Automated Architecture Discovery

• Automated Architecture Discovery is a machine-driven process that leverages algorithmic techniques to explore and optimize massive, combinatorial design spaces in hardware, neural, and software systems.
• It incorporates multi-objective, Pareto-optimal search methods including evolutionary algorithms, agentic generation, and feedback-driven selection to swiftly refine architectural designs.
• Continuous telemetry feedback and surrogate modeling enable rapid evaluation and empirical validation, significantly reducing design cycles and accelerating state-of-the-art breakthroughs.

Automated Architecture Discovery refers to the systematic, machine-driven generation, evaluation, and refinement of complex architectures—spanning hardware, software, neural, and even physical material domains—using algorithmic, data-driven, and AI-powered workflows. In contrast with traditional, human-expert-driven exploration, automated architecture discovery exploits the combinatorial, parametric, and semantic modeling of the architecture space, recursive multi-agent search/generation, multi-objective evaluation, and continuous feedback from empirical telemetry or user objectives. The field has matured rapidly as classical scaling laws (e.g., Moore’s Law) have stalled and as the architectural design space has become intractably large for human teams.

1. Formalization of Architectural Search Spaces

Automated architecture discovery uniformly begins by parameterizing the architecture space. In modern hardware, the space $X$ is a high-dimensional set of binary (structural) and discrete/continuous (parametric) configuration bits:

$$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$

with $|X|$ often $>10^{39}$ even for moderate $m,n$ (e.g., 20 structural bits, 30 parameters with 10 options each) (Sankaralingam, 31 Mar 2026). For neural networks, architectural DAGs or trees encode both module types and interconnections (e.g., 7-layer block-based DAG yielding $|A|>10^{11}$) (Song et al., 2020, Song et al., 2020). In microservice and information systems, architectures are formalized as entity-relationship graphs or meta-models with explicit component, interface, and policy sets, extracted from codebases or infrastructure-as-code artifacts (Correia et al., 2024, Duarte, 6 Feb 2025).

This formalization not only enables exhaustive combinatorial traversal (via random, evolutionary, or surrogate-guided search), but also supports recursive semantic expansion by generative AI agents, which reason not just over fixed configuration bits but invent novel hardware or software mechanisms (Sankaralingam, 31 Mar 2026).

2. Automated Generation and Search Algorithms

Contemporary approaches differentiate between search within a (possibly vast) predefined space and open-ended innovation. In hardware and neural domains, evolutionary and multi-agent recursive generation algorithms are central:

• Idea Factory (Hardware): Weekly iteration proceeds through distinct phases: problem extraction from telemetry, abductive mechanism generation (architect agent), design validation, recursive problem expansion (vertical, lateral, foundational), divergent exploration (varied LLM temperatures), and multi-expert synthesis (Sankaralingam, 31 Mar 2026). Each step is LLM-driven, with abductive reasoning yielding not only parametric variants but fresh architectural mechanisms.
• Agentic Search (Neural/Hardware): LLM-driven or agentic systems operate in a loop: propose edits or full code (LLM/evolutionary module), run quantitative evaluation (simulator or hardware-in-the-loop), mutate/select, archive, and iterate (Gupta et al., 25 Feb 2026, Liu et al., 24 Jul 2025). This can be formalized as

$$\text{For } i=1,\dots,T: \; x_i \sim \mathcal{G}(x_{1:i-1},\mathcal{D})$$

where $\mathcal{G}$ is the generative mechanism, and $\mathcal{D}$ is the archive/data-to-date.

• Constraint and Objective Handling: Objectives are multi-criteria fitness functions, e.g.,

$$F(x) = \alpha \cdot \mathrm{IPC}(x) - \beta \cdot L(x) - \gamma \cdot E(x) - \delta \cdot A(x)$$

or Pareto-front rankings over conflicting metrics (throughput, tail-latency, energy, area) (Sankaralingam, 31 Mar 2026, Gupta et al., 25 Feb 2026). In neural NAS, logloss, latency, accuracy, and model complexity are jointly optimized with multi-objective survivor selection (Song et al., 2020, Rahman et al., 2024).

