Modality-Aware Dual-Track Surrogate in MGNAS

• The paper introduces MADTS, a surrogate-based acceleration scheme in MACC-MGNAS that reduces expensive full-model evaluations by approximating local fitness.
• MADTS employs Gaussian process surrogates for modality-specific blocks and randomized sampling for fusion blocks to balance local and global signals.
• Empirical results on the VulCE dataset demonstrate that MADTS cuts GPU-hours significantly while maintaining competitive F1-score performance.

Modality-aware Dual-Track Surrogate (MADTS) is the worker-side acceleration mechanism in MACC-MGNAS, the modality-aware cooperative co-evolutionary framework proposed for multimodal graph neural architecture search (MGNAS). In that framework, the central bottleneck is that each candidate architecture would otherwise require decoding into an MGNN, training, and validation; MADTS is introduced to reduce evaluation cost and accelerate worker-side evolution while preserving the Coordinator’s global evaluation as the ground truth (Wang et al., 23 Sep 2025).

1. Position in MACC-MGNAS and problem addressed

MADTS appears inside a three-part system composed of MACC, MADTS, and SPDI. MACC is the modality-aware cooperative co-evolution framework; MADTS is the surrogate-based acceleration mechanism; SPDI is the similarity-based population diversity indicator used for adaptive exploration and exploitation control. The overall setting is multimodal graph neural architecture search for vulnerability co-exploitation prediction, where heterogeneous and multimodal vulnerability data are processed by MGNNs, and where architecture design is difficult because modality-specific blocks and cross-modal fusion blocks must be coordinated at each layer (Wang et al., 23 Sep 2025).

The method is organized around a coordinator-worker structure. The coordinator maintains the global chromosome population

$$\mathcal{P}_t = \{\mathbf{C}_1, \mathbf{C}_2, \ldots, \mathbf{C}_N\},$$

decomposes each chromosome into modality-specific and fusion blocks, dispatches blocks to workers, receives elites, reassembles chromosomes, and performs global evaluation by decoding and evaluating complete MGNNs. The workers are multiple modality workers, one per modality, together with one fusion worker. MADTS operates at the worker side, not at the coordinator side, and is explicitly described as approximating local fitness of modality and fusion blocks so that fewer expensive full-model evaluations are required (Wang et al., 23 Sep 2025).

The rationale for MADTS is tied to a specific mismatch in multimodal cooperative co-evolution. Existing surrogate methods are described as assuming more homogeneous search spaces, whereas this search space is heterogeneous: modality-specific blocks have their own local semantics and local validation subsets, fusion blocks are structurally different and do not have an obvious standalone local fitness, and local block quality and global assembled-architecture quality are related but not identical. This suggests that a single surrogate over all candidates, or a purely local proxy, would be misaligned with the actual optimization objective. MADTS addresses that mismatch by combining local worker signals with coordinator feedback while leaving final authority to global evaluation (Wang et al., 23 Sep 2025).

Component	Role
MACC	Structural search framework with modality-aware decomposition
MADTS	Worker-side efficiency mechanism
SPDI	Coordinator-side diversity controller

2. Dual-track and modality-aware design

The term “dual-track” refers to two distinct routes corresponding to the two types of gene blocks created by MACC decomposition. Track 1 is the modality-specific block surrogate track. Track 2 is the fusion block surrogate/proposal track. This separation is not cosmetic; it follows from the claim that fusion blocks do not have a standalone local fitness, whereas modality-specific blocks do (Wang et al., 23 Sep 2025).

For modality-specific blocks, MADTS uses a Gaussian process surrogate. For each modality block $\mathbf{C}^{(m)}$,

$$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$

where $S^{(m)}$ is the modality-$m$ surrogate regressor and $\theta^{(m)}$ its parameters. This is an explicit surrogate-learning track with online archive updates and retraining. By contrast, for the fusion block $\mathbf{C}^{(\mathrm{fus})}$, the method states that “no standalone local fitness exists”; instead, candidate fusion blocks are generated via randomized sampling, their quality is assessed only through global fitness, and fusion archive updates rely solely on Coordinator feedback without retraining (Wang et al., 23 Sep 2025).

