Adaptive and Ordered Fusion

• Adaptive/Ordered Fusion is a class of methods that dynamically integrates multiple data sources using context-sensitive gating and order-aware processing.
• It employs adaptive gating and attention mechanisms to assign dynamic weights to each source, enabling improved performance under varying conditions.
• Applications span multi-modal perception, sequential recommendation, and continual learning, where preserving input order is critical for performance.

Adaptive and Ordered Fusion

Adaptive and ordered fusion encompasses a class of methodologies for integrating heterogeneous feature representations or model weights in a way that explicitly conditions the fusion process on the input data, task requirements, or system state. Unlike static concatenation or summation, adaptive fusion leverages learned or data-driven gating, attention, or weighting mechanisms to dynamically select, emphasize, or interpolate among sources. Ordered fusion further handles sequential or application-specific constraints, preserving, exploiting, or reasoning about the order or hierarchy among fused elements. Across modalities, domains, and tasks, these principles support improved representation, robustness, efficiency, and adaptability in deep learning, vision, language, perception, and sequential modeling.

1. Core Principles of Adaptive Fusion

At the mathematical core, adaptive fusion seeks to compute the fused representation $f$ as a context- or input-dependent function of $n$ sources $\{x_i\}_{i=1}^n$, instead of a static operation:

• Adaptive Gating: Assigns a per-source, per-channel, or spatially varying gate $g_i$ (often via a shallow network or attention mechanism), such that

$f = \sum_{i=1}^n g_i(x_1,\ldots,x_n) \odot x_i$

where $g_i$ may be normalized (e.g., softmax or sigmoid) and can depend on concatenations or projections of the sources (Mungoli, 2023).

• Attention-Based Fusion: Computes data-dependent compatibility scores (e.g., via $$e_i = \text{ReLU}(W_\text{att} f_\text{proj}(x_i) + b_\text{att})$$) and normalizes them to weights.
• Hybrid Structures: Combine local and global feature extractors with a gating mechanism (e.g., the combination of CNN and BiLSTM branches gated per-sample as in AVP-Fusion (Wen et al., 25 Dec 2025), or cross-modal attention with learned gates as in AG-Fusion (Liu et al., 27 Oct 2025)).
• Learned Alpha in Weight Fusion: In continual learning, adaptive fusion includes learning model weight interpolation factors (e.g., $$\theta_\text{fused} = \alpha \theta_\text{new} + (1-\alpha) \theta_\text{old}$$, with $\alpha$ optimized via task-driven loss, not statically set by class counts (Sun et al., 2024)).

Adaptive fusion is thus parameterized, differentiable, and sensitive to both local and global context, often explicitly mitigating issues of source unreliability, domain shift, or conflicting evidence via dynamic selection.

2. Ordered and Sequential Fusion Mechanisms

Ordered fusion architectures incorporate explicit mechanisms to exploit, encode, or respect the intrinsic sequence or hierarchical relationships among sources:

• Sequential Order Preservation: In time-aware modeling (e.g., sequential recommendation), the fusion operator and its gating maintain strict left-to-right order, as in the TASIF model which never permutes time-indexed user events, preserving both temporal and side attribute orderings throughout all fusion layers (Luo et al., 30 Dec 2025).
• Recurrent or Cumulative Gating: Ordered variants of AFF (R-AFF, Cumulative AFF) propagate fused context incrementally, such that at each location or timestep, the gating weights are functions of the previous context or previously fused elements—supporting auto-regressive or causal fusion (Mungoli, 2023).
• Multi-Stage, Layered, or Pipeline Fusion: Some architectures partition fusion into sequential stages: e.g., CLDyN for image fusion (Yang et al., 10 Apr 2026) first freezes a backbone, then applies closed-loop feedback from downstream tasks to insert semantic corrections at different layers, iteratively refining fused features and leveraging the dependency structure among feature levels.
• Per-Branch Ordered Fusion: In LAFB for multi-modal SOD, fusion is performed across both schemes (each tailored for a challenge, e.g. scale, center bias, ambiguity) and encoder layers, and the learned selection weights can reveal a hierarchy or order of branch usage relevant to the specific data distribution (Wang et al., 2024).
• Adaptive Frequency-Domain Fusion: Order-aware filtering is also embedded in spectral/frequency decoupling (AdaWAT (Wang et al., 21 Aug 2025)) and frequency domain filtering (adaptive frequency filter in TASIF (Luo et al., 30 Dec 2025)), where separate frequency bands, temporally or spatially, are dynamically fused based on input statistics or learned criteria.

