MOSAIC: A Multi-Domain Paradigm

• MOSAIC is a multi-domain framework that integrates quantum lattice models, advanced astronomical instrumentation, computational geometry, and AI systems to address complex scientific challenges.
• In quantum matter, mosaic lattices employ modulated potentials to reveal robust topological phases and Majorana modes through even/odd dichotomy analyses.
• In AI and robotics, MOSAIC frameworks optimize multi-modal learning and agent-based simulations, enabling scalable, real-world applications and efficient system design.

Mosaic (MOSAIC) refers to a diverse set of technical concepts, methods, and systems across domains—physics, astronomy, computational geometry, machine learning, computer vision, robotics, agent-based simulation, and system design. The term is used for specialized lattice models (often with modulated potentials or hoppings), high-multiplex astronomical instrumentation, agentic architectures, computational geometry algorithms, and several advanced methods for learning, generation, and simulation. The following sections synthesize the principal definitions, frameworks, outcomes, and technical architectures from recent arXiv literature.

1. Mosaic Lattices in Quantum Matter

“Mosaic lattice” models refer to 1D or higher-dimensional physical systems where either on-site potentials or hopping amplitudes are selectively modulated at regular or quasi-regular intervals. In one-dimensional mosaic-Kitaev chains, onsite potentials act only on equally spaced sites; periodic (commensurate), quasiperiodic, and random mosaic patterns can all be constructed. The prototypical Hamiltonian for a mosaic-Kitaev chain is

$$H = \sum_{j=1}^{L-1}\Bigl[ -t\,c_{j+1}^\dagger c_j + \Delta\,c_{j+1}^\dagger c_j^\dagger + \textrm{H.c.} \Bigr] - \mu\sum_{j=1}^L c_j^\dagger c_j + \sum_{j=1}^L V_j c_j^\dagger c_j,$$

where $V_j$ is only nonzero for $j = n p$ with modulation interval $p$, and can take periodic, quasiperiodic, or random forms (Zeng et al., 2020, Zeng et al., 2021).

In the off-diagonal mosaic lattice, hopping terms are modulated at regular intervals, leading to complex localization and topological phenomena (Zeng et al., 2021). For commensurate modulations ($\alpha \in \mathbb{Q}$), Berry (Zak) phases can rigorously characterize phase transitions and the existence of zero and nonzero-energy edge modes. For incommensurate (quasiperiodic) modulations, Anderson localization results purely from off-diagonal disorder.

A profound result in the mosaic-Kitaev model is an even/odd dichotomy: when the onsite modulation interval $p$ is even, the topological superconducting phase persists for any finite modulated onsite potential (i.e., Majorana zero modes are robust to arbitrary disorder); when $p$ is odd, a critical potential strength exists at which a topological-to-trivial phase transition occurs, with analytical expressions for $V_c$ and explicit transfer matrix construction for topological invariants (Zeng et al., 2020).

2. MOSAIC in Astronomical Instrumentation

MOSAIC designates the premier multi-object spectrograph (MOS) for the 39 m Extremely Large Telescope (ELT). The instrument’s concept, architecture, and technical innovations are documented in depth (Hammer et al., 2016, Puech et al., 2018, Morris et al., 2018, Sánchez-Janssen et al., 2020, Rodrigues et al., 2016). MOSAIC integrates high-multiplex, multi-Integral Field Unit (multi-IFU) modes for simultaneous intermediate- and high-resolution spectroscopy in visible and near-infrared regimes, over a patrol field of $40\ \mathrm{arcmin}^2$.

The primary observing modes are:

• HMM (High-Multiplex Mode): $\sim$200 single-aperture fibre feeds for faint-object surveys.
• HDM (High-Definition Mode): $V_j$0 MOAO-fed IFUs for small-scale, spatially resolved studies.
• VIFU and HMM-Vis/HMM-NIR: Flexible IFU and fibre bundle configurations for visible and near-IR.

