Dual Path Architecture

• Dual Path Architecture is a family of neural designs that splits processing into two distinct paths, enabling both feature reuse and new feature exploration.
• It is applied across fields such as image recognition, speech processing, and neuromorphic computing, where parallel processing enhances local and global context modeling.
• Empirical results demonstrate improved accuracy and efficiency by balancing complementary computations through adaptive routing and coordinated fusion.

Searching arXiv for the cited papers on dual-path architectures to ground the article in current research. Dual path architecture denotes a family of neural designs in which two coordinated processing pathways are maintained within the same model, but the precise meaning of “path” is domain-dependent. In some works it refers to residual-style feature re-use coupled with dense-style feature exploration, as in Dual Path Networks (Chen et al., 2017); in others it denotes factorized sequence modeling along two axes such as intra-/inter-chunk, time/frequency, or temporal/spectral dimensions (Chen et al., 2020, Zhao et al., 2023, Rika et al., 18 Dec 2025); elsewhere it describes parallel branches specialized for global versus local context, semantic versus structural cues, or public versus covert objectives (Lee et al., 2024, Li et al., 2019, Yu et al., 5 May 2026). The unifying idea is the explicit separation of complementary computations into two paths and the subsequent recombination of their representations or decisions.

1. Terminological scope and main lineages

The expression “dual path architecture” does not identify a single canonical topology. Instead, recent arXiv literature uses it for several recurring design patterns that share a two-stream inductive bias.

Lineage	Path definition	Representative paper
Residual–dense hybrid	feature re-use via addition + feature exploration via concatenation	(Chen et al., 2017)
Axis-factorized sequence modeling	local/global or time/frequency processing along different axes	(Chen et al., 2020, Zhao et al., 2023, Rika et al., 18 Dec 2025)
Parallel branch specialization	two branches for different modalities, scales, or tasks	(Li et al., 10 Oct 2025, Lee et al., 2024, Zhou et al., 2022)

Historically, the best-known early formulation is Dual Path Networks, which reinterpret ResNet and DenseNet within a higher-order RNN view and then combine a residual path $y^k$ with a dense path $x^k$ through

$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$

followed by $h^k = g^k(r^k)$ (Chen et al., 2017). In that formulation, the dual path is explicitly tied to the coexistence of shared and non-shared transformations.

A second major lineage appears in long-sequence speech modeling. The end-to-end speech separation DPTNet chunks a sequence into $D \in \mathbb{R}^{N \times K \times P}$ and alternates an intra-transformer over chunk length $K$ with an inter-transformer over chunk index $P$, thereby separating local and global context modeling (Chen et al., 2020). Related work extends the same principle from chunk structure to time–frequency structure: D2Former applies conformer blocks along time and frequency in the complex STFT domain (Zhao et al., 2023), and DPDFNet uses a dual-path RNN whose intra-path runs across frequency bins while the inter-path runs across time (Rika et al., 18 Dec 2025).

A third lineage is parallel specialization. In speaking style recognition, one path preserves the serial ASR$\rightarrow$style dependency while a second path performs frame-synchronous acoustic–linguistic interaction (Li et al., 10 Oct 2025). In panoramic activity recognition, one path is individual-to-global and the other is individual-to-social (Lee et al., 2024). In real-time object detection, one path maintains low-resolution semantics while the other preserves high-resolution detail (Zhou et al., 2022). This suggests that the term has broadened from a specific residual–dense topology into a general architectural principle for pairing complementary computations.

2. Core structural principles

Across these variants, several structural principles recur. The first is complementarity: the two paths are not intended to be redundant replicas, but to encode distinct inductive biases. Dual Path Networks explicitly assign one path to feature re-use and the other to new feature exploration (Chen et al., 2017). DDPNet makes the distinction spatial, using a standard $3\times3$ branch and a dilated $3\times3$ branch inside each Dual-Path Module to aggregate local and larger-scale context in semantic segmentation (Yang et al., 2020). DPNet for detection adopts an analogous division at the backbone level, with a low-resolution path for high-level semantics and a high-resolution path for low-level object details (Zhou et al., 2022).