The term “modality-aware” denotes that the method preserves modality heterogeneity rather than collapsing all blocks into a single homogeneous problem. The full chromosome is decomposed as

$$\mathbf{C} = \big( \mathbf{C}^{(1)}, \mathbf{C}^{(2)}, \ldots, \mathbf{C}^{(M)}, \mathbf{C}^{(\text{fus})} \big),$$

and each modality has its own search subspace, its own surrogate, and its own local validation signal. The local fitness for modality $m$ is

$$F^{(m)}_{\mathrm{local}}(\mathbf{C}^{(m)}) = \mathrm{ModuleEval}\big(\mathbf{C}^{(m)}; \mathcal{D}_{val}^{(m)}\big).$$

Operationally, modality awareness is realized through separate gene groups, separate surrogates $\mathbf{C}^{(m)}$0, separate local datasets $\mathbf{C}^{(m)}$1, and separate treatment of the fusion block (Wang et al., 23 Sep 2025).

A common misconception is to treat MADTS as a replacement for coordinator evaluation. It is not. The paper distinguishes sharply between surrogate-assisted local guidance and the “true” full-architecture objective. MADTS filters and prioritizes local candidates; the coordinator’s global evaluation remains the authoritative fitness signal (Wang et al., 23 Sep 2025).

3. Mathematical formulation and score construction

The real optimization target is global architecture-level validation. For a complete chromosome $\mathbf{C}^{(m)}$2,

$\mathbf{C}^{(m)}$3

where $\mathbf{C}^{(m)}$4 is the decoded MGNN and $\mathbf{C}^{(m)}$5 is the validation dataset. The local modality score is auxiliary rather than final; it is used to update surrogates and guide worker-side evolution (Wang et al., 23 Sep 2025).

For modality workers, MADTS does not rely only on raw local scores. It defines a fused modality score that combines local and global evidence: $\mathbf{C}^{(m)}$6 The balancing weight is adaptive: $\mathbf{C}^{(m)}$7 The notation specifies estimated variance of local scores, estimated variance of global scores, and a small constant $\mathbf{C}^{(m)}$8 for numerical stability. The exact estimation procedure for these variances is not specified beyond that notation (Wang et al., 23 Sep 2025).

Candidate proposal in modality workers is acquisition-driven: $\mathbf{C}^{(m)}$9 Here $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$0 is the modality-specific search space and $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$1 is an acquisition function. The paper does not specify the exact form of $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$2, and it also does not provide the explicit GP kernel, GP training loss, uncertainty estimates, trust-region logic, or switching thresholds. The surrogate input is described only as “block encodings” (Wang et al., 23 Sep 2025).

Archive maintenance differs by track. For modality workers, the bounded archive is updated by fusing scores, removing duplicates, keeping top-$$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$3 elites, and retraining surrogates. For the fusion worker, archive updates depend solely on Coordinator feedback and occur without retraining. This asymmetry is one of the defining properties of MADTS as written (Wang et al., 23 Sep 2025).

4. Worker-side workflow and computational role

The MADTS-relevant workflow is generation-based. In each global generation $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$4, the coordinator decomposes chromosomes into modality and fusion blocks and dispatches them to workers. Modality workers compute local block evaluations on $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$5, incorporate global coordinator feedback, retrain their Gaussian process surrogates on updated archives, and use acquisition-driven proposal to choose promising next blocks. The fusion worker generates fusion blocks by randomized sampling because no standalone local fitness is defined. After $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$6 local-search steps, each worker returns top-$$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$7 elites to the coordinator (Wang et al., 23 Sep 2025).

The coordinator then reassembles the returned elites as

$$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$8

decodes the resulting full chromosomes into MGNNs, performs true global evaluation, and applies survivor selection for the next generation. MADTS therefore influences candidate proposal, worker-side screening, archive update, and elite retention, but the paper is less explicit about whether surrogate scores directly control parent selection, crossover acceptance, or mutation acceptance in the GA operators (Wang et al., 23 Sep 2025).

Its efficiency claim is straightforward. By approximating local block quality, workers do not need expensive full MGNN evaluation for every local candidate. Only top proposals and elites are returned for global evaluation, making the divide-and-conquer search practical. The complexity discussion attributes part of the per-generation cost to surrogate updates and acquisition over bounded archives: $$\hat{F}^{(m)}(\mathbf{C}^{(m)}) = S^{(m)}(\mathbf{C}^{(m)};\,\theta^{(m)}),$$9 per generation, and

$S^{(m)}$0

over $S^{(m)}$1 generations, with the explicit emphasis that $S^{(m)}$2. This is the structural reason MADTS matters: it reduces pressure on the dominant decoding/training/evaluation cost (Wang et al., 23 Sep 2025).