3. Canonical Architectures and Mathematical Formulations

Multiple adaptive and ordered fusion designs have demonstrated effectiveness:

Framework	Mechanism Summary	Adaptive/Ordered Aspects
AVP-Fusion (Wen et al., 25 Dec 2025)	CNN–BiLSTM branches + per-instance gating	Per-sample, channel-wise dynamic gating; reused across tasks (transfer learning)
AWF (Sun et al., 2024)	Trainable alpha for model weight fusion	Alternating training to balance old/new knowledge
AG-Fusion (Liu et al., 27 Oct 2025)	Cross-modal attention, window-based gating	Bidirectional cross-attention, spatially varying pixel-wise gates
AFF (Mungoli, 2023)	Softmaxed attention weights, optional meta-gating	Modular insertion; supports R-AFF for ordered fusion
S-AdaFusion (Qiao et al., 2022)	Spatial summary + trainable selector conv	Fixed per-source order; pooling and padding for variable size
LAFB (Wang et al., 2024)	Bank of fusion schemes bet-weighted by AEM module	Per-layer, per-challenge adaptive weighting; weights reveal ordered importance

Mathematically, per-sample adaptive fusion modules generally instantiate as:

$$\begin{align*} \text{(a) Score computation:} &\quad e_i = a(W_\text{att} f_\text{proj}(x_i) + b_\text{att}) \ \text{(b) Normalization:} &\quad \alpha_i = \frac{\exp(e_i)}{\sum_j \exp(e_j)} \ \text{(c) Fusion:} &\quad f = \sum_{i=1}^n \alpha_i \cdot f_\text{fuse}(x_i) \end{align*}$$

where $n$0 and $n$1 are typically $n$2 convs or FCs, and $n$3 is a nonlinearity (Mungoli, 2023).

Weight fusion in continual learning proceeds via

$n$4

with $n$5 alternately trained against task loss (Sun et al., 2024).

4. Practical Applications Across Domains

Adaptive/ordered fusion has become a backbone for numerous advanced systems:

• Bioinformatics: Panoramic fusion spaces in peptide classification (AVP-Fusion) aggregate LLM embeddings with multi-descriptor features, dynamically balancing motif-level and context-level predictors (Wen et al., 25 Dec 2025).
• 3D Perception: Point-wise and window-wise fusion (MVAF-Net (Wang et al., 2020), AG-Fusion (Liu et al., 27 Oct 2025)) enables robustness to spatially localized sensor degradation.
• Multi-modal Saliency/Object Detection: LAFB enables robust salient object detection over RGB-Thermal/Depth images, adapting scheme weighting to scene-specific challenge regimes (Wang et al., 2024).
• Sequential Recommendation and Time-Series: TASIF's frequency domain adaptive filtering and strictly ordered multi-sequence gating ensure both denoising and preservation of event order, with significant efficiency gains over fully pairwise cross-attention (Luo et al., 30 Dec 2025).
• Cooperative Perception in Connected Vehicles: S-AdaFusion and C-AdaFusion provide order-sensitive, learned selection over architectural axes, achieving improved perception compared to fixed aggregation (Qiao et al., 2022).
• Model Fusion for Continual/Lifelong Learning: AWF provides mechanisms for training-task aware blending of model weights, reducing catastrophic forgetting and providing fine-grained adaptation (Sun et al., 2024).

