Single-Source Model: Core Concepts

• Single-Source Model is a modeling approach where one privileged origin—be it a task, modality, domain, or physical emitter—defines the system's dependency structure and robustness criteria.
• It is applied in varied fields such as fMRI task transfer, multimodal robustness, domain generalization, graph algorithms, and astrophysics, demonstrating specialized use cases in each area.
• Empirical studies and algorithmic formulations reveal its effectiveness in isolating directed transfers and failure modes, while highlighting limitations in addressing many-to-one interactions and supervision constraints.

Searching arXiv for papers related to “single-source model” across the domains represented in the provided material. Single-Source Model is a field-dependent term for formulations organized around one privileged source: a source task, one modality, one labeled domain, one source node, one emitter, one astrophysical accelerator, or one initial site. Across recent arXiv literature, the common feature is not a single shared formalism but a single distinguished origin that defines the admissible dependence structure, robustness criterion, or explanatory mechanism (Xia et al., 24 Jun 2026, Yang et al., 2022, Cho et al., 19 May 2025, Luo et al., 2023, Erlykin et al., 2015, Engelen et al., 2013).

1. Semantic scope across research areas

The term is used in several technically distinct ways.

Domain	Meaning of “source”	Representative papers
fMRI taskonomy	One source task state reused for one target task	(Xia et al., 24 Jun 2026)
Multimodal robustness	One corrupted modality while other modalities remain clean	(Yang et al., 2022, Kim et al., 2019)
Domain generalization	One labeled source domain with no target data during training	(Cho et al., 19 May 2025, Arsenos et al., 2024)
Audio curation	One salient sound event without other sound-event types	(Yang et al., 13 May 2026)
Graph and network algorithms	One query/source node or one infection source	(Luo et al., 2023, Assadi et al., 23 Jul 2025, Hu et al., 2014)
Physical origin models	One nearby accelerator, one emitter, or one master flavor source	(Erlykin et al., 2015, Barr et al., 2015, Engelen et al., 2013)

In some literatures, the source is a computational primitive. In others, it is a threat model, a data-purity assumption, or a physical hypothesis. The phrase therefore denotes a modeling stance rather than a universal architecture.

2. Directed pairwise transfer in cognitive taskonomy

In reconstruction-based fMRI taskonomy, the single-source model is defined explicitly as a directed transfer relation $s \rightarrow t$: an encoder $E_s$ learned by masked fMRI reconstruction on source task $s$ is frozen and reused for low-data reconstruction on target task $t$, while only a target decoder is retrained on $1\%$ of target-task training data (Xia et al., 24 Jun 2026). The common self-supervised objective is masked fMRI reconstruction on samples $x \in \mathbb{R}^{R \times T}$ with $R=360$ cortical regions and $T=20$ time frames, evaluated through a normalized transfer distance

$d_t(\{s\})=\frac{L_t(\{s\})}{L_t^{\mathrm{gold}}},$

and visualized through within-target standardized affinities. Because all source-target relations are scored by the same held-out reconstruction criterion, the resulting taskonomy is a directed graph rather than an undirected similarity matrix.

The experimental setting uses 23 Human Connectome Project task states from 7 paradigms, with train/test splits in a 4:1 ratio and no subject overlap. Exhaustive evaluation of distinct ordered task pairs yields $23 \times 22 = 506$ single-source transfer models, forming part of a total of 1,127 trained models once gold and multi-source variants are included (Xia et al., 24 Jun 2026).

The main single-source findings are that transfer is strongly directional and paradigm structured. The five motor states form a coherent within-category block, with especially strong transfer among homologous left-right effectors, but comparatively weak transfer to many non-motor targets. The paper interprets this as a shared sensorimotor execution system overlaid by effector-specific representations. Working-memory states require a more careful directional reading: as targets, they can receive support from a broader range of non-motor sources, but as sources they are not uniformly dominant, and several transferred less broadly than sources such as math or neutral in the prior single-source analysis (Xia et al., 24 Jun 2026).