The second principle is factorization of dependency structure. DPTNet for speech separation does not apply a single transformer over the full sequence; it splits modeling into intra-chunk and inter-chunk attention, which makes extremely long sequences tractable while retaining direct pairwise interaction within each path (Chen et al., 2020). DPDFNet adopts an equivalent decomposition over spectro-temporal tensors, with a spectral intra-path and a temporal inter-path (Rika et al., 18 Dec 2025). D2Former makes the factorization fully complex-valued and uses both dual-path self-attention and dual-path encoder/decoder processing (Zhao et al., 2023).

The third principle is coordinated fusion rather than mere coexistence. In DPN, the two path outputs are recombined inside every block (Chen et al., 2017). In SFANet, the attention map path gates the density path by

$x^k$0

so one path directly modulates the other before the density prediction (Zhu et al., 2019). In the serial-parallel SSR model, the two paths remain architecturally distinct during training and are combined only at the decision level (Li et al., 10 Oct 2025). In the LLM scaling architecture of 2026, both deep and wide paths are always evaluated and then mixed by learned per-token gates, yielding

$x^k$1

which exposes compute and capacity as separate routing dimensions (Frey et al., 28 May 2026).

A fourth principle is parameter efficiency through selective sharing. DPN shares residual-path channels while allowing dense growth (Chen et al., 2017). Looped-deep plus one-shot-wide transformer blocks scale compute with shared parameters in the deep path and capacity with unshared parameters in the wide path (Frey et al., 28 May 2026). DMP-SNN uses a high-dimensional sparse fast path and a low-dimensional dense slow memory path, replacing $x^k$2 recurrent structure with $x^k$3 memory coupling when $x^k$4 (Sun et al., 8 Dec 2025). This suggests that dual-path design is often a mechanism for separating parameter growth from representational growth.

3. Domain-specific instantiations

In speech and audio, dual-path architectures are especially prominent because speech naturally exhibits both long temporal structure and strong local or cross-band structure. The serial-parallel SSR model couples a serial Whisper–adaptor–LLM pipeline with a parallel Acoustic-Linguistic Similarity Module. The serial path captures utterance-level temporal dependency through ASR followed by style labeling, whereas the parallel path aligns acoustic frames and linguistic tokens by attention and computes multi-space cosine similarities for frame-synchronous style classification (Li et al., 10 Oct 2025). PDPCRN extends the older dual-path CRN idea by adding two parallel bottleneck branches—DPRNN plus self-attention, and depthwise convolution plus DPRNN—with bi-directional inter-branch interactions for multi-channel enhancement (Pan et al., 2023). D2Former makes the entire time–frequency pipeline complex-valued, using time and frequency paths in both conformer blocks and encoder/decoder modules, then combines a masking decoder and a spectral-mapping decoder (Zhao et al., 2023). DPDFNet places dual-path RNN blocks into the encoder of DeepFilterNet2, with bidirectional spectral modeling and causal temporal modeling (Rika et al., 18 Dec 2025).

In vision, the dual-path idea appears in several distinct forms. DPN combines residual and dense connectivity for image classification and then transfers successfully to Places365 and PASCAL VOC detection and segmentation (Chen et al., 2017). DPNet for real-time detection maintains a low-resolution path and a fixed $x^k$5-resolution high-resolution path connected by bi-directional fusion modules (Zhou et al., 2022). DPTNet for scene text detection runs local-window self-attention and depth-wise convolution in parallel inside every hybrid block, then adds channel and spatial bi-directional interaction modules (Lin et al., 2022). SFANet uses one path to predict attention maps and another to regress density maps for crowd counting (Zhu et al., 2019). The mammogram dual-path network separates texture/context processing from segmentation-derived shape modeling, using a Locality Preserving Learner and a Conditional Graph Learner (Li et al., 2019). SPDP-Net for panoramic activity recognition defines paths by granularity: individual-to-global and individual-to-social, both repeated over multiple transformer layers (Lee et al., 2024). MDFA-Net for battery RUL prediction uses a shallow multiscale branch and a deep encoder branch (Lv et al., 16 Dec 2025).