5. Robustness, Interpretability, and Empirical Impact

Adaptive/ordered fusion architectures yield empirically validated benefits:

• Robustness to Input or Model Degradation: Adaptive gating mitigates the impact of unreliable sensors (AG-Fusion (Liu et al., 27 Oct 2025)) or feature branches (MVAF-Net (Wang et al., 2020)), significantly improving AP on challenging or occluded scenarios.
• Fine Control of Tradeoffs: In incremental segmentation, AWF enables direct, dynamic negotiation between old and new class performance per training regime (Sun et al., 2024).
• Systematic Handling of Specialized Challenges: LAFB adapts to center bias, clutter, or modality ambiguity by learning challenge-wise branch importances; CLDyN implements closed-loop feedback to align fusion with actual downstream task gradients (Yang et al., 10 Apr 2026).
• Interpretability: Gating vectors (e.g., in SSAFB (Dhiman et al., 8 Apr 2026)) and per-source fusion weights are amenable to local and global feature attribution analyses (LIME, PFI), enabling inspection of fusion decisions.
• Ablation and Sensitivity Studies: Removal or randomization of adaptive fusion consistently degrades performance by 2–4 mAP points, 1–3 F-score, or more, confirming the functional necessity of conditional fusion (see (Mungoli, 2023, Wang et al., 2024, Sun et al., 2024)).

6. Open Problems and Research Trajectories

Current and future avenues in adaptive/ordered fusion include:

• Increasing Fusion Granularity and Hierarchy: Layer-wise, spatially-varying, or block-wise adaptive factors are underexplored in continual learning (Sun et al., 2024).
• Hybrid Fusion Modes: Combining adaptive spectral and spatial gating, or integrating frequency-domain and task feedback, as in AdaSFFuse (Wang et al., 21 Aug 2025) or CLDyN (Yang et al., 10 Apr 2026), supports generalization to new domains.
• Efficient Scaling with Modalities: Efficient "guide-not-mix" schemes addressing quadratic cost in number of modalities or attributes—exemplified by the linear-complexity ASIF (Luo et al., 30 Dec 2025)—enable practical deployment in high-dimensional multi-source contexts.
• Unsupervised or Self-Supervised Adaptive Fusion: bridging the training-signal gap for unlabelled or transfer learning situations, as in DRF (Gu et al., 2021).
• Order-Invariant Adaptive Fusion and Set-Based Fusion: Extending from order-sensitive to permutation-invariant fusions for fully flexible architectures, especially in dynamically sized multi-agent or multi-modal settings.

7. Summary Table: Prominent Adaptive/Ordered Fusion Models

Model/Framework	Adaptive Mechanism	Ordered Logic	Key Domains
AVP-Fusion (Wen et al., 25 Dec 2025)	Per-input gated CNN+BiLSTM	Two-stage transfer	Biosequence analysis
AFF (Mungoli, 2023)	Attention, meta-gating	R-AFF, Cumulative	Vision, language, graph
AWF (Sun et al., 2024)	Trainable $n$6 weight	Stepwise, alternated	Continual learning
S-AdaFusion (Qiao et al., 2022)	Spatial/task-wise selectors	Channel-wise ordered	Cooperative vehicle
AG-Fusion (Liu et al., 27 Oct 2025)	Bidirectional cross-attn	Window-pixel adaptive	3D detection
LAFB (Wang et al., 2024)	Multi-branch + AEM module	Layered, challenge	Multimodal SOD
MVAF-Net (Wang et al., 2020)	Pointwise attention	Modality order fixed	LiDAR-Camera fusion
TASIF (Luo et al., 30 Dec 2025)	Frequency-domain filtering	Sequence preserved	Sequential rec
AdaSFFuse (Wang et al., 21 Aug 2025)	AdaWAT + SSM blocks	Band-ordered, pipeline	Image fusion
CLDyN (Yang et al., 10 Apr 2026)	Task-driven semantic inject	Closed-loop, feedback	Multi-task fusion

Adaptive and ordered fusion represent a broad methodological class fundamental to modern deep learning, enabling context-sensitive, efficient, and robust integration of heterogeneous or sequential inputs. Their continued evolution is expected to underpin future advances in multi-modal learning, continual adaptation, robust perception, and beyond.