MemANS: Event-Driven In-Memory Agentic Name Search for Efficient Multimodal Misinformation Detection on a Balanced Fakeddit Benchmark

1. Introduction

1.1. Background and Motivation

The proliferation of misleading and deceptive content on social-media platforms has transformed misinformation detection into a central research problem for multimedia computing. Contemporary deceptive posts rarely rely on text alone; they combine titles, imagery, engagement statistics, and community context into composite media objects that must be interpreted jointly. Consequently, the relevant scientific question is no longer whether a model can classify text accurately, but whether a multimedia system can reason over heterogeneous signals in a manner that is both effective and operationally scalable [1,2].

A substantial body of work has established that multimodal fake-news detection can outperform unimodal alternatives. Early systems such as SpotFake demonstrated that combining textual and visual evidence improves classification relative to text-only baselines [3]. Subsequent studies refined this intuition by examining richer fusion mechanisms, multi-image settings, and attention-based pooling architectures [4,5]. The introduction of the Fakeddit benchmark marked a particularly important milestone because it supplied a public, fine-grained multimodal dataset for six-way misinformation classification rather than only binary true–false prediction [1]. Building on that benchmark, Segura-Bedmar and Alonso-Bartolome confirmed that multimodal approaches often yield clear performance gains, especially for classes in which visual inconsistency or image-dependent manipulation plays a substantial role [6].

More recent literature has pushed the field beyond static feature concatenation. Dynamic fusion models now attempt to adapt cross-modal weighting to instance-specific conditions, thereby addressing the fact that different samples demand different evidence-combination strategies [7]. Expert-oriented systems, including modality-interactive mixture-of-experts architectures, suggest that specialist decomposition can be more effective than uniform processing under certain multimodal conditions [8]. Self-learning multimodal detectors have further indicated that model adaptability and iterative improvement remain active research directions [9]. Taken together, these studies suggest that the next stage of multimodal misinformation detection will be defined not merely by adding more modalities, but by designing architectures that determine when and how each modality should matter.

1.2. Research Gap and Problem Statement

Despite this progress, an important conceptual limitation remains. Even in recent high-performing systems, adaptation is typically treated as an internal model behaviour: fusion weights, attention coefficients, latent expert mixtures, or modality interactions are learned inside a unified predictive mechanism. What is much less common is an explicit service-resolution abstraction in which the system begins from a default resolver, observes event conditions, and then selectively escalates to named specialists. In other words, the literature has improved multimodal fusion, but it has not sufficiently explored multimodal reasoning as a name-resolution process over callable experts.

This distinction matters for at least three reasons. First, unconditional fusion may waste computation on samples that do not require multimodal escalation. Second, when adaptation is internal and opaque, it is difficult to explain why a particular modality was emphasised for a particular case. Third, if multimodal reasoning is framed as a monolithic fusion block, extending the system to accommodate new specialists requires re-engineering the entire architecture rather than registering an additional named expert. These limitations are summarised in Table 1, which contrasts the dominant problem dimensions in prior literature with the response offered by the present work.

Table 1 clarifies that the gap is not a lack of multimodal methods, but a lack of systems-level abstractions that make adaptive reasoning explicit, interpretable, and efficient. The present article develops that missing perspective through the notion of Agentic Name Search (ANS). In this work, ANS is conceived as a domain name system (DNS)-like abstraction for artificial intelligence: a computational layer that resolves a query to an appropriate named expert rather than assuming that every specialist must be invoked in every case. This interpretation is informed by the recent Agent Name Service proposal of Huang, Narajala, Habler, and Sheriff, which formalises DNS-inspired, capability-aware agent discovery for interoperable AI systems [10]. From this perspective, multimodal misinformation detection becomes a problem of dynamic capability resolution: a default text–metadata expert handles ordinary posts, while posts exhibiting cross-modal uncertainty are escalated to a specialist designed for ambiguity-rich cases.

The resulting system, termed MemANS, is event-driven because routing is triggered by observable disagreement and confidence conditions, and in-memory because the routing decision relies only on internal posterior state, confidence, and metadata already generated in the current inference pass. The in-memory property is conceptually important. Much contemporary discourse around adaptive AI systems focuses on retrieval-augmented generation (RAG) and external memory stores. Those approaches are valuable in knowledge-intensive settings, but they are not the only way to support adaptive behaviour. In multimedia verification, it may be more effective to perform routing over internal capability cues rather than retrieving external documents. Recent multimodal agent research on ToolMem reinforces this design intuition by showing that learnable tool-capability memory can improve capability selection without collapsing adaptation into generic external retrieval [11]. MemANS therefore represents a beyond-RAG design in a narrow but meaningful sense: adaptation is driven not by document retrieval or nearest-neighbour lookup, but by compact internal memory of what the active experts believe and how strongly they believe it.

The study is guided by three research questions. RQ1 asks whether an event-driven ANS architecture can outperform the strongest single-stage baseline on a balanced multimodal misinformation benchmark. RQ2 asks whether MemANS can surpass unconditional multimodal fusion while using fewer average expert invocations. RQ3 asks whether the final routing policy can be justified empirically through structured search rather than through manual preference. These questions lead to the following expectations: that ANS-style escalation will improve macro-F1 over a strong default resolver, that it will more clearly outperform always-on fusion, and that a structured policy search will identify the strongest routing configuration.

