MR-Search: Efficient Retrieval in MR & Mixed Reality

• MR-Search is a comprehensive framework combining retrieval algorithms and context-aware interaction in Magnetic Resonance and Mixed Reality environments.
• Experimental findings reveal that physical clutter and dual-tasking elevate reaction times and subjective workload, while repeated spatial layouts can improve search efficiency by around 20%.
• Algorithmic strategies such as summed-area tables, multi-resolution zoom searches, and meta-reinforcement learning enable efficient region retrieval and adaptive search in complex MR contexts.

MR-Search encompasses a spectrum of methodologies and systems for efficient, robust, and context-adaptive search tasks in Magnetic Resonance (MR) environments and Mixed Reality (MR). The term refers to a family of approaches including retrieval of medical images, fast region search in biomedical volumes, secure ranked document retrieval, meta-reinforcement learning for agentic in-context search, and context-optimized search interface design in MR visualization. These approaches address algorithmic, statistical, and human–machine factors fundamental to information retrieval and context-aware interaction in MR-centric applications.

1. Experimental Paradigms and Variable Manipulations

MR-Search is studied in both computational and human-computer interaction contexts. In mixed reality MR environments, visual search is typically formalized with multi-factor experimental designs. For example, in the 2 × 2 × 2 factorial design of (Lai et al., 21 Feb 2026), independent variables include physical environment complexity (simple/complex furnishings), virtual element depth layout (fixed plane/variable depth), and secondary task presence (single-task/dual-task with concurrent auditory counting). Spatial layout configuration (repeated/novel) introduces contextual cueing conditions per block.

Dependent variables encompass log-transformed reaction time (RT), accuracy (as correct orientation classification), and subjective workload via NASA-TLX six-dimension subscales. Hardware and software implementations use stereoscopic head-mounted displays (e.g., Meta Quest 3), auto-recorded key logs, and controlled open-plan lab environments.

2. Statistical Models and Analysis Methods

The MR-Search paradigm employs rigorous statistical frameworks for isolating and interpreting effects. Linear mixed-effects models (LMMs) are deployed for continuous outcomes like log(RT):

$$\log(RT_{ijk\ell}) = \beta_0 + \beta_1\,\mathrm{ComplexEnv}_i + \beta_2\,\mathrm{DiffDepth}_j + \beta_3\,\mathrm{DualTask}_k + \beta_4\,\mathrm{RepeatedLayout}_\ell + \text{interactions} + u_{\mathrm{participant}} + \varepsilon_{ijk\ell}$$

Generalized linear mixed-effects models (GLMMs; binomial logit link) analyze trial-level accuracy, while cumulative link mixed models (CLMMs) address ordinal NASA-TLX ratings. All fixed effects, including up to four-way interactions, are modeled, with participant random intercepts isolating individual variation. Tukey-adjusted estimated marginal means (EMMs) and Cohen’s $d$ quantify contrasts, while odds ratios (OR) summarize categorical effects.

3. Key Findings on Visual Search and Cognitive Factors

In MR visual search, physical scene clutter (complex vs. simple) increases RT by approximately 17% ($\exp(0.157)-1$), and variable virtual depth introduces additional RT cost, although depth and environment effects interact sub-additively (i.e., combined cost is less than additive; $B=-0.056,\,p=4.54\times10^{-3}$). Repeating spatial layouts yields strong contextual cueing, with repetition affording an ≈20% RT improvement ($\beta=-0.425,\,d\approx0.2$). However, these contextual benefits are reduced by complexity and depth variation. Dual-tasking elevates subjective workload across all NASA-TLX subscales (e.g., $p=1.9\times10^{-6}$ for mental demand), even when global RTs do not significantly worsen, indicating dissociation of objective and subjective costs.

Accuracy decrements via environmental complexity ($\mathrm{OR}=0.716,\,p=3.73\times10^{-4}$) but not depth or task, except that spatial repetition offsets dual-task-induced accuracy drops ($\mathrm{OR}=1.289,\,p=0.0163$) (Lai et al., 21 Feb 2026).

4. Implicit Spatial Memory and Contextual Cueing

MR-Search establishes that implicit memory of spatial regularities robustly accelerates visual search. Across all tested environments, log(RT) reductions for repeated vs. novel virtual layouts ($p<10^{-6}$ to $10^{-3}$, $d$0–0.48) are observed, yet forced-choice recognition reveals high error rates (30–45 %), confirming that the memory is implicit. This dissociation mandates system designs that leverage layout consistency for performance without relying on users’ explicit awareness or recall (Lai et al., 21 Feb 2026).

5. Design Implications for MR Search Systems

Interface guidelines emergent from MR-Search findings include:

• Preserve fixed spatial layouts of critical virtual elements to leverage implicit contextual cueing, especially where multitasking is unavoidable.
• In simple physical settings, anchor all digital content to a single depth plane to maximize search speed. In complex, cluttered settings, introduce depth stratification to segment virtual from real clutter, exploiting sub-additive interaction for efficiency recovery.
• Avoid unnecessary depth variation under heavy cognitive/multitasking loads, as these re-introduce RT penalties.
• For highly cluttered real-world backgrounds, reduce virtual content density or employ depth-based grouping. Where possible, physically simplify real-world backgrounds (e.g., dim irrelevant objects).
• Account for the dissociation between measurable search efficiency and perceived workload: dual-task scenarios can be objectively efficient but subjectively taxing, so workload-aware design is paramount.
• Leverage layout regularities to offset dual-task and clutter costs, but do not rely on explicit user training or recognition—benefits are implicit.

6. Systemic and Algorithmic Generalizations

MR-Search methodologies extend beyond human–computer interaction. Algorithmic MR-Search frameworks, as in fast tumor region retrieval (Takeshima et al., 1 Oct 2025), federated model search (Wu et al., 2023), or encrypted ranked retrieval (Liu et al., 2019), share a focus on reducing search complexity, maximizing discrimination, and adapting credit or weight assignments in uncertain and multi-context environments. Techniques such as summed-area tables for O(1) region-summarization, multi-resolution zoom search in large parameter spaces, and meta-RL with explicit self-reflection harness both computational and behavioral regularities for search optimization.

7. Practical Impact and Future Directions

MR-Search results redefine MR interface design and retrieval frameworks by quantifying distinct contributions of environmental, spatial, and cognitive variables to task efficiency and subjective load. Future research directions include domain-generalization in feature learning for CBIR (Tobari et al., 2 Jan 2025), automated architecture discovery for MR image reconstruction (Wu et al., 2023, Yan et al., 2020), and meta-agent frameworks for autonomous search strategy development (Xiao et al., 11 Mar 2026). Ongoing work also explores agentic search in MR sequence development, with LLM-driven autoresearch frameworks that automatically generate and validate MR pulse sequences using controller–developer–validator loops and structured feedback (Zaiss et al., 14 Apr 2026). These findings collectively establish MR-Search as both a methodological paradigm and a foundational principle in the optimization of MR-centric search tasks in cognitive, computational, and agentic systems.