The paper also states the central limitation of the single-source model. It captures directed pairwise dependencies under a controlled low-data adaptation protocol, but it is inherently pairwise. Many-to-one relations are not fully captured by pairwise taskonomy alone, and budget-constrained supervision allocation can prioritize tasks that are not the strongest average outgoing single sources. This is the paper’s main reason for extending from $E_s$0 to source-set-to-target relations $E_s$1 (Xia et al., 24 Jun 2026).

3. Single-source as fault isolation, adversarial regime, and source purity

In multimodal fusion, single-source usually denotes a failure model rather than a transfer primitive. In the single-source adversarial setting, a multimodal model still receives all $E_s$2 modalities at inference time, but the attacker perturbs only one modality $E_s$3, leaving the remaining $E_s$4 modalities clean (Yang et al., 2022). The 2022 paper shows that standard mid- to late-fusion models are vulnerable in precisely this regime: on EPIC-Kitchens, concatenation and mean fusion reduce top-1 action accuracy to $E_s$5 under visual, motion, and audio perturbations; on KITTI, perturbing one depth modality can collapse AP; and on MOSI, text-only attacks sharply degrade both binary and 7-class sentiment accuracy (Yang et al., 2022). The proposed defense augments the fusion model with an odd-one-out detector and a robust fusion layer containing $E_s$6 experts, one per omitted modality plus one all-modality expert, thereby using cross-modal consistency to suppress the suspect source.

An earlier formulation studies the same general problem under the name single-source robustness and reaches a similar conclusion: robustness is not guaranteed even in a linear fusion model, because shared information may be allocated asymmetrically across sources (Kim et al., 2019). That paper defines a worst-case single-source-noise objective, MaxSSN, and proposes two training procedures, TrainSSN and TrainSSNAlt, together with a latent ensemble layer (LEL). On KITTI 3D detection with RGB and LIDAR, both MaxSSN-style training and LEL improve worst-case single-source robustness while largely preserving clean-data performance (Kim et al., 2019).

A distinct use of the term appears in audio curation. In "FSD50K-Solo" (Yang et al., 13 May 2026), a single-source clip is defined operationally as an audio clip whose labeled event is present in isolation, without other types of sound events. Multi-source clips include strong background noise, overlapping sound events, and clips in which the target event is only sparsely present in time. The paper constructs synthetic supervision from diffusion-generated clean event clips, trains a BEATs-plus-BiLSTM-plus-MLP binary classifier with binary cross-entropy, and filters FSD50K into FSD50K-Solo with 32,880 retained samples. The reported performance is 93.47% accuracy on the generated test set and 95.51% accuracy on an expert-curated Bose Sound Events test set (Yang et al., 13 May 2026). Here, single-source is neither a robustness regime nor a transfer relation; it is a property of source purity in the data itself.

4. Single-source domain generalization

In single-source domain generalization, the source is a training domain. The defining assumption is that training uses data from only one labeled source domain, while evaluation is on multiple unseen target domains with no target data available during training (Cho et al., 19 May 2025, Arsenos et al., 2024). This is stricter than multi-source DG because domain invariance cannot be inferred directly from multiple environments.

"PEER pressure" (Cho et al., 19 May 2025) studies augmentation-based single-source DG and argues that target-domain performance universally fluctuates during training because the model cannot accumulate knowledge learned from diverse augmentations. PEER addresses this by separating a frozen task model from a trainable proxy model. The proxy learns on original and augmented source samples, while the task model is updated only through uniform parameter-space averaging over saved proxy snapshots. Representation alignment is imposed through a mutual-information-style regularizer, implemented by default with Barlow Twins. The method is evaluated on PACS, Digits, Office-Home, and VLCS, with $E_s$7, $E_s$8, and $E_s$9. The paper reports gains over the same-backbone state of the art of 2.30% on PACS and 0.96% on Digits, and large gains over plain RandAugment on Office-Home and VLCS (Cho et al., 19 May 2025).