1.3. Contribution Statement

The article makes four concrete contributions. First, it introduces MemANS, an event-driven, in-memory ANS architecture for multimodal misinformation detection. Second, it constructs a balanced, fully image-available Fakeddit benchmark suitable for transparent evaluation in resource-constrained environments. Third, it presents a reproducible empirical comparison among default, fusion, and ANS-routed systems with bootstrap confidence intervals and paired t-tests. Fourth, it advances a multimedia systems interpretation in which adaptive reasoning is treated as explicit expert resolution rather than merely internal multimodal fusion.

1.4. Paper Organisation

The remainder of the article proceeds as follows. Section 2 synthesises the literature on multimodal fake-news detection, dynamic fusion, and agentic systems, and formalises the research gap. Section 3 details the benchmark construction, feature engineering, routing policy, and evaluation protocol. Section 4 reports the empirical findings, including the main system comparison and statistical testing. Section 5 interprets the results in relation to prior work and discusses implications for multimedia system design. Section 6 concludes the article and outlines promising future directions.

2. Related Work

2.1. Evolution of Multimodal Misinformation Detection

The detection of deceptive online content has shifted from unimodal textual analysis toward architectures that jointly reason over language, imagery, and social metadata. This trajectory reflects a growing recognition that misinformation is a multimedia phenomenon in which text and visual signals interact in ways neither channel alone can capture.

2.1.1. Early Multimodal Baselines and the Emergence of Joint Text–Image Reasoning

The foundational premise that multimodal fusion outperforms text-only approaches was established by early systems demonstrating joint text–image classification. Singhal and colleagues proposed SpotFake, a multimodal framework combining textual embeddings with visual features, showing that shallow cross-modal concatenation improves over isolated baselines [3]. Wang and colleagues introduced EANN, which employed adversarial training to learn event-invariant representations, addressing the tendency of models to overfit to specific news topics rather than learning generalisable deceptive cues [12]. Khattar and colleagues proposed MVAE, a multimodal variational autoencoder modelling text and image through a shared latent representation, offering a generative perspective on fusion that early discriminative baselines did not address [13]. Singh, Ghosh, and Sonagara further demonstrated that multimodal analysis yields more robust detection than either modality in isolation, a finding replicated across subsequent studies [2]. These works collectively established that fake-news detection is a multimedia inference problem in which cross-modal interaction carries discriminative signal.

2.1.2. Fine-Grained Benchmarks and the Shift from Binary to Multi-Class Detection

The methodological significance of early studies was amplified by the Fakeddit benchmark, a public multimodal dataset that departed from the binary true–false paradigm by supporting six-way fine-grained misinformation classification [1]. Unlike earlier datasets that conflated distinct deception strategies, Fakeddit distinguishes among true, satire, false connection, imposter content, manipulated content, and misleading content, enabling researchers to examine not only whether multimodal signals help but which classes benefit and under what conditions. Segura-Bedmar and Alonso-Bartolome leveraged this benchmark to demonstrate that multimodal systems substantially improve over text-only approaches and that visual information is particularly valuable for classes such as manipulated content and false connection, where visual–semantic inconsistency is a primary deceptive mechanism [6]. Wu and colleagues further addressed class imbalance by proposing a balanced multimodal learning framework with hierarchical fusion, showing that rebalanced training improves both fairness and classification quality [14]. These advances shifted the field from asking whether multimodal models work to asking how their benefits are distributed across heterogeneous categories.

2.1.3. Advanced Fusion Architectures: Progressive Fusion, Entity-Oriented Methods, and Structured Pooling

Attention then turned toward improving the mechanism of fusion itself. Kumari and Ekbal proposed AMFB, which introduced structured attention over multimodal feature interactions, demonstrating that factorised bilinear pooling captures richer cross-modal correlations than concatenation [5]. Jing and colleagues advanced progressive fusion networks, in which cross-modal features are integrated through stage-wise refinement rather than a single operation, suggesting that fusion quality depends on depth of interaction [15]. Qi and colleagues proposed an entity-enhanced framework leveraging named entities as bridging structures between modalities, showing that entity-aware alignment improves detection when semantic correspondence is manipulated [16]. Zhou and colleagues pursued multi-grained fusion at multiple semantic and visual granularities, arguing that deception can manifest at the level of objects, scenes, or entire narratives [17]. Li and colleagues surveyed these strategies, documenting the transition from concatenation to attention-based, graph-based, and adversarial fusion methods, while noting that reproducibility remains unresolved [18]. Together, these studies indicate that the relationship between modalities is class-dependent, context-dependent, and amenable to progressively more sophisticated formulations.

2.2. Dynamic and Adaptive Fusion

The proliferation of advanced fusion architectures raised a subsequent question: if different samples exhibit different cross-modal relationships, should the fusion mechanism itself vary by instance? This question animated a research thread on dynamic and adaptive fusion, in which the central objective is not merely to design a stronger fusion block but to make the fusion process responsive to input-specific conditions.

2.2.1. from Static Concatenation to Dynamic Weighting

The dynamic fusion model MDF-FND, proposed by Lv and colleagues, is a particularly important contribution in this direction [7]. Rather than applying fixed modality weights, MDF-FND learns to adaptively mix semantic-level representations according to instance-specific evidence, treating fusion as a decision problem rather than a purely representational one. Giachanou, Zhang, and Rosso further showed that visual evidence can be distributed across multiple images in a single post, reinforcing the view that one-size-fits-all fusion is inherently suboptimal when image context varies by instance [4]. The implication is that multimodal misinformation detection should be conceptualised not as information accumulation but as adaptive evidence composition, in which each modality’s contribution is moderated by sample-specific ambiguity.

