
• $\hat{X}$6 for $\hat{X}$7 N with $\hat{X}$8 N/Hz.

Tangential force magnitude is derived from a sinusoidal magnetic response

$\hat{X}$9

with angle

$Y = X R,\qquad x = C X,$0

and components recovered by projection:

$Y = X R,\qquad x = C X,$1

The reported metrics emphasize decoupling fidelity. Temperature accuracy is stated as less than $Y = X R,\qquad x = C X,$2 in calibration and $Y = X R,\qquad x = C X,$3 versus a commercial thermocouple in static mixed-loading tests, with load-induced temperature error below 1%. Normal-force accuracy is 0.01 N for 0–1 N and 0.03 N above 1 N, with average error 0.03 N (at most 3%) across $Y = X R,\qquad x = C X,$4–$Y = X R,\qquad x = C X,$5, contrasted with 13.2% thermal drift in a commercial strain sensor at $Y = X R,\qquad x = C X,$6. Tangential-force resolution is 0.01 N in the initial phase, and handover detection is reported at approximately 0.4 s based on a rapid decrease in tangential force (Yang et al., 2024).

Here, DTP is neither feature disentanglement nor multimodal attention. It is an engineering doctrine: decouple the sensing physics first, then infer contact state, then feed the resulting variables into slip detection, grasp stabilization, impedance control, or task-level decision-making.

5. Vision-LLMs: inference-stage and training-stage decoupling

In fine-grained visual reasoning, DTP reappears as an explicit separation between where-to-look and how-to-answer. Perceive-to-Reason (P2R) formulates the task as a two-stage process executed by a single shared-parameter VLM. In the Perceiver role, the model predicts question-relevant boxes

$Y = X R,\qquad x = C X,$7

constructs an annotated image and crops,

$Y = X R,\qquad x = C X,$8

and then, in the Reasoner role, generates the final answer

$Y = X R,\qquad x = C X,$9

PRA-GRPO alternates between Perceiver-focused and Reasoner-focused GRPO updates using only final-answer supervision, with binary reward

$Y \approx \hat{X}R$0

and group-relative advantage

$Y \approx \hat{X}R$1

The method uses no ground-truth boxes, no IoU rewards, and no chain-of-thought supervision. Reported results include 93.2% on V-Star, 81.9% on HR-Bench-4K, and 80.5% on HR-Bench-8K for P2R-4B; on V-Star, self-generated boxes after PRA-GRPO are within 0.5% of oracle boxes (93.2% versus 93.7%). The paper also reports that DAPO led to one-box collapse and worse spatial performance than GRPO on V-Star Spatial, 85.8% versus 94.7% for the 4B model (Li et al., 1 Jul 2026).

A second VLM line applies the same principle at the curriculum level rather than at inference time. “From Seeing to Thinking” decomposes post-training into three sequential stages: visual perception, textual reasoning, and visual reasoning. Stage 1 uses perception-focused question answering synthesized from DOCCI captions and filtered so that a base VLM fails when reading the image but succeeds when reading the caption; Stage 2 uses ORZ-Math-13k for purely textual reasoning; Stage 3 uses visual reasoning data from CLEVR-Math, GeoQA-170K, Math PUMA, DocVQA, and ArxivQA. All stages are trained with RLVR and GRPO. The paper reports that 86.9% of Qwen3-VL-8B’s wrong answers on three visual math datasets are attributed by an external judge to perception errors made at the outset. Staged training yields 1.5% higher reasoning accuracy with 20.8% shorter reasoning traces than merged training, and the stage order matters: starting with visual reasoning degrades performance, whereas perception-first schedules perform best. For Qwen3-VL-8B, the staged model reports a 65.77% overall average versus 64.67% for merged training and 62.19% for the base model (Wu et al., 19 May 2026).

The two papers decouple at different loci. P2R hard-separates localization and answering during inference, while staged VLM post-training separates capabilities during optimization. Both are driven by the same diagnosis: perception, not merely long chain-of-thought, is often the bottleneck.

6. Empirical pattern, misconceptions, and limitations

Across these domains, DTP is supported not only by final benchmarks but also by ablations that directly test the necessity of the decoupling step. In MossFuse, removing $Y \approx \hat{X}R$2 drops CAVE performance from PSNR 42.05 and SAM 6.57 to 37.33 and 10.04, and replacing the aggregation blocks with simple concatenation+conv or cross-attention is also worse. In LLISR, the full FSD+SDR+CSR configuration reaches 23.15/0.817/0.124 at $Y \approx \hat{X}R$3 on RELLISUR, compared with a baseline of 18.51/0.765/0.387. In night-time segmentation, removing the disturb operation drops NightCity-fine performance from 64.2% to 62.5%. In P2R, replacing direct CoT with P2R inference raises Qwen3-VL-4B on V-Star from 81.7% to 89.0% even before role-aware RL training (Du et al., 2024, He et al., 28 Mar 2026, Wei et al., 2023, Li et al., 1 Jul 2026).

Several misconceptions are clarified by the surveyed work. First, DTP is not synonymous with unsupervised learning. MossFuse is self-supervised, night-time DTP is fully supervised, P2R uses answer-only RL, and staged VLM post-training relies on verifiable rewards rather than unlabeled data (Du et al., 2024, Wei et al., 2023, Wu et al., 19 May 2026). Second, DTP is not a claim that the decoupled variables are perfectly independent. Each domain uses explicit regularization, calibration, or architectural constraints because the decomposition is only approximately recoverable. Third, DTP is not identical to removing one factor and ignoring it; several methods decouple a factor precisely so that it can be reintroduced in a controlled manner, as with illumination in IAParser or complementary features in MossFuse.

The limitations are correspondingly domain-specific. MossFuse notes reduced accuracy under large sensor mismatch, non-Gaussian blur or noise, rare or highly non-linear spectra, and scene changes between modalities. Night-time segmentation states that perfect disentanglement remains hard in extremely dark scenes and that non-semantic illumination can still mislead the attention mask. F3T identifies elastomer creep, ion-gel dehydration, large-shear nonlinearity, and dense-array scaling as open issues. The LLISR paper leaves full losses and many implementation details to supplementary material and does not enumerate failure cases in the main text. P2R reports added latency from two-stage inference and reward sparsity from binary answer-only supervision, while staged VLM post-training highlights scale limits and dependence on fine-grained caption pipelines (Du et al., 2024, Wei et al., 2023, Yang et al., 2024, He et al., 28 Mar 2026, Li et al., 1 Jul 2026, Wu et al., 19 May 2026).

Taken together, these works indicate that DTP is best understood as a methodological family for controlling interference: modality redundancy in fusion, lighting entanglement in segmentation, cross-talk in tactile sensing, frequency entanglement in low-light reconstruction, and capability entanglement in VLMs. The unifying claim is not that one universal decomposition exists, but that explicit separation of confounded factors can make subsequent perception, parsing, reconstruction, reasoning, or control both more accurate and more diagnostically interpretable.