Geneses Framework in Computational Modeling

• Geneses Framework is a family of computational architectures that co-evolve systems and usage schemes via a dialectical loop between instruments and practices.
• It integrates applications ranging from visual generative modeling and speech enhancement to cognitive memory and automated bioinformatics.
• It emphasizes configurability, transparency, and empirical validation through iterative co-construction, driving state-of-the-art performance across diverse fields.

The Geneses Framework designates a family of computational and methodological architectures characterized by the co-evolution (or genesis) of entities—resources, instruments, models, or representations—together with their corresponding schemes of utilization, interpretation, or control. Across domains, “Geneses” frameworks instantiate dialectical or generative processes in which instruments are simultaneously shaped by, and shaping of, practice, theory, or structure. Key applications range from object-centric generative modeling in vision and audio, to documentational didactics, scientific automation and bioinformatics, interdisciplinary software for the humanities, and large-scale simulation platforms.

1. Theoretical Foundations: Instrumental and Documentational Geneses

The seeds of the Geneses Framework are grounded in Rabardel’s instrumental approach, distinguishing between artefacts (material or digital resources) and instruments (artefacts plus utilization schemes). The concept of documentational genesis formalizes this process by which a document is produced as the union of selected resources and the teacher's invariant scheme of utilization:

$$\text{Document} = \text{Resources} + \text{Scheme of Utilization}$$

Documentational genesis is governed by two interdependent flows:

• Instrumentalization: Users appropriate, adapt, combine, and refine artefacts to fit situated needs;
• Instrumentation: The affordances and constraints of artefacts in turn modify users’ practices and operational beliefs.

This duality generates a dynamic documentation system—a structured, evolving network of interrelated documents and professional knowledge—which can be modeled formally as a dialectical loop:

1. Document₁ (Resources₁ + Scheme₁) produces new resources R₂
2. Document₂ forms from R₂ + Scheme₂
3. The system evolves as $\{\text{Document}_i\}_{i=1}^N$, recursively incorporating prior outputs (Gueudet et al., 2010).

In interdisciplinary digital humanities research, this genesis paradigm underpins collaborative software frameworks, such as timeline-based visualization tools, that co-evolve with user practices—emphasizing configurability and shareability via lightweight, domain-specific languages (DSLs) and schemas directly reflecting scholarly workflows (Aubert et al., 2023).

2. Computational Architectures in Generative Modeling

The Geneses paradigm surfaces in state-of-the-art generative models that prioritize the sequential, object-centric decomposition and synthesis of complex scenes or signals.

Visual Scene Modeling

The GENESIS architecture for visual scene modeling represents an image as a spatial Gaussian mixture model (GMM) over $K$ object-centric components. Each component is parametrized by:

• Latent masks $(z_k^m)$, decoded to $\alpha_k$ spatial masks using stick-breaking normalization;
• Latent appearances $(z_k^c)$, feeding into a neural decoder $g_\mu$ for mean images.

The image likelihood is expressed as:

$$p(x \mid \{z_k\}) = \sum_{k=1}^K \alpha_k(x; z_k^m) \ \mathcal{N}(x;\, \mu_k(z_k^c),\, \sigma_x^2 I)$$

Sequential inference and an autoregressive prior over $\{z^m_k\}_{k=1}^K$ yield improved compositional sampling and coherent scene generation. The training objective is the variational ELBO adapted for this multi-component structure (Engelcke et al., 2019).

Unified Generative Speech Enhancement and Separation

The Geneses framework for speech enhancement and separation employs latent flow matching in VAE space, integrating multi-modal inputs. For a noisy input $y(t)$, the system extracts framewise SSL features $\mathbf{c}$ and predicts latent flow via a multi-modal diffusion Transformer (MM-DiT):

$$\frac{d\mathbf{x}_t}{dt} = \mathbf{v}_\theta(\mathbf{x}_t, \mathbf{c}, t)$$

The ODE is solved from noise to clean speech latents, which are decoded to separate speaker waveforms. Training uses a mean-squared flow-matching objective, and the architecture achieves significant gains in perceptual, separation, and intelligibility metrics under both additive and complex degradations (Asai et al., 26 Jan 2026).