2.2.2. the Limitation of Model-Internal Adaptation

Despite the appeal of dynamic fusion, a critical limitation persists: adaptation usually remains model-internal. Whether through attention coefficients, gating mechanisms, or latent weighting, the decision to emphasise one modality is learned within a unified block and is difficult to inspect operationally. There is an important distinction between a system that weights modalities differently and a system that resolves a case to a different processing pathway on the basis of observable evidence. The literature has made considerable progress on the former, but comparatively little on the latter. Even the most sophisticated dynamic fusion models do not fully separate the question of when multimodal escalation is warranted from the question of how modalities are combined once escalation occurs. This matters for large-scale platforms, where efficiency and interpretability are first-class concerns alongside accuracy.

2.2.3. Mixture-Of-Experts Approaches

A related direction addresses monolithic fusion through mixture-of-experts (MoE) decomposition. Liu and colleagues proposed a modality-interactive MoE architecture, in which specialised subnetworks handle distinct modality interactions and a gating mechanism determines expert activation [8]. Shen and colleagues extended this through GAMED, a knowledge-adaptive framework selecting expert pathways according to semantic domain cues, showing that expert-centric modelling improves performance under heterogeneous distributions [19]. Kou and colleagues explored potential-features fusion networks combining explicit and latent feature families through an MoE-style routing layer, suggesting that the space of multimodal interactions is too large for any single fusion operator to cover optimally [20]. However, the MoE perspective does not automatically solve the resolution problem: the expert mechanism remains a latent, highly parameterised learning strategy rather than a transparent, named, and callable routing architecture. The gating function may adapt, but it does not typically resolve a case to an independently interpretable specialist in the way a service-oriented system would.

2.3. Agentic Systems and Capability Routing

Parallel to developments in multimodal fusion, the broader field of agentic AI has introduced frameworks treating adaptive behaviour as capability discovery and routing rather than internal model reweighting. These frameworks provide the intellectual foundation for reframing multimodal misinformation detection as a name-resolution problem.

2.3.1. Agent Name Service and Capability-Aware Agent Discovery

Huang and colleagues formalised the Agent Name Service (ANS) as a DNS-inspired abstraction for AI systems, in which agents register capabilities under symbolic names and queries are resolved to appropriate named providers rather than handled by monolithic models [10]. This treats interoperability and capability routing as first-class architectural concerns, shifting the design question from “how powerful is a single model?” to “how does the system know which model to invoke?” The ANS abstraction is conceptually close to the routing problem in multimodal misinformation detection, where the system must decide whether a default resolver suffices or whether a multimodal specialist should be consulted. Xie and colleagues surveyed large multimodal agents, documenting systems combining perception, reasoning, and tool use, and noting that effective capability routing remains a central unsolved challenge [21].

2.3.2. Learnable Capability Memory in Multimodal Agents

A related strand emphasises memory-assisted capability selection. Xiao and colleagues proposed ToolMem, a learnable tool-capability memory architecture that records historical success patterns of different tools and uses this memory to improve future capability selection without collapsing adaptation into generic external retrieval [11]. Hu and colleagues provided a broader conceptual treatment of memory in AI agents, distinguishing between external knowledge stores, latent sequence memory, and compact capability state, and arguing that different tasks demand different memory architectures [22]. For multimedia verification, where the primary challenge is deciding whether image evidence should alter a textual interpretation, these memory-aware routing frameworks offer a design vocabulary that the misinformation-detection literature has not yet adopted.

2.3.3. beyond RAG: In-Memory Capability Routing as an Alternative to External Retrieval

Much contemporary discourse around adaptive AI centres on retrieval-augmented generation (RAG), in which external documents are retrieved to enrich reasoning [23]. These approaches are valuable in knowledge-intensive settings, but they are not the only viable mechanism for adaptive behaviour. In multimedia verification, the critical issue is frequently not missing knowledge but multimodal ambiguity: the system already possesses the relevant image and text, yet must decide whether the image justifies a departure from the default textual interpretation. In such settings, adaptation may be more effectively performed through in-memory capability routing over internal state than through external document retrieval. MemANS is positioned within this beyond-RAG design space: its routing decisions are driven entirely by in-process expert posteriors, confidence signals, and disagreement events, rather than by nearest-neighbour retrieval or textual augmentation. This design choice is a principled assertion that some adaptive tasks require capability selection more than knowledge retrieval.

2.4. Positioning Against Prior Work

The preceding synthesis reveals that the literature has improved multimodal fusion, introduced dynamic weighting, and begun to decompose models into experts, yet it has not adequately formalised multimodal reasoning as explicit event-driven resolution among named experts using in-memory capability state. Table 2 compares MemANS against nine of the closest prior systems across five dimensions central to the present contribution: routing mechanism, efficiency strategy, interpretability, reproducibility, and benchmark.

