Dual-Mode Evidence System

• Dual-mode evidence systems are frameworks that differentiate and fuse individual (local) and collective (aggregated) evidence to enhance decision-making under uncertainty.
• They integrate explicit expert data with algorithmic interpolation, balancing fine-grained and global insights to overcome limitations of traditional models.
• These systems are applied in AI, distributed coordination, and integrated sensing, offering modular, adaptive solutions for complex, uncertain environments.

A dual-mode evidence system is a class of frameworks and methodologies in which two complementary modalities or operational modes of evidence—often individual/local and collective/aggregated—are structurally distinguished, formally coordinated, and algorithmically fused to support more expressive, reliable, and nuanced reasoning or decision-making. This paradigm occurs across epistemic logic, uncertain reasoning, vision, scientific inference, distributed systems, and artificial intelligence. Dual-mode systems typically enable the interplay between fine-grained (agent-level, tabular/explicit, local, or semantically focused) evidence and coarse-grained (common knowledge, interpolated/derived, global, or aggregated) evidence, incorporating both deductive and inductive or adaptive operations, and often facilitate expert override, iterative refinement, or joint optimization.

1. Foundational Logical Principles and Formalisms

Dual-mode evidence systems have formal roots in multi-agent justification logics, particularly those developed to express both agent-centric and common evidence (Bucheli et al., 2010). In these logics, explicit evidence terms (e.g., $t$) are syntactically attached to assertions, yielding formulas such as ${t}_C A$ ("$t$ justifies $A$ as common knowledge"). The syntax distinguishes operator modalities: for each agent ($i$), for mutual knowledge ($E$), and for common knowledge ($C$). Group knowledge is constructed using combinators such as tupling, projection, application, sum, co-closure, and induction:

• Application: $$A\rightarrow B\in E_*(w,t),\;A\in E_*(w,s)\Rightarrow B\in E_*(w,t\cdot s)$$
• Sum: $E_*(w,t)\cup E_*(w,s)\subseteq E_*(w,t+s)$
• Co-Closure: $${t}_C A\rightarrow {t}_E A, \;{t}_C A\rightarrow {t}_E {t}_C A$$
• Induction: $$A\wedge {t}_C (A\rightarrow {s}_E A)\rightarrow {ind\{t,s\}_C A}$$

Kripke-style AF-model semantics assign to each evidence term an intensional set of justified formulas, enforcing closure properties matching syntactic operations. As a result, these systems are sound, complete, and possess the finite model property. Restricting to agent-only operations yields conservative extensions of prior bimodal logics, and forgetful projection maps the system to standard modal logics like S4 with explicit realisers for modalities.

2. Algorithmic Duality: Evidence Acquisition and Combination

The dual-mode approach frequently emerges in the context of computational evidence fusion under uncertainty. Modifiable combining functions (Cohen et al., 2013) embody the duality by bridging expert-specified tabular evidence and algorithmic, interpolated evidence derivation. A typical workflow:

• Mode 1: Explicit/Tabular—Acquire categorical evidence via expert-annotated prototypical cases (e.g., medical conditions under different risk profiles).
• Mode 2: Interpolated/Derived—Use algorithmic interpolation (often Bayesian) to fill in intermediate evidence values, as in

$P(a) = P(a|e\land r)P(e)P(r) + \cdots$

Experts may locally override any interpolated cell, providing flexible, explainable, and robust fusion. This directly addresses brittleness in classical tabular systems and the potential misfit of pure interpolation to experiential knowledge.

3. Hierarchical and Distributed Architectures

Dual-mode evidence systems prominently feature in hierarchical and agentic architectures. Vision systems leveraging Dempster–Shafer theory (Li et al., 2013) instantiate a dual process:

• Bottom-Up (Accumulation)—Parallel feature extraction assigns belief masses to low-level cues; these are hierarchically fused (e.g., Dempster’s rule) to yield composite beliefs about object hypotheses.
• Top-Down (Verification)—Hypothesis confirmation propagates expectations downward, seeking or adjusting underlying feature evidence. Such systems achieve robust perception via confidence aggregation and dynamic verification, supported by mathematical mechanisms such as basic probability assignments (BPAs) and evidence mapping:

$Bel(o) = \sum m_e(A_i)m_s(B_j)$

This dual processing is essential for pattern disambiguation under ambiguity or incomplete input.