In scientific computing and operator learning, DPNO transfers the residual–dense dual-path concept to neural operators. One path evolves as

$x^k$6

while the other uses dense concatenation,

$x^k$7

before fusion by a post-processing network (Wang et al., 17 Jul 2025). This is a direct operator-level analogue of DPN rather than a metaphorical use of the term.

In neuromorphic computing, the Dual Memory Pathway architecture separates fast spiking dynamics from a slow low-dimensional memory state $x^k$8. The slow state obeys

$x^k$9

and re-enters the fast pathway through a memory current $$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$0 (Sun et al., 8 Dec 2025). Here the dual path is not spatial or modal but dynamical: sparse fast events versus dense slow memory.

In systems work, the meaning becomes infrastructural rather than neural. DualPath for agentic LLM inference introduces two KV-cache loading paths—storage$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$1prefill and storage$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$2decode$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$3prefill—to balance saturated storage NICs on prefill engines against underused resources on decoding engines (Wu et al., 25 Feb 2026). This suggests that “dual path architecture” can designate a systems-level dataflow topology as well as a network block.

4. Interaction, routing, and fusion mechanisms

A central differentiator among dual-path architectures is how the two paths communicate. Some designs fuse features directly. DPN fuses by addition plus concatenation semantics inside the block implementation (Chen et al., 2017). DDPNet concatenates the outputs of a standard and a dilated branch, then reuses them through dense connectivity in subsequent layers (Yang et al., 2020). SFANet uses multiplicative spatial gating from the attention path to the density path (Zhu et al., 2019).

Other designs rely on structured cross-attention or similarity. In the SSR serial-parallel architecture, the parallel path uses frame–token attention

$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$4

to align text tokens with audio frames, then computes cosine similarities in $$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$5 subspaces before a transformer classifier (Li et al., 10 Oct 2025). PDPCRN adds channel interaction from the convolutional branch to the self-attention branch and spatial interaction in the reverse direction, implemented with small convolutional gating modules (Pan et al., 2023). DPTNet for scene text detection uses the same bidirectional pattern, with channel interaction from the convolution path to the attention path and spatial interaction from the attention path back to the convolution path (Lin et al., 2022).

Some models keep the paths separate until late fusion. The SSR system trains with a dual loss

$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$6

with $$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$7 and $$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$8, and then fuses posteriors during inference as

$$x^{k} \triangleq \sum_{t=1}^{k-1} f_t^{k}(h^t),\qquad y^{k} \triangleq y^{k-1} + \phi^{k-1}(y^{k-1}),\qquad r^{k} \triangleq x^{k} + y^{k},$$9

(Li et al., 10 Oct 2025). SPDP-Net similarly couples its two paths through shared individual representations and multi-task supervision rather than explicit cross-attention between global and social tokens (Lee et al., 2024).

Routing can also be discrete or adaptive. The covert semantic communication framework shares a ResNet-34 backbone across an Explicit path and a Stego path, but equips each residual block with Gumbel-Softmax gates $h^k = g^k(r^k)$0 and $h^k = g^k(r^k)$1, allowing path-specific block execution while regularizing policy similarity and enforcing contrastive feature alignment (Yu et al., 5 May 2026). The compute/capacity LLM block learns per-token deep and wide gates and additionally defines a deep-share statistic

$h^k = g^k(r^k)$2

which makes routing directly interpretable (Frey et al., 28 May 2026). DualPath for LLM inference makes analogous routing decisions at the systems level: a global scheduler chooses whether a request should use the PE read path or the DE read path based on token load and storage queue state (Wu et al., 25 Feb 2026).