5. Empirical evidence, limitations, and scope

The strongest direct evidence for MADTS comes from ablation. On the VulCE dataset, the ablation compares “w/o MADTS” against full MACC-MGNAS. The “w/o MADTS” configuration reports Mean F1 $S^{(m)}$3, Best F1 $S^{(m)}$4, and GPU-hours $S^{(m)}$5. The full configuration reports Mean F1 $S^{(m)}$6, Best F1 $S^{(m)}$7, and GPU-hours $S^{(m)}$8. The paper’s own interpretation is that removing the surrogate has little effect on accuracy but degrades efficiency substantially, increasing GPU-hours from 3.00 to 4.87; it concludes that MADTS mainly contributes to reducing computational cost by guiding sampling and reducing expensive real evaluations (Wang et al., 23 Sep 2025).

At the framework level, MACC-MGNAS as a whole achieves an F1-score of 81.67% within only 3 GPU-hours on VulCE, outperforming the state-of-the-art competitor by 8.7% F1 while reducing computation cost by 27%. Those gains are not attributable to MADTS alone, because MACC and SPDI are also part of the integrated system. The ablation isolates MADTS primarily as an efficiency mechanism rather than an accuracy mechanism (Wang et al., 23 Sep 2025).

Several details remain unspecified. The paper does not provide the explicit GP kernel, GP training procedure, archive size beyond “bounded,” precise top-$S^{(m)}$9 management details, exact acquisition function $m$0, exact randomized sampling rule for fusion blocks, exact duplicate-removal definition, or exact construction of surrogate input features beyond block encodings. It also notes a broader limitation: “evolutionary NAS remains costly on very large graphs despite the surrogate acceleration.” A plausible implication is that MADTS makes cooperative co-evolution practicable, but does not remove the scaling difficulties of evolutionary NAS on large graph workloads (Wang et al., 23 Sep 2025).

6. Conceptual neighborhood and interpretation

MADTS belongs to a broader family of modality-aware and dual-track designs, but adjacent methods are not equivalent to it. “Modality-Aware SAM” proposes M-SAM, not MADTS, and is described as a modality-aware asymmetric surrogate optimization method with a dual-track update structure over a single model: a sharpness-aware surrogate on the dominant modality and standard loss optimization on non-dominant modalities (Nowdeh et al., 28 Oct 2025). “DT-GOL” is not multimodal, yet it is described as a strong conceptual and architectural analogue because it combines a surrogate-learning idea with a decoupled dual-track architecture separating stable delayed-ground-truth learning from speculative current adaptation (Wang et al., 22 Jun 2026). These cases show that “dual-track” can denote either structural separation of subproblems, as in MADTS, or separation of optimization/timing roles, as in M-SAM and DT-GOL.

In multimodal tracking, several architectures preserve modality-specific pathways and use lightweight cross-track interaction rather than uniform fusion. HMAD is described as an efficient, hierarchical, modality-aware RGB-D tracker with RGB and depth preserved in parallel and reweighted through a global aggregation-and-distribution mechanism (Xu et al., 24 Apr 2025). DMTrack uses a frozen backbone with dual adapters, one per modality branch, plus progressive cross-modal prompting (Li et al., 3 Aug 2025). MDTrack explicitly combines modality-aware fusion with decoupled RGB/X temporal propagation, which directly aligns with the “modality-aware” and “dual-track” parts of the term, though not with MADTS as a search surrogate (Wang et al., 10 Mar 2026). UASTrack, MTNet, VMDA, and DMTracker all preserve modality distinction while using compact adaptive modules rather than exhaustive full-model retraining (Wang et al., 25 Feb 2025, Hou et al., 24 Aug 2025, Xu et al., 30 Jun 2025, Gao et al., 2022).

The closest conceptual reading is therefore narrow rather than generic. In its strict sense, MADTS denotes the worker-side, modality-aware, dual-track surrogate method inside MACC-MGNAS: modality blocks are modeled with learnable Gaussian process surrogates informed by local and global signals, while fusion blocks are proposed by randomized sampling and judged only by coordinator feedback (Wang et al., 23 Sep 2025). In a broader editor’s sense, the term also sits within a larger design pattern in which modality heterogeneity is preserved, uniform optimization is avoided, and lightweight auxiliary mechanisms are used to guide expensive multimodal reasoning or search.