Table 2 reveals a systematic pattern. Prior systems fall into two categories: static fusion architectures that always invoke all modalities (SpotFake, EANN, AMFB, Progressive Fusion, Entity-Enhanced, Self-Learning) and dynamic internal architectures that adapt fusion weighting or expert activation through learned gates (MDF-FND, MoE-FND, GAMED). Neither category captures the architectural position of MemANS. Static systems lack adaptation entirely. Dynamic systems embed adaptation within a monolithic graph, making it difficult to interpret as a discrete routing decision and impossible to extend by adding independently named specialists. MemANS occupies a distinct design point: it is explicitly event-driven (routing triggered by observable disagreement and confidence), named-expert-based (default resolver and specialist are independently callable), and in-memory (routing depends only on internal state from the current inference pass). Furthermore, MemANS is the only system in this comparison designed for high reproducibility in lightweight environments, employing TF-IDF features, compact image descriptors, and logistic-regression experts suitable for Google Colab Free execution.

The research gap can be stated precisely. Existing multimodal misinformation detectors have established the importance of multimodal evidence, improved fusion quality, and begun to incorporate expert decomposition, yet they do not adequately address explicit event-driven resolution among named experts using low-cost in-memory capability state. As a result, the field lacks empirical evidence on whether an Agentic Name Search architecture can match or exceed stronger multimodal baselines while reducing indiscriminate specialist usage, and on whether such routing can be justified through structured search rather than manual preference [6,7,24,25]. This article addresses that gap by testing a concrete ANS formulation, named MemANS, on a balanced and reproducible Fakeddit benchmark.

3. Methodology

3.1. Problem Formulation

This paper addresses fine-grained multimodal misinformation detection as a six-way supervised classification task. Each instance is a social-media post represented as an ordered triplet

$$x_i=(t_i,v_i,m_i),$$

(1)

where $t_i$ denotes the cleaned textual title, $v_i$ the associated image, and $m_i$ the structured metadata vector. The target label $y_i$ belongs to $\mathcal{Y}=\{0,1,2,3,4,5\}$, corresponding to true, satire, false connection, imposter content, manipulated content, and misleading content [1]. The text $t_i$ is lowercased, punctuation-normalised, and stop-word filtered before feature extraction. The image $v_i$ is resampled to a fixed spatial grid, and $m_i$ is a real-valued vector of engagement cues. This triplet formulation ensures that all systems operate on the same input space [1,6]. The task is to learn a classifier $f:\mathcal{X}\to \mathcal{Y}$ whose performance is evaluated by macro-level metrics that treat all six classes equally.

3.2. Dataset and Benchmark Construction

3.2.1. Fakeddit Selection Rationale and Public Accessibility

The source dataset is Fakeddit, a public multimodal benchmark introduced by Nakamura, Levy, and Wang [1]. Fakeddit satisfies four criteria essential for this study: public accessibility, multimodal content, fine-grained six-way annotation, and established recognition in the literature. Unlike binary datasets, its six-way taxonomy makes routing behaviour and specialist performance more interpretable.

3.2.2. Availability-Aware Balanced Subset

The full Fakeddit corpus includes instances with missing or broken image links. For a study comparing text-only, image-only, text–metadata, fusion, and routed systems on identical instances, image availability is a methodological prerequisite. The benchmark was reconstructed into a balanced, fully image-available subset by drawing class-balanced candidate pools, enforcing image availability through programmatic download and validation, and pruning an oversampled test candidate pool of 1320 instances (220 per class) to a final image-available set. The resulting splits are summarised in Table 3.

The balanced design eliminates class-skew confounding for macro-level metrics. By enforcing image availability across all splits, every comparison in Section 4 is conducted over the same multimodal population.

3.2.3. Image-Availability Enforcement and Split Reproducibility

Image availability was enforced by programmatic download of URLs referenced in the Fakeddit metadata, followed by validation of file integrity (non-empty, decodable by Pillow). Instances with missing or corrupted images were excluded. Split identifiers and per-instance image-path mappings were serialised as JSON artifacts. Any researcher who begins from the public Fakeddit metadata and applies the same filtering and balancing logic with the same random seed will obtain the identical split.

3.3. Feature Engineering

The feature design prioritises clarity, auditability, and computational modesty. All channels are summarised in Table 4.

3.3.1. Text Representation

The text modality is represented by TF-IDF features extracted from the cleaned title field. The vectoriser uses unigram and bigram tokenisation, English stop-word removal, max_features$=12,000$, and min_df$=2$. The standard sublinear tf–idf formulation is used, yielding a sparse representation of dimensionality $d_t=12,000$.

3.3.2. Image Representation

Each image is resampled to $32\times 32$ RGB pixels and encoded through grayscale flattening (1024 dimensions), per-channel mean and standard deviation (6 dimensions), per-channel histograms with 8 bins (24 dimensions), a grayscale histogram with 8 bins (8 dimensions), and simple edge statistics from a Sobel-filtered image (2 dimensions). The total image dimension is $d_v=1,064$.

3.3.3. Metadata Representation and Rationale for Lightweight Features

The metadata channel consists of num_comments, score, upvote_ratio, token_count, and char_count, all standardised to zero mean and unit variance. The use of handcrafted, interpretable features is a deliberate methodological commitment. When evaluating a routing architecture, every component should be independently auditable; opaque pretrained embeddings confound the attribution of performance differences to the routing layer versus the representation layer. By holding the feature surface constant and transparent, any observed gain can be attributed to architectural design [1].