"CUDGNet" (Arsenos et al., 2024) uses a different single-source DG strategy. It assumes one labeled source domain $s$0, generates fictitious domains $s$1 through a transformation component, EFDMix style transfer, and Gaussian latent perturbations $s$2, and learns source–generated invariance with InfoNCE. The generator is trained adversarially through

$s$3

while the overall task model minimizes a variational objective plus contrastive loss. On CIFAR-10-C, the reported average accuracy is 85.53%, exceeding the prior best 78.45%; on PACS, training on Photo and testing on Art Painting, Cartoon, and Sketch yields 57.32% average accuracy, slightly above the strongest listed baseline at 57.17% (Arsenos et al., 2024). In this literature, the single-source model is the data regime itself.

5. Source-rooted algorithmic and network formulations

Several algorithmic papers use single-source in the literal graph-theoretic sense of a distinguished query or origin node. In single-source SimRank, the task is to compute or approximate $s$4 for one fixed source node $s$5. The MPC algorithm in (Luo et al., 2023) is the first with provable additive error $s$6 that achieves $s$7 communication rounds while using strongly sublinear memory per machine $s$8, overcoming the apparent $s$9 round barrier.

The same source-rooted pattern governs semi-streaming shortest paths. In (Assadi et al., 23 Jul 2025), the source is a designated root $t$0, and the output is a $t$1-approximate shortest path tree. The paper gives a randomized algorithm using

$t$2

and proves that any semi-streaming algorithm achieving even a constant-factor approximation requires $t$3 passes (Assadi et al., 23 Jul 2025).

In single-source unsplittable flow, the source is a common origin $t$4 from which all terminal demands are routed. The planar acyclic result in (Traub et al., 2023) shows that a fractional single-source flow can be rounded to an unsplittable one with

$t$5

for all arcs without costs, and with exact cost preservation at the price of a $t$6 additive deviation in the costed setting. The proof uses a noncrossing path decomposition and a structured discrepancy problem on a non-interleaving partition (Traub et al., 2023).

Two further papers place single-source at the hypothesis level. In single-snapshot source localization, a single source is exactly the sparsity-order-one case, $t$7, in a complex underdetermined regression model. The c-LARS-GIC procedure computes the exact knot sequence of the complex Lasso path, evaluates generalized information criteria at those knots, and selects $t$8 when the data support a single-source explanation (Tabassum et al., 2018). In epidemic inference under the SIRI model, the single-source assumption is encoded by binary initial conditions $t$9 with exactly one active source over the candidate set. The HISS estimator then performs approximate MAP inference over source node and unknown elapsed time using state propagation with infection, recovery, and reinfection parameters (Hu et al., 2014).

6. Physical, astrophysical, and high-energy uses

In astrophysics, the Single Source model denotes a source-plus-background hypothesis. The local cosmic-ray flux is modeled as the sum of a smooth Galactic background and one nearby, relatively recent source, typically a supernova remnant (Erlykin et al., 2015, Erlykin et al., 2016). In the antiproton paper, the motivating datum is the approximately constant AMS-02 $1\%$0 ratio from about $1\%$1 to $1\%$2, which the authors regard as difficult to explain by standard secondary production alone. They compute a secondary background from proton and helium interactions, subtract it from the measured ratio, and interpret the residual as a Single Source antiproton component. Among Vela, Monogem, and Geminga, Monogem is described as the closest match in the sub-TeV regime, though the conclusion is framed as compatibility rather than proof (Erlykin et al., 2015). In the secondary-nuclei paper, the same framework is extended to boron and lithium, with a background escape grammage

$1\%$3

and enhancement factors at $1\%$4 of $1\%$5 for $1\%$6, $1\%$7 for $1\%$8, and $1\%$9 or $x \in \mathbb{R}^{R \times T}$0 for Li depending on the background construction (Erlykin et al., 2016).