Key technical features include step-tilted, non-telecentric focal planes, fibre positioner arrays, cross-beam pairing for accurate sky subtraction, MOAO and GLAO adaptive optics, rapid sub-array detector readouts, and modular spectrograph designs. Science programs include galaxy evolution, reionization epoch studies, mapping of the intergalactic and circumgalactic medium, resolved stellar population analysis, and exoplanet demographics. The architecture supports spectral resolutions up to $V_j$1 and wavelength coverage from 0.37–2.5 μm (Puech et al., 2018, Sánchez-Janssen et al., 2020). MOSAIC is the reference for ELT-class large-area, multiplexed spectroscopy and is planned to launch in the late 2020s.

3. MOSAIC in Computational Geometry

Mosaic in computational geometry refers to methods for constructing area-closed partitions of spherical surfaces by intersection with Cartesian grids—a problem arising in conservative data transfer between volumetric and spherical boundary representations (e.g., in MHD, atmospheric, or space weather models). The algorithmic stages are:

1. Identification of intersecting Cartesian cells via corner-straddle and face-penetration geometric tests.
2. Construction of spherical “prepatches” representing exact cell-sphere intersections, including all degenerate geometric configurations (e.g., doubly-crossing edges, lens-shaped patches).
3. Sequential “splicing” by colatitude (θ-patching) and then azimuth (φ-patching), with explicit treatment of polar singularities and exact great-circle/meridian intersection computations.
4. Guarantee of area closure to roundoff via additive spherical-polygon area formulae, supporting conservative coupling.

The reference implementation (mdi-mosaic, Java/Maven) produces machine-precise mosaics, robustly handling thousands of cells and patch boundaries without holes or overlaps (Counts et al., 14 May 2026).

4. MOSAIC Frameworks in Machine Learning and AI

Multiple recent systems employ the MOSAIC acronym:

a. Multi-Objective Slice-Aware Iterative Curation for Alignment

MOSAIC provides a closed-loop framework for optimizing the mixture of fine-tuning data to achieve Pareto-optimal tradeoffs between multi-turn safety (XGuard metric), minimal over-refusal (OrBench), and instruction-following under constraints (IFEval), all under strict data budgets. The framework:

• Decomposes evaluation into slices and atomic metrics via a unified L1–L3 interface.
• Iteratively updates mixture ratios and bucket weights using slice-level failure profiles.
• Empirically achieves internal XGuard increase from 2.76 to 4.67 with negligible OrBench and IFEval tradeoff, outperforming random static mixtures and generalizing better on external adversarial and capability benchmarks (Dou et al., 19 Mar 2026).

b. Multi-Subject Personalized Generation: Correspondence-Aware Synthesis

MOSAIC in personalized generation targets the challenge of synthesizing images conditioned on multiple distinct reference subjects while preserving semantic correspondence and feature disentanglement. The system introduces:

• Explicit point-to-point semantic correspondence annotation (SemAlign-MS dataset).
• A semantic correspondence attention loss that enforces correct alignment between references and generated output.
• A multi-reference disentanglement loss that orthogonalizes attention subspaces, preventing feature/identity blending as the number of references increases.

State-of-the-art results are achieved for $V_j$2 subjects, with plug-and-play LoRA design and strong identity consistency even as scene complexity increases (She et al., 2 Sep 2025).

c. Multi-Modal Supervision-Aware Contrastive Learning

MoSAiC optimizes for multi-modal and multi-label satellite imagery, jointly leveraging intra- and inter-modality contrastive invariance as well as explicit multi-label supervised contrast. The architecture comprises dual modality encoders, paired projection heads, and late fusion for classification:

• Intra-modality: SimCLR-style invariance within each modality.
• Inter-modality: Alignment of co-registered SAR and optical images.
• Multi-label supervision: Extension of SupCon to multi-hot labels, pulling together patches that share any class.

Empirically, MoSAiC achieves large gains over both fully supervised and classic self-supervised methods for multi-label macro/micro F1 and cluster coherence, handling high inter-class similarity and label scarcity in large remote-sensing datasets (Gupta et al., 11 Jul 2025).