3.4. Base Experts and Baselines

All predictive components are implemented with multinomial logistic regression using the scikit-learn LogisticRegression estimator with the lbfgs solver, multinomial loss, and max_iter $=1000$. This isolates the contribution of adaptive routing from large-scale representation learning. The regularisation C is $2.0$ for the specialist and $1.0$ for all other experts, reflecting the policy search in Section 3.6. The five experts are summarised in Table 5.

The text–metadata expert is the default resolver; the fusion expert represents the conventional assumption that multimodal evidence should always be jointly exploited. The specialist expert is trained exclusively on training instances where the text and image experts disagree in their argmax predictions, enabling conditional escalation for ambiguity-rich cases.

3.5. Memans Architecture

3.5.1. Dns-Like Resolution Layer over Named Experts

MemANS treats inference as a name-resolution process over a registry of named experts, analogous to DNS. A default resolver handles ordinary queries; escalation to a named specialist occurs only when an observable event signals that the default path may be insufficient. The architecture is depicted conceptually in Figure 1. The resolution layer is in-memory because routing decisions depend entirely on quantities computed in the current inference pass: expert posteriors, disagreement status, and confidence values.

3.5.2. Disagreement Event

Let $p_i^{\left(T\right)}\in {\mathbb{R}}^6$ and $p_i^{\left(V\right)}\in {\mathbb{R}}^6$ denote the posterior probability vectors of the text and image experts for instance i. A disagreement event is defined as

$$d_i=\mathbb{I}\left[arg\underset{c\in \mathcal{Y}}{max}{p}_i^{\left(T\right)}\left(c\right)\ne arg\underset{c\in \mathcal{Y}}{max}{p}_i^{\left(V\right)}\left(c\right)\right],$$

(2)

where $\mathbb{I}[·]$ is the indicator function. The test disagreement rate is 0.7976, indicating that cross-modal ambiguity is prevalent rather than exceptional.

3.5.3. Confidence Gate

Not every disagreement event warrants escalation. A confidence gate filters events by the strength of multimodal evidence. The fusion expert produces posterior $p_i^{\left(F\right)}\in {\mathbb{R}}^6$, and its maximum value serves as the confidence signal:

$$c_i=\underset{c\in \mathcal{Y}}{max}{p}_i^{\left(F\right)}\left(c\right).$$

(3)

This gate distinguishes weak disagreement from strong disagreement based on the fusion model’s confidence despite cross-modal mismatch.

3.5.4. Routing Equation and Posterior Selection

The routing decision is

$$r_i=\mathbb{I}\left[d_i=1\wedge c_i\ge \tau \right],$$

(4)

where $\tau$ is the confidence threshold. The final MemANS posterior is selected by substitution:

$$p_i^{\left(M\right)}=\left\{\begin{array}{cc}{p}_i^{\left(S\right)}\hfill & \mathrm{if}{r}_i=1,\hfill \\ p_i^{\left(TM\right)}\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$

(5)

where $p_i^{\left(TM\right)}$ and $p_i^{\left(S\right)}$ denote the posteriors of the default text–metadata expert and the specialist expert, respectively. The predicted label is ${\widehat{y}}_i=arg{max}_c{p}_i^{\left(M\right)}\left(c\right)$. The selected threshold is $\tau =0.40$, yielding a validation route rate of 0.7226 and a test route rate of 0.7202. On average, MemANS uses 1.7202 expert invocations per instance, compared with 2.0000 for the always-on fusion baseline.

3.5.5. Inference Routing Procedure

Algorithm 1 formalises the MemANS inference procedure. It is deterministic: given the same trained experts and threshold, it produces identical routing decisions for every instance.

This formulation yields three architectural properties: transparent routing (escalation is explainable by disagreement and confidence), bounded expert usage (invoked only under joint event conditions), and name-resolution abstraction (resolves to a named expert rather than reweighting latent states).

3.6. Policy Search and Validation

The final routing policy was not selected heuristically. A structured search was performed over specialist masks (disagreement-based and text-length-based triggers), specialist feature spaces (text-only, metadata-only, text–metadata), confidence sources (fusion maximum, text–metadata maximum, image maximum), and regularisation and threshold values. Thresholds were evaluated over a grid that preserved meaningful variation in route rate. The strongest validated configuration is summarised in Table 6.

3.7. Evaluation Metrics and Statistical Testing

The study reports accuracy, weighted precision, weighted recall, weighted F1, macro-F1, macro-precision, macro-recall, MCC, and weighted OvR AUC. Macro-F1 is the primary metric because it treats all six classes equally. Precision P, recall R, and F1-score are defined as

$$P={\displaystyle \frac{TP}{TP+FP}},R={\displaystyle \frac{TP}{TP+FN}},$$

(6)

$$F_1=2·{\displaystyle \frac{P·R}{P+R}}.$$

(7)

MCC provides a balanced summary under multiclass conditions, and weighted OvR AUC captures ranking quality across classes. Bootstrap estimation is employed over macro-F1 differences: for each system pair, 10,000 resamples of the test set are drawn with replacement; the 2.5th and 97.5th percentiles form a 95% CI, and the proportion of non-positive differences yields a two-sided bootstrap p value [25,26]. Confidence intervals are preferred because they convey the magnitude and uncertainty of the difference [24]. As a complementary procedure, paired t-tests are conducted over per-instance correctness vectors, and standardised effect estimates (Cohen’s d) are also reported.