These variants show that dual-path architecture is not only about drawing two arrows in a block diagram. It also concerns the algebra of interaction: additive fusion, concatenative growth, multiplicative gating, attention-mediated exchange, posterior fusion, or explicit routing.

5. Empirical behavior, efficiency, and trade-offs

A recurrent empirical theme is improved accuracy–efficiency trade-offs. In image classification, DPN-92 surpasses ResNeXt-101 (32×4d) on ImageNet-1k with 37.8M parameters and 6.5 GFLOPs versus 44.3M and 8.0 GFLOPs, and DPN-98 matches or exceeds ResNeXt-101 (64×4d) with about 26% fewer parameters and about 25% fewer FLOPs (Chen et al., 2017). On Places365, PASCAL VOC detection, and segmentation, DPN-92 also exceeds DenseNet, ResNet, and ResNeXt baselines (Chen et al., 2017).

In speaking style recognition, the serial-parallel dual-path model reaches 87.73% accuracy versus 67.34% for the OSUM baseline, while reducing parameter count from about 7.35B to about 0.855B, i.e., an 88.4% reduction. Its ablations show that the full dual path outperforms the serial-only ASR+STYLE path (73.83%) and the parallel-only ALSM path (84.17%), indicating complementarity rather than mere redundancy (Li et al., 10 Oct 2025).

In object detection, DPNet achieves 30.5% AP on MS COCO test-dev and 81.5% mAP on Pascal VOC 2007 with nearly 2.5M parameters, 1.04 GFLOPs, and 164 FPS and 196 FPS, respectively, for $h^k = g^k(r^k)$3 inputs. Its ablations attribute a +3.0 AP gain over a single low-resolution path to the dual-path backbone itself, before adding lightweight attention modules (Zhou et al., 2022). For scene text detection, DPTNet-Normal reaches 88.2 F1 on MSRA-TD500 and 85.5 F1 on CTW1500 in the reported ablations, outperforming ResNet backbones while remaining competitive in speed (Lin et al., 2022). DDPNet reports 75.3% mIoU at 52.6 FPS on Cityscapes test with only 2.52M parameters (Yang et al., 2020).

In speech enhancement, DPDFNet-2, -4, and -8 show monotonic improvement in the PRISM aggregate metric from 0.95 to 1.00 as the number of dual-path blocks increases, with DPDFNet-4 at 2.84M parameters and 2.36G MACs, and DPDFNet-8 at 3.54M parameters and 4.37G MACs (Rika et al., 18 Dec 2025). D2Former reaches state-of-the-art VoiceBank+Demand results with 0.87M parameters and benefits measurably from both complex dual-path attention and dual-path encoder/decoder design (Zhao et al., 2023). DPTNet for monaural separation reaches 20.6 dB SDR on WSJ0-2mix (Chen et al., 2020).

In activity recognition, SPDP-Net reports 46.5% overall F1 on JRDB-PAR, and the authors attribute gains over hierarchical baselines to spatio-temporal proximity modeling plus the dual-path transformer (Lee et al., 2024). In operator learning, DPNO reports relative improvements of over 30% on some standard PDE cases, with Burgers and Darcy experiments showing substantial relative $h^k = g^k(r^k)$4 error reductions over baseline DeepONet and FNO implementations (Wang et al., 17 Jul 2025). In neuromorphic sequence modeling, the dual memory pathway architecture achieves competitive or better accuracy with 40–60% fewer parameters than equivalent state-of-the-art spiking neural networks and yields more than a 4× increase in throughput and over a 5× improvement in energy efficiency in the proposed hardware stack (Sun et al., 8 Dec 2025). At the systems level, DualPath improves offline agentic LLM inference throughput by up to $h^k = g^k(r^k)$5 and online serving throughput by an average factor of $h^k = g^k(r^k)$6 without violating SLOs (Wu et al., 25 Feb 2026).