3. Generative Episodic-Semantic Integration in Memory Modeling

The "GENESIS" model in cognitive neuroscience formalizes semantic-episodic memory interaction via dual VAEs and retrieval-augmented generation (RAG):

• Cortical-VAE encodes semantic structure with controllable capacity constraint $C_\text{sem}$;
• Hippocampal-VAE compresses episodic keys, indexed via temporal and semantic cues ($\phi$-weighted);
• RAG retrieves stored episodic codes based on cosine similarity for reconstructive generation.

Capacity constraints ($C_\text{sem}$, $C_\text{epi}$) directly modulate rate-distortion trade-offs, prototypical collapse, recognition/interference effects, and gist-based distortions. Episodic replay and generative recombination account for constructive simulation and memory flexibility (D'Alessandro et al., 17 Oct 2025).

4. Automation and Knowledge Genesis in Science and Bioinformatics

Robot Scientist for Systems Biology

"Genesis" in AI-driven science represents an integrated system spanning hypothesis generation, experimental execution, data analysis, and formal model revision:

• Hardware: 1,008 parallel $\mu$-bioreactors for microbial experiments;
• Automated Analytics: AutonoMS for high-throughput IM-MS, pipelines for quality control, normalization, and Genesis-DB for semantic data integration;
• Model-Revision Ontology: RIMBO documents all discrete model changes in OWL;
• Relational Learning: LGEM+ performs logic-based, abductive model improvement and reconciliation with experimental phenotypes, e.g., via abductively generated FOL clauses and flux-balance analysis constraints (Tiukova et al., 2024).

Constraint-Based Causal Discovery in Multi-omics

The GENESIS algorithm leverages the fixed causal precedence of genotypes to constrain causal discovery in gene-expression networks. The method constructs an ancestrality matrix $M_{ij}$ through constraint-based, forward induction, using a series of marginal and conditional independence tests with growing Markov blankets. The result is a provably sound, sparse partial ancestral order that restricts downstream search complexity and aligns with biological directionality (Asiedu et al., 21 May 2025).

5. Interdisciplinary Extensions and Instrument Genesis in Humanities

The Geneses framework in the digital humanities operationalizes instrumental genesis in collaborative software artifacts, exemplified by Advene’s visualization system:

• Data is structured as XML annotations with controlled vocabularies and hierarchies;
• Visualization configuration is embedded in a lightweight, URL-parametrized DSL, promoting reproducibility and interpretability;
• Iterative user–developer co-construction fosters adaptation and refinement, aligning data schemas with domain methodologies (eMAEX) and scholarly annotation workflows;
• This framework has supported cross-study synthesis, visual argumentation, and scalable comparison of video annotation data (Aubert et al., 2023).

6. Physical Modeling and Sound Synthesis

GENESIS3 (G3) extends the mass–interaction network paradigm for musical composition using the CORDIS-ANIMA formalism:

• Models are authored as scalable, modular graphs of masses (nodes) and interaction laws (edges), assembled and manipulated via a dedicated Tcl-based scripting language (PNSL) with pattern-matching;
• The system implements multithreaded simulation, zoomable user interfaces, adaptive module grouping, and multimodal visualization/sound export;
• The architecture brings compositional control over physical modeling, supporting extensive algorithmic sound design and computational creativity (0911.4642).

7. Synthesis: Common Patterns and Principles

Despite substantial domain heterogeneity, Geneses frameworks share several structural and conceptual invariants:

• Dialectical co-evolution: Systems and their usage schemes are continuously co-constructed, whether at the level of software, instrument, or scientific model.
• Explicit formalizations: Each framework codifies the transformation of resources/information into operational structures—be it scene latents, documentation systems, or causal graphs—often with provable soundness or generative capacity guarantees.
• Instrumentation/Instrumentalization Loop: The bi-directional mediation of tool affordances and user practices is central, extending across educational, scientific, engineering, and creative domains.
• Configurability, Transparency, and Extensibility: Architectures are deliberately lightweight, scriptable, and accessible to non-programmer domain experts, supporting continuous adaptation and methodological alignment.
• Empirical validation: Across visual, audio, biological, and documentational applications, Geneses frameworks achieve state-of-the-art or benchmarked performance, demonstrating robustness, compositional coherence, and/or scientific utility.

Collectively, Geneses frameworks operationalize a unifying design principle: the genesis of knowledge, structure, or signal is a recursive, generative process coupling system, instrument, and interpretive practice within a formal, extensible infrastructure.