3.8. Reproducibility Protocol

All experiments use a fixed global random seed of 42 for NumPy, scikit-learn, and Python’s built-in random module. The implementation uses Python 3.11, NumPy 1.26, pandas 2.2, scikit-learn 1.5, SciPy 1.13, and Pillow 10.4. Preprocessing is fully deterministic. The complete pipeline is packaged as a Google Colab-ready notebook designed for the free-tier runtime; it reproduces benchmark construction, feature extraction, expert training, policy search, final evaluation, statistical testing, and figure generation. The end-to-end construction and evaluation workflow is depicted in Figure 2. All artifacts—trained model coefficients, split identifiers, evaluation tables, routing summaries, and significance-test outputs—are preserved in the project workspace.

4. Results and Analysis

4.1. Main System Comparison

Table 7 reports the performance of MemANS and the two strongest baselines on the final balanced Fakeddit test benchmark. All systems are evaluated on identical instances drawn from the fully image-available, six-class test set (840 observations, 140 per class). The average number of experts invoked per test instance is also reported to capture operational cost.

The most striking finding is that the text–metadata expert is the strongest default baseline, not the always-on fusion expert. Across all reported metrics, the text–metadata system outperforms the fusion system by margins ranging from 0.0331 (MCC) to 0.0371 (macro-F1). This indicates that unconditional multimodal combination does not guarantee improved performance, even on a benchmark in which every instance possesses a corresponding image. Forcing image evidence into all decisions appears to introduce noise that dilutes the predictive strength of the text–metadata resolver, consistent with prior evidence that the utility of visual signals in misinformation detection is class-dependent and context-sensitive [6,7].

MemANS achieves the strongest absolute performance on every reported metric. Its macro-F1 of 0.6182 exceeds the text–metadata expert by 0.0079 and the fusion expert by 0.0454. The accuracy advantage follows the same ordering: 0.0083 above the text–metadata baseline and 0.0440 above the fusion baseline. Critically, MemANS attains these improvements while invoking, on average, 1.7202 experts per instance—a reduction of 0.2798 relative to the fusion expert’s fixed cost of 2.0000. MemANS thus occupies an intermediate operational position between single-stage default inference and unconditional multimodal escalation, and that intermediate position coincides with the highest classification quality observed in the study. Figure 3 visualises both the predictive ranking and the efficiency–performance trade-off. This pattern supports the central architectural hypothesis: selective expert resolution can outperform both indiscriminate fusion and default-only inference.

4.2. Statistical Comparison

Table 8 presents the inferential comparison of MemANS against the two baselines, and Figure 4 visualises the corresponding effect estimates and bootstrap intervals. Bootstrap estimation over 10,000 resamples yields 95% confidence intervals and two-sided bootstrap p values for the mean macro-F1 difference; paired t-tests are conducted over per-instance correctness vectors.

The comparison against the fusion baseline is unequivocal. The mean macro-F1 improvement of 0.0454 is accompanied by a 95% confidence interval that lies entirely above zero ($[0.0198,0.0710]$), a bootstrap p value of 0.0000, and a paired t-test p value of 0.0010. These results provide strong statistical evidence that MemANS outperforms unconditional multimodal fusion on the final test benchmark. The mean accuracy difference of 0.0440 reinforces the same conclusion. The practical interpretation is that event-driven specialist resolution is preferable to always-on fusion when the underlying visual features are compact and the image channel is not uniformly informative.

The comparison against the text–metadata baseline is more nuanced. The mean macro-F1 difference of 0.0079 is favourable to MemANS, but the 95% bootstrap confidence interval $[-0.0012,0.0173]$ crosses zero, and both the bootstrap p value (0.0980) and the paired t-test p value (0.0896) exceed the conventional 0.05 threshold. Consequently, the gain over the strongest default baseline cannot be declared statistically significant at standard levels. This does not diminish the result; rather, it places it in appropriate epistemic context. The text–metadata expert is already a strong and efficient resolver, and improving upon it with lightweight visual signals is inherently difficult. The observed positive trend suggests that ANS-style routing carries incremental value even against a strong default, but the current implementation should be treated as an architectural proof of concept rather than a saturated performance endpoint.

4.3. Routing Behaviour

Table 9 characterises the event-driven routing behaviour of MemANS on both the validation and test splits, and Figure 5 compares the disagreement and route rates across splits. The disagreement rate measures the proportion of instances in which the text and image experts produce different predicted classes; the route rate measures the proportion that satisfy both the disagreement condition and the fusion-confidence threshold, thereby triggering specialist invocation.

Two observations merit attention. First, the cross-modal disagreement rate is high: approximately 79–80% of instances in both validation and test exhibit a mismatch between the text expert’s and image expert’s top predictions. This indicates that multimodal misinformation detection on the Fakeddit benchmark involves substantial cross-modal ambiguity rather than straightforward additive evidence. Second, the confidence threshold filters the disagreement events into a smaller but still large specialist-eligible subset: roughly 72% of instances are routed. The agreement between validation and test route rates (0.7226 vs. 0.7202) is close, suggesting that the threshold generalises stably across splits. The operational significance is that MemANS avoids specialist invocation in approximately 28% of test cases, retaining the computational and predictive benefits of the default resolver without incurring the cost or noise of multimodal escalation.