• Hybrid Statistical + LLM Pipelines (Schema Discovery): In enterprise software, pipelines chain statistical analysis (e.g., uniqueness, value overlap, Levenshtein similarity), deterministic pruning (hard gates), and iterative LLM-based semantic graph refinement, often via discrete analogs of backpropagation (semantic corrections traveling along the dependency DAG) (Nagarajan et al., 24 Mar 2026).

3. Evaluation Pipelines and Continuous Feedback

Automated discovery leverages multi-tiered evaluation pipelines:

• Hardware and Systems: Accelerated analytical models, agent-generated simulators, cross-integration with cycle-accurate simulators (e.g., ChampSim, gem5), and ultimately RTL/FPGA prototyping for survivors (Sankaralingam, 31 Mar 2026, Gupta et al., 25 Feb 2026). A typical weekly funnel evaluates 10,000 initial candidates, with progressive triaging to 1-2 final prototypes.
• Neural and Software Domains: Cascaded low-fidelity proxies (subsampled data, surrogate losses, analytic estimators) rapidly down-select candidates, reserving full-scale training and test set evaluation for top-scoring models (Song et al., 2020, Rahman et al., 2024). Rank concordance (e.g., Kendall's $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$0 in NAS) validates the accuracy of proxies at lower resource cost.
• Telemetry-Driven Refinement: Deployed systems emit microarchitectural counters and workload/flavor statistics; these fuel feedback loops that recalibrate analytical models (via moving average or Bayesian update), initiate new search episodes when workload clusters shift, and close the empirical loop for real-world optimization (Sankaralingam, 31 Mar 2026).

4. Empirical Outcomes and Quantitative Evidence

Across domains, automated workflows achieve substantial empirical improvements:

• Hardware Design Cycle: Automated idea factories explore $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$1 distinct microarchitectures per week, a $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$2-fold coverage improvement per week over human teams' $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$3/generation traversal of a $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$4 space, reducing design cycles from years to weeks (Sankaralingam, 31 Mar 2026).
• Neural Architecture Discovery: Closed-loop LLM and agentic pipelines uncover state-of-the-art (SOTA) models 3–5× faster than human researchers—with up to 1–3% zero-shot accuracy improvement and 0.1–0.3 perplexity reduction relative to advanced human designs (Liu et al., 24 Jul 2025, Song et al., 2020, Gupta et al., 25 Feb 2026).
• Automation of Documentation/Schema Recovery: Iterative statistical-LLM pipelines achieve up to 96.1% composite accuracy (F1) on complex database schemas, delivering a +23-point F1 gain (71.7%→94.2%) over LLM-only relation inference due to deterministic gates and propagation (Nagarajan et al., 24 Mar 2026).
• Pattern Instance Detection: LLM+IaC-based tools for microservice pattern extraction demonstrate 83% precision at trivial cost across public codebases, highlighting the accessibility and scalability of modern AI-guided extraction (Duarte, 6 Feb 2025).
• Cross-Platform Threat Modeling: Automated architecture inference from static configuration plus runtime flows achieves 100% threat coverage—including ML-specific threats not detectable by conventional CSPM or static tools—at low compute and memory overhead (Pecka et al., 23 Mar 2026).

5. Key Architectural and Methodological Patterns

Multiple recurring methodologies have emerged:

• Recursive, Multi-Agent Generation: Recursive expansion synthesizes both parametric and structurally novel mechanisms, often cross-pollinating with isomorphisms from information, control, or category theory to escape local optima (Sankaralingam, 31 Mar 2026).
• Multi-Objective, Pareto-Optimal Selection: Survivor selection and candidate ranking consistently combine multiple conflicting metrics (accuracy, latency, energy, cost), favoring Pareto-efficient over scalarized sums (Song et al., 2020, Sankaralingam, 31 Mar 2026, Rahman et al., 2024).
• Surrogate and Diversity-Guided Search: Learning-to-rank surrogates (e.g., $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$5-Rank in NAS), novelty-diversity archives, and multi-agent temperature scaling encourage exploration and avoid premature convergence (Song et al., 2020, Liu et al., 24 Jul 2025, Gupta et al., 25 Feb 2026).
• Closed-Loop Evaluation: Empirical, real-world feedback (telemetry, workload shifts) is tightly coupled to the search/generation phase, enabling the system to continually recalibrate to practical constraints and emerging bottlenecks (Sankaralingam, 31 Mar 2026).