4. Modal Distinction and Uncertainty in Evidence Interpretation

Advanced dual-mode frameworks handle uncertainty not only in the truth of claims but also in the interpretation of evidence itself (Bjorndahl et al., 2019). Here, an evidence state $e$ is not a fixed set but a family of functions $I_e$ assigning, for each world $x$, a possibly different set $I_e(x)$. Consequently:

• Mode 1 (Objective): $E\varphi$ ("evidence entails $\varphi$") is evaluated by actual entailment: $I_e(x)\subseteq\llbracket\varphi\rrbracket_e$
• Mode 2 (Epistemic): $K\varphi$ ("agent knows $\varphi$") is true if $U_e\subseteq\llbracket\varphi\rrbracket_e$, with $U_e$ the set of worlds coherent with evidence $e$ Bi-modal logics may be further enriched with belief (KD45 modality) and knowability (topology-inspired interior operation), offering rigorous expressive power for modeling empirical error, measurement, and the epistemics of imperfect evidence.

5. Applications: Reasoning, Verification, and Networked Coordination

Dual-mode evidence systems underpin a variety of real-world and research applications:

• Distributed Coordination: In models of coordinated attack (Bucheli et al., 2010), absence of common evidence terms impedes achieving collective action, despite abundance of individual evidence, demonstrating the necessity of explicit common-mode aggregation in distributed protocols.
• Fact/Claim Verification: Modern NLP frameworks (e.g., TwoWingOS (Yin et al., 2018), dual-view models (Wu et al., 2021), multimodal RED-DOT (Papadopoulos et al., 2023)) employ joint dual-wing architectures for simultaneous evidence extraction and verification, combining individual fragment analysis with global, holistic evaluation and attention-based integration.
• Scientific Discovery: Multi-agent and iterative frameworks such as BioDisco (Ke et al., 2 Aug 2025) use dual-mode evidence—knowledge graph structure and automated literature retrieval—for hypothesis generation, with iterative scoring and temporal evaluation to assess novelty and groundedness.
• Integrated Sensing and Communication: Dual-mode ISAC networks (Nabil et al., 19 Feb 2025) exploit monostatic (local) and multistatic (cooperative) sensing, leveraging spatial diversity to compensate for hardware imperfections and maximize sensing/communication throughput.

6. Tradeoffs, Coordination, and Performance Considerations

Dual-mode evidence systems expose tradeoffs and design implications arising from their bidirectional structure:

• Aggregation Thresholds: Transition from individual to group/common knowledge depends on the ability to construct common evidentiary terms, as illustrated in both logic and distributed systems (Bucheli et al., 2010).
• Resource Allocation: In networked ISAC (Nabil et al., 19 Feb 2025), dual-mode sensing's effectiveness is conditioned on parameters such as BS density, power split, beamwidth, and self-interference cancellation; optimal performance is a function of both monostatic and multistatic gains.
• Bias and Robustness: Balancing collective (aggregated, potentially more stable) modes with individual (potentially biased or noisy) inputs, and employing loss functions to penalize divergence (e.g., inconsistency loss (Wu et al., 2021)), enhances both interpretability and reliability.

7. Evaluation, Adaptation, and Future Directions

Evaluation methodologies for dual-mode evidence systems increasingly employ rigorous, statistical, and longitudinal approaches:

• Paired Comparison Models: For scientific hypothesis generation, paired evaluation with Bradley–Terry models provides uncertainty-quantified performance assessment across system configurations (Ke et al., 2 Aug 2025).
• Temporal Validation: Simulating historical knowledge cutoffs and examining subsequent literature for confirmation tests a system's ability to anticipate scientific discoveries, anchoring claims of novelty and support in strict empirical settings.
• Modularity and Extensibility: Many frameworks are designed for modularity, supporting custom model, database, or component integration without loss of core dual-mode architectural virtues.

In sum, the dual-mode evidence system paradigm offers a powerful, extensible, and analytically grounded approach to the challenges of supporting, combining, interpreting, and leveraging evidence in domains characterized by uncertainty, decentralization, or complex coordination requirements.