4.4. Policy Search Evidence

The final MemANS configuration was not chosen through manual preference or post-hoc rationalisation. A structured policy search was conducted over candidate specialist masks, specialist feature spaces, confidence sources, regularisation values, and thresholds. The selected policy—disagreement-triggered specialist, text–metadata feature space, fusion confidence as the gating signal, regularisation $C=2.0$, and threshold $\tau =0.40$—emerged as the strongest validated configuration among all tested alternatives. Candidate policies based on short-text or very-short-text regimes approached the performance of the final design but did not surpass it. This outcome is methodologically important because it prevents the final system from being interpreted as a hand-tuned heuristic: the chosen policy is the empirically best-validated policy from a defined search space. The search result also carries a substantive theoretical implication. The dominant routing signal is not a superficial metadata partition or a length heuristic, but a cross-modal event condition: disagreement between the text expert and the image expert. This supports the central conceptual position that adaptive multimedia reasoning is best framed as resolution under evidence conflict rather than as generic conditional modelling.

4.5. Class-Level Patterns

The present study does not claim that all six misinformation classes benefit equally from ANS-style routing. Prior Fakeddit literature has established that categories such as manipulated content and false connection are especially sensitive to visual inconsistency, whereas others are more text-dominant [6]. The class-level behaviour observed in the current benchmark is consistent with that broader pattern. The confusion-matrix artifacts indicate that classification errors remain concentrated in semantically adjacent categories—particularly misleading content, false connection, and imposter content—where textual surfaces overlap and the structure of deception is subtle rather than overt. The lightweight image descriptors employed in this study are sufficient to expose some cross-modal tensions but not rich enough to eliminate all ambiguity in these adjacent categories. This is one reason the study interprets MemANS as a proof of architectural viability rather than as a performance ceiling.

4.6. Summary of Empirical Findings

The empirical evidence supports four concise statements. First, the strongest default baseline on the final benchmark is the text–metadata expert, not the always-on fusion expert. Second, MemANS improves upon that strong default, but the margin is modest and the confidence interval for the difference crosses zero. Third, MemANS clearly and significantly outperforms unconditional multimodal fusion, with a statistically supported macro-F1 gain of 0.0454 and a paired t-test p value of 0.0010. Fourth, the final routing policy emerged from structured search rather than from intuition, and cross-modal disagreement was identified as the dominant routing signal. Collectively, these findings provide empirical support for the proposition that Agentic Name Search is a viable and practically meaningful multimedia-systems abstraction for multimodal misinformation detection.

5. Discussion

5.1. Reinterpreting Multimodal Reasoning as Expert Resolution

The dominant paradigm in multimodal misinformation detection has treated progress as a representation problem: how to align visual and textual signals and engineer more expressive fusion operators [18]. The present findings suggest the challenge is also architectural—not merely how modalities should be fused, but under what conditions multimodal escalation should occur at all. MemANS addresses this by reframing inference as resolution among named experts. In conventional fusion, the system combines all evidence continuously. In MemANS, the system begins with a default resolver and asks whether a named specialist should be consulted. This shift makes multimodal inference more interpretable and aligned with agentic architectures [10]. Table 10 contrasts the two architectural lenses across five dimensions.

This comparison reveals that MemANS is an alternative systems abstraction. The routing mechanism operates over internal capability state, making the architecture lightweight and low-latency. This architectural interpretation extends beyond fake-news detection to video event interpretation, cross-modal search, and multimodal content moderation, where an ANS-style design governs escalation without imposing the fixed cost of unconditional fusion on every instance.

5.2. Why MemANS Outperforms Unconditional Fusion

The most robust empirical conclusion is that MemANS substantially outperforms the always-on fusion baseline. On the final test benchmark, MemANS achieves a macro-F1 of 0.6182, whereas the fusion expert attains 0.5732—a difference of 0.0454 that is statistically supported at $p<0.001$. This suggests that visual information is most effective when invoked selectively. MemANS avoids degradation by using the image expert as an event detector that measures whether visual and textual reasoning disagree, while fusion confidence determines whether the disagreement is sufficiently structured to justify specialist escalation. Rather than treating the image channel as a universal contributor, MemANS treats it as a conditional diagnostic signal consulted only when the default resolver faces evidence conflict. This interpretation aligns with prior evidence that the benefit of multimodality is context-sensitive and class-dependent [6], and extends that reasoning by showing that context-sensitivity can be formalised not only through dynamic fusion weights [7] but through an explicit routing architecture.

5.3. Why the Gain Over the Default Is Modest

The more modest gain over the text–metadata baseline is equally important to interpret. MemANS improves upon the text–metadata expert by 0.0079 in macro-F1 (0.6182 vs. 0.6103), but the 95% bootstrap confidence interval $[-0.0012,0.0173]$ crosses zero, and neither test reaches conventional significance. The text–metadata baseline is already a strong default resolver, and improving upon it with lightweight visual features is inherently difficult. That MemANS still improves on this baseline, albeit cautiously, is meaningful: it demonstrates that the routing principle carries predictive value beyond the strongest single-modality default. The confidence interval crossing zero makes clear that the current implementation should be regarded as an architectural proof of concept, not a final form.