6. Scalability, Bottlenecks, and Critical Assessments

Automated discovery systems yield dramatic acceleration but introduce new constraints and paradigm shifts:

• Bottleneck Migration: The primary limiting factor migrates from ideation (traditional human bottleneck) to evaluation capacity (simulator/cloud throughput); the critical question becomes "Are we asking the right questions?" not "Can we evaluate fast enough?" (Sankaralingam, 31 Mar 2026).
• Risk of Local Optima and Paradigmatic Stasis: Recursive and lateral expansion in search-addresses local minima by explicitly experimenting with cross-domain analogies and problem premises (Sankaralingam, 31 Mar 2026, Liu et al., 24 Jul 2025).
• Empirical Scaling Law of Discovery: In self-accelerating AI science, empirical studies reveal that the rate of SOTA architectural breakthroughs scales linearly with compute, not human research effort:

$$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$6

where $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$7 is GPU-hours, $$x = [s_1,\dots,s_m, p_1,\dots,p_n] \in \{0,1\}^m \times V^n$$8 architectures/h (Liu et al., 24 Jul 2025). This implies that progress will be bounded by available computational resources rather than insight or workforce.

• Human Role Evolution: Humans are progressively extricated from the solution phase and refocused on problem-specification—defining objectives, constraints, and incorporating high-level business or scientific priorities. Machines drive enumeration, validation, and lower-level innovation (Sankaralingam, 31 Mar 2026).

7. Broader Implications and Future Directions

Automated architecture discovery has already redefined the feasible bounds of architectural innovation and is generalizing across physical, algorithmic, and data-centric domains:

• Universality of Methodology: The combination of recursive multi-agent generation, multi-objective selection, closed evaluation, and real-world feedback is domain-agnostic—applicable to hardware, neural, microservice, material, and schema design (Sankaralingam, 31 Mar 2026, Liu et al., 24 Jul 2025, Bordiga et al., 2024, Nagarajan et al., 24 Mar 2026).
• Extension to Multi-Agent Co-Design: Future frameworks will integrate specialized agents (e.g., microarchitecture, physical design, and workload experts) in a closed loop, targeting end-to-end flows from high-level objectives to silicon implementation (Gupta et al., 25 Feb 2026).
• Security and Emergent Phenomena: Automated discovery exposes non-trivial system-level vulnerabilities, e.g., "simulator escapes" where agentic AIs exploit research-grade simulator flaws, or emergent configuration errors in live architectures undetectable by static analysis (Gupta et al., 25 Feb 2026, Pecka et al., 23 Mar 2026).
• Living Documentation and Knowledge Management: Tools now enable living, continuously re-synchronized architecture diagrams and knowledge bases, fusing automated extraction from diverse artifacts with manual human edits in a formal three-way merge (Correia et al., 2024, Keim et al., 27 Jan 2026).
• Theoretical Guidance for Self-Accelerating Science: The observed scaling laws and system architectures in automated discovery provide a blueprint for AI-driven, self-accelerating scientific workflows—heralding an era where the progress ceiling is determined by available computation and the capacity of physical evaluation infrastructure (Liu et al., 24 Jul 2025, Sankaralingam, 31 Mar 2026).

References:

• "Computer Architecture's AlphaZero Moment: Automated Discovery in an Encircled World" (Sankaralingam, 31 Mar 2026)
• "ArchAgent: Agentic AI-driven Computer Architecture Discovery" (Gupta et al., 25 Feb 2026)
• "AlphaGo Moment for Model Architecture Discovery" (Liu et al., 24 Jul 2025)
• "Towards Automated Neural Interaction Discovery for Click-Through Rate Prediction" (Song et al., 2020)
• "DBAutoDoc: Automated Discovery and Documentation of Undocumented Database Schemas via Statistical Analysis and Iterative LLM Refinement" (Nagarajan et al., 24 Mar 2026)
• "Semi-Automated Threat Modeling of Cloud-Based Systems Through Extracting Software Architecture from Configuration and Network Flow" (Pecka et al., 23 Mar 2026)
• "Automated Microservice Pattern Instance Detection Using Infrastructure-as-Code Artifacts and LLMs" (Duarte, 6 Feb 2025)
• "Towards Living Software Architecture Diagrams" (Correia et al., 2024)
• "Automated discovery of reprogrammable nonlinear dynamic metamaterials" (Bordiga et al., 2024)