5. Agentic and System Architectures: Social Simulation, Robotics, and Clinical Annotation

MOSAIC has also been adopted for large-scale system architecture:

• In assistive robotics, a modular system for collaborative cooking leverages LLMs (for language and planning), vision-LLMs (for open-vocabulary object retrieval), RL and IK-based control policies, and multi-agent task scheduling for fluid human–robot interaction and coordinated multi-robot management. System evaluation demonstrates 68.3% completion over complex collaborative trials (Wang et al., 2024).
• In social simulation, MOSAIC constitutes a simulation platform using LLM-powered agents embedded in a social graph to study content dissemination, engagement, and moderation (community-based, third-party, and hybrid fact-checking). The system tracks emergent dynamics, content veracity, engagement metrics, and the divergence between agent-reported rationales and global outcomes. Results highlight the efficacy of hybrid moderation protocols for suppressing misinformation and the mismatch between individual reasoning and emergent dynamics (Liu et al., 10 Apr 2025).
• In clinical communication coding, an agentic architecture (LangGraph orchestrated) decomposes large-scale multi-framework annotation into Plan, Update, Annotation, and Verification agents. The system achieves F1 ≈ 93%, exceeding typical human inter-rater reliability, and features extensibility, on-the-fly codebook updating, and transparent RAG-based prompt construction (Yang et al., 9 Dec 2025).

6. Design Principles in Simulation, Workflow, and Community Systems

The mosaic (“tesserae”) paradigm has influenced co-simulation tools and creative platforms. “Tesserae” groupings in mosaik allow visual, systematic creation and wiring of entity sets in simulation GUIs, simplifying large-scenario design with automatic multidirectional data flow consistency. The abstraction enables drag-and-drop creation, always mapping to the same entity/connect API as script-based approaches, thereby preserving flexibility and reproducibility (Schulte et al., 15 Apr 2026).

In creative online communities, Mosaic shifts the focus from outcome-based sharing to processes documented via work-in-progress (WIP) snapshots and reflections. Quantitative studies reveal engagement patterns, feedback norms, and reduced apprehension in sharing unfinished work. The system design foregrounds process and discovery, signaling critique appetite and supporting technique-based exploration (Kim et al., 2016).

Table: Representative Technical Domains and MOSAIC Systems

Domain	System/Model	Reference
1D Topological Phases	Mosaic-Kitaev, off-diagonal lattices	(Zeng et al., 2020, Zeng et al., 2021)
Astronomical Instrumentation	ELT MOSAIC spectrograph	(Sánchez-Janssen et al., 2020, Puech et al., 2018, Hammer et al., 2016)
Comput. Geometry	Area-closed spherical mosaics	(Counts et al., 14 May 2026)
Multi-Objective ML	Slice-aware data curation/alignment	(Dou et al., 19 Mar 2026)
Vision/Generation	Personalized multi-reference image synth	(She et al., 2 Sep 2025)
Remote Sensing	Multi-label, multi-modal contrastive CL	(Gupta et al., 11 Jul 2025)
Robotics	Modular human–robot collaborative system	(Wang et al., 2024)
Social Simulation	LLM-based agent simulation, moderation	(Liu et al., 10 Apr 2025)
Clinical NLP	Agentic multi-codebook annotation system	(Yang et al., 9 Dec 2025)
Simulation Design	Tesserae-based co-simulation GUI	(Schulte et al., 15 Apr 2026)
Creative Process	Online process-oriented art community	(Kim et al., 2016)

7. Impact and Future Directions

Mosaic concepts unify a range of advances in specialized lattice engineering, astronomical instrumentation, system interaction protocols, and representation learning. In quantum matter, the mosaic modulation enables new forms of disorder-robust topological protection and localization. In astronomy, MOSAIC sets the standard for high-multiplex multi-band spectroscopy in the era of extremely large telescopes. Closed-loop, slice-aware MOSAIC frameworks offer actionable and interpretable optimization in multi-objective supervised learning for AI alignment. Systems design in robotics and agent-based simulation leverages modular decomposition, open-vocabulary abstraction, and agent orchestration for dynamic, extensible workflows.

Continuing research explores scaling (to higher dimensions, larger agent systems), increased automation (online annotation learning, error recovery), cross-domain transfer (from visual correspondences to multimodal contexts), and new modes of user/system interaction, driven by the advances consolidated under the “mosaic” conceptual umbrella.