These results do not imply that dual-path design is uniformly cheaper. Some models increase MACs or require more complex schedulers and fusion policies. DPDFNet’s MACs rise from 0.36G in the baseline to 4.37G in DPDFNet-8 (Rika et al., 18 Dec 2025), and DualPath’s systems gains depend on a CNIC-centric traffic manager plus a global scheduler (Wu et al., 25 Feb 2026). The empirical record instead suggests that dual paths often improve the balance among accuracy, parameter count, FLOPs, latency, or memory bandwidth by allocating different roles to different paths.

6. Limitations, misconceptions, and future directions

A common misconception is that “dual-path architecture” names a single transferable block type. The literature does not support that view. In DPN it means residual plus dense connectivity (Chen et al., 2017); in DPTNet it means intra- and inter-transformer sequence modeling (Chen et al., 2020); in SSR it means serial ASR-driven reasoning plus parallel acoustic–linguistic interaction (Li et al., 10 Oct 2025); in covert semantic communication it means public and steganographic routing policies over shared residual blocks (Yu et al., 5 May 2026). The term is therefore best understood as a design family rather than a unique blueprint.

Another misconception is that the second path is always a lightweight auxiliary head. Some works make the two paths highly asymmetric, but others give both paths substantial modeling power. The wide path in the compute/capacity LLM block has its own enlarged FFN, while the deep path is recursively applied $h^k = g^k(r^k)$7 times (Frey et al., 28 May 2026). In DPDFNet, both spectral and temporal recurrent paths are integral to every dual-path block (Rika et al., 18 Dec 2025). In SPDP-Net, both individual-to-global and individual-to-social paths are repeated across all DPATr layers (Lee et al., 2024).

Several limitations recur across papers. DPN notes that the residual–dense split and growth-rate choices are heuristic and that the design space is not exhaustively explored (Chen et al., 2017). The SSR serial path remains vulnerable to ASR errors, and the fixed fusion weights $h^k = g^k(r^k)$8, $h^k = g^k(r^k)$9 may limit robustness; the paper explicitly mentions dynamic weighting as a possible future improvement (Li et al., 10 Oct 2025). SPDP-Net couples its paths only through shared individual features rather than explicit cross-attention between global and social tokens (Lee et al., 2024). DPTNet for speech separation acknowledges that dual-path chunking makes some interactions indirect rather than fully global (Chen et al., 2020). DPDFNet shows that higher dual-path depth improves quality but can violate real-time constraints on smaller NPUs (Rika et al., 18 Dec 2025). DualPath for inference assumes disaggregated prefill/decode serving with physically separated compute and storage networks (Wu et al., 25 Feb 2026).

Future directions in the surveyed literature are correspondingly diverse. The SSR work proposes extending the serial-parallel bimodal architecture to emotion recognition and sound event detection (Li et al., 10 Oct 2025). DPNO suggests applying dual-path composition to additional neural operator families and calls for theoretical analysis of approximation and generalization (Wang et al., 17 Jul 2025). The dual memory pathway paper proposes dual fast/slow abstractions as a broader co-design template for real-time neuromorphic systems (Sun et al., 8 Dec 2025). The adaptive covert communication model points toward broader use of dual-path routing under joint accuracy–privacy or accuracy–covertness objectives (Yu et al., 5 May 2026). The LLM compute/capacity work suggests that token-wise allocation between deep and wide paths can make scaling laws more structurally explicit (Frey et al., 28 May 2026).

Taken together, the literature indicates that dual-path architecture is best understood as a controlled decomposition of learning or dataflow into two complementary streams with a designed mechanism of interaction. The specific axes vary—feature re-use versus exploration, local versus global context, serial versus parallel multimodal reasoning, high versus low resolution, fast versus slow memory, or compute versus capacity—but the central research question remains the same: how to split heterogeneous inductive biases across two paths and recombine them more effectively than a monolithic single-path model.