5.4. Implications for Event-Driven Multimedia Systems

The present findings contribute to the conversation about event-driven architecture in multimedia AI systems. In many real-world platforms, it is undesirable to invoke every processing component on every sample. MemANS offers a concrete design principle: use multimodal disagreement and confidence as operational events that trigger specialist resolution. On the final test set, the route rate of 0.7202 means that MemANS avoids specialist invocation in approximately 28% of cases. A platform processing large volumes of content may classify rapidly while reserving more expensive reasoning for ambiguous cases. The default-resolver-plus-specialist pattern is a systems-design paradigm that balances predictive quality with operational efficiency.

5.5. Implications for Beyond-RAG Design

Retrieval-augmented generation (RAG) has become a dominant pattern for adaptive behaviour in language-centric systems, yet not every adaptive system requires external document retrieval. In multimedia verification, it may be more effective to ask whether the system should consult an additional expert capability rather than retrieve additional text. MemANS is closer in spirit to capability-memory designs such as ToolMem than to classical document-centric retrieval pipelines [11]. This highlights an underexplored alternative in agentic AI: adaptation through capability routing rather than external knowledge retrieval [21]. The in-memory design of MemANS also speaks to a broader discussion about memory in adaptive systems. Recent work on agent memory has emphasised internal state maintenance for long-horizon reasoning and tool selection [22]. MemANS contributes by demonstrating that even compact in-memory state can support meaningful adaptive behaviour, an alternative that may be especially attractive where the primary challenge is multimodal ambiguity rather than missing knowledge.

5.6. Limitations

Several limitations bound the scope of the contribution and should be acknowledged explicitly. First, the current implementation relies on handcrafted image descriptors and logistic-regression models; this supports reproducibility but constrains the performance ceiling, and a pretrained vision transformer could expose richer cross-modal signals. Second, the empirical evaluation is conducted on a single public benchmark, Fakeddit, so the findings may not generalise to other datasets or languages, and cross-benchmark validation is needed. Third, the in-memory design uses a fixed routing rule rather than a learned long-horizon memory structure: the threshold $\tau =0.40$ was selected through structured policy search but remains static, whereas a learned capability memory could adapt to distribution shifts more effectively [11]. Fourth, the specialist operates within the text–metadata feature space rather than incorporating a more sophisticated learned multimodal representation, so it cannot leverage deep cross-modal interactions available to a neural fusion module. These limitations define the scope precisely: the study establishes the viability of ANS-style routing under transparent and reproducible conditions.

5.7. Future Directions

Four directions are especially promising. The most immediate is to strengthen the expert pool while preserving the routing abstraction; pretrained vision encoders, transformer-based text experts, and multimodal contrastive specialists could be inserted into the same ANS framework without altering the routing logic. A second direction is to introduce learned capability memory, where routing decisions evolve from historical expert success rather than from a fixed disagreement rule alone, an approach supported by recent work on tool-capability memory [11]. A third direction is cross-benchmark validation across other multimodal misinformation datasets and multimodal agent benchmarks, to test whether the routing principle generalises across data distributions and languages. A fourth direction is resource-aware evaluation, in which latency, memory footprint, and throughput are compared explicitly against fusion-heavy baselines, moving beyond average expert counts to wall-clock operational metrics.

6. Conclusions

This article introduced MemANS, an event-driven, in-memory Agentic Name Search architecture for multimodal misinformation detection. By recasting multimodal reasoning as resolution among named experts rather than as internal feature fusion, the study moved beyond the dominant paradigm of unconditional multimodal combination and demonstrated that selective specialist activation can produce stronger classification results on a balanced, fully image-available Fakeddit benchmark [1]. On the final test set of 840 instances, MemANS achieved 0.6182 macro-F1, 0.6167 accuracy, 0.5405 Matthews correlation coefficient, and 0.8726 weighted one-vs-rest AUC, establishing that the architecture is not merely conceptually distinct but empirically competitive on a fine-grained six-class misinformation task.

The study also contributed a reproducible benchmark and a transparent statistical framework. It constructed a balanced Fakeddit benchmark with 5400 training, 840 validation, and 840 test instances, ensuring that all comparisons were performed over identical fully multimodal instances, and adopted bootstrap confidence intervals and paired t-tests to guard against spurious conclusions. Against the always-on fusion baseline, MemANS delivered a mean macro-F1 improvement of 0.0454, a 95% confidence interval entirely above zero ($[0.0198,0.0710]$), and a paired t-test p value of 0.0010, providing strong evidence for the superiority of event-driven routing. The gain over the strongest default text–metadata baseline ($\Delta =0.0079$) was positive but not statistically significant, indicating a promising yet modest trend that should be interpreted with appropriate caution.

Finally, the article advanced a multimedia-systems interpretation in which adaptive reasoning is treated as explicit expert resolution rather than internal multimodal fusion [7,8]. By using cross-modal disagreement and fusion confidence as operational events, the architecture offers transparent routing logic, bounded expert usage (1.7202 experts on average versus 2.0000 for unconditional fusion), and natural extensibility through named-specialist registration, a perspective informed by the Agent Name Service proposal for capability-aware agent discovery [10]. The most important empirical conclusion is that event-driven routing outperforms unconditional fusion on the final benchmark, and that Agentic Name Search is a viable multimedia-systems abstraction for adaptive social-media analysis. Future work should strengthen the expert pool with transformer-based and pretrained vision experts, replace the fixed disagreement rule with learned capability memory [11], validate across additional benchmarks, and measure latency, memory, and throughput against fusion-heavy baselines.