In flavor theory, "single source" means a unique origin of flavor violation. The SU(5)-based model in (Barr et al., 2015) assumes that all observable flavor violation comes from mixing between the three ordinary chiral families and extra vectorlike $x \in \mathbb{R}^{R \times T}$1 fermions. After integrating out the heavy states, all low-energy flavor-changing effects are governed by one master matrix

$x \in \mathbb{R}^{R \times T}$2

which enters $x \in \mathbb{R}^{R \times T}$3, $x \in \mathbb{R}^{R \times T}$4, and $x \in \mathbb{R}^{R \times T}$5 (Barr et al., 2015). The same matrix also controls scalar-mediated flavor-changing processes.

In charged-particle source physics, the phrase refers to an actual single emitter. The isolated single-atom electron source model in (Engelen et al., 2013) treats one photoelectron emitted from one atom and then propagated classically in a Coulomb-plus-uniform-field potential,

$x \in \mathbb{R}^{R \times T}$6

From closed expressions for asymptotic transverse velocities and trajectories, the paper derives effective source temperature and effective source size as functions of acceleration field and ionization laser energy, and shows that the hydrogenic model also describes near-threshold photoionization of rubidium with good accuracy (Engelen et al., 2013).

A different physical use appears in the single-source stochastic sandpile model. There, the source is the origin of the initial mass: $x \in \mathbb{R}^{R \times T}$7 grains are placed at $x \in \mathbb{R}^{R \times T}$8 on $x \in \mathbb{R}^{R \times T}$9, a site topples when it has at least $R=360$0 grains, and each toppling sends a random $R=360$1 drawn from $R=360$2 to each of its four neighbors (Selig et al., 2022). The paper studies the radius

$R=360$3

and avalanche number

$R=360$4

and reports the asymptotic forms

$R=360$5

together with a phase transition in the conditioned-binomial case at the scale $R=360$6 (Selig et al., 2022).

7. Recurrent distinctions and conceptual limits

Several distinctions recur across these literatures. First, single-source does not usually mean single-explanation sufficiency. In fMRI transfer, the pairwise model $R=360$7 is informative but insufficient for many-to-one relations and budget-constrained allocation (Xia et al., 24 Jun 2026). In multimodal learning, redundant inputs do not automatically yield robustness to one corrupted source; ordinary fusion can still collapse under single-source perturbation (Yang et al., 2022, Kim et al., 2019). In single-source DG, one labeled domain does not solve model selection: target-domain performance can fluctuate during training, which is precisely the phenomenon PEER was designed to address (Cho et al., 19 May 2025).

Second, the source is not always an information source in the same sense. It may be a task representation, a modality under attack, a training environment, a graph root, a candidate infection origin, a cosmic-ray accelerator, a master flavor structure, or an isolated emitter. This suggests that the term is best interpreted as a family resemblance concept: a single privileged origin constrains the model, but the privileged object and the operative mathematics change with the field.

Third, single-source formulations often expose asymmetry more clearly than multi-source or averaged models. The fMRI taskonomy paper stresses that $R=360$8 and $R=360$9 are different experiments (Xia et al., 24 Jun 2026). The multimodal-robustness papers distinguish robustness to corruption of source $T=20$0 from robustness to corruption of source $T=20$1 (Yang et al., 2022, Kim et al., 2019). The algorithmic papers distinguish single-source problems from all-pairs variants not just by size but by structure: the fixed root $T=20$2 changes both the computational objective and the proof strategy (Luo et al., 2023, Assadi et al., 23 Jul 2025).

A plausible implication is that single-source models are most valuable when one needs to isolate directed dependence, one-source failure modes, or one-origin explanatory hypotheses. They are less reliable when the real phenomenon is intrinsically many-to-one, jointly multimodal, or globally allocative.