MAS-CPA for Cosmological Parameter Analysis

• The paper introduces a robust MAS-CPA framework that modularizes cosmological workflows using specialized agents and hybrid optimization methods.
• It employs PSO, grid-based algorithms, and MCMC integration to efficiently explore high-dimensional parameter spaces with near-perfect parallel scaling.
• MAS-CPA significantly improves speed, scalability, and reproducibility, reducing human oversight while streamlining next-generation cosmological analyses.

A Multi-Agent System for Cosmological Parameter Analysis (MAS-CPA) integrates specialized autonomous or semi-autonomous agents—most often software entities, but also swarms of candidate solutions in stochastic optimization—through structured and cooperative interaction schemes to automate, accelerate, or enhance the inference of fundamental cosmological parameters from observational or simulated data. These architectures have emerged as essential tools to manage the increasing complexity, data volume, and theoretical demands characteristic of modern cosmology. Implementations span a wide spectrum, from bio-inspired optimization (Particle Swarm Optimization) and parallel grid-based methods, to LLM-driven agentic toolchains that orchestrate literature-to-simulation pipelines and terabyte-scale ensemble inference.

1. Multi-Agent Architectures in Cosmological Analysis

MAS-CPA frameworks typically modularize cosmological workflows into distinct agent roles, each specializing in a core analytical or computational function. Examples include:

• PSO-based Systems: Here, each "particle" is a cognitive agent representing a parameter vector $x_i \in \mathbb{R}^n$, evolving through stochastic, memory-based protocol. Communication occurs via star (global best) or ring (local best) topologies, with group adaptation accelerating convergence to high-probability regions (Hernández et al., 7 Aug 2025, Prasad, 2014, Prasad et al., 2011).
• LLM-driven Agentic Pipelines: Frameworks such as cmbagent deploy LLM-based planners, retrieval agents, coders, and executors. Managerial agents recursively decompose high-level objectives (e.g., "infer cosmological parameters from ACT DR6") into granular sub-tasks distributed across retrieval-augmented, code-writing, and execution agents (Laverick et al., 2024).
• Simulation Setup and Extraction Systems: SimAgent couples a physics-reasoning agent (for literature parsing), a simulation validator agent (for software compliance), and a tool-execution agent (for analysis automation), communicating via synchronous, round-based message passing to deliver executable simulation configurations (Zhang et al., 11 Jul 2025).
• TB-scale Ensemble Analysis: InferA leverages a supervisor agent orchestrating planning, data loading, SQL, Python, visualization, QA, and documentation agents, enabling scalable, provenance-traceable inference on HACC ensemble datasets and similar (Tam et al., 14 Oct 2025).
• Agentic ML Investigation: AI Cosmologist organizes planning, coding, execution, analysis, and synthesis agents under a looped hypothesis-generation and refinement protocol, automatically generating and selecting machine learning pipelines for cosmological regression or classification tasks (Moss, 4 Apr 2025).

A common thread is the explicit division of labor and communication protocol—either via synchronous message passing or implicit state-sharing—to realize robust, scalable, and sometimes human-in-the-loop inference pipelines.

2. Optimization and Sampling Methodologies

MAS-CPA systems employ a diverse range of underlying inference strategies, notably:

• Particle Swarm Optimization (PSO): Each agent (particle) updates its velocity and position using the canonical rule,

$$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$

$x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$

where $w$ is the inertia weight, $c_1$, $c_2$ are acceleration coefficients, $r_1$, $r_2$ are uniform random numbers, $p_i$ is particle $i$'s personal best, and $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$0 is the global or local neighborhood best (Hernández et al., 7 Aug 2025, Prasad et al., 2011, Prasad, 2014). The collective evaluates the chi-squared-based likelihood $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$1 or its negative log-likelihood equivalent for observational fit.

• Parallel Grid-based "Snake" Algorithm: A master-slave architecture dynamically visits high-likelihood "surface" cells of a multi-dimensional parameter grid, using a surface-tracking and priority queue approach to achieve nearly perfect parallel scaling while avoiding low-likelihood computation (Mikkelsen et al., 2012). Each "slave" agent processes a unique parameter cell, directly returning likelihoods.
• Markov Chain Monte Carlo (MCMC) Integration: MAS can embed or orchestrate traditional MCMC engines (e.g., Metropolis-Hastings), either directly (as in cmbagent, using cobaya and getdist) or as a post-processing refinement over a region identified by other agents (Laverick et al., 2024, Hernández et al., 7 Aug 2025). This allows agents to efficiently combine global optimization (PSO or grid) with Boltzmann-sampling-based posterior mapping.
• ML Regression/Emulation: In certain agentic systems, cosmological parameter inference is formulated as a supervised regression, with neural networks $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$2 trained to predict $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$3 given simulation input, optimized via MSE losses approximating maximum-likelihood estimation (Moss, 4 Apr 2025).

Significantly, several works recommend hybrid approaches: PSO or grid-based search quickly locates maxima or high-probability regions, which then serve as starting points or proposal regions for a sampling-based method (e.g., MCMC), balancing computational efficiency and statistical rigor.

3. Data Structures, Cosmological Models, and Parameterization

MAS-CPA frameworks are designed to be model-agnostic, supporting $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$4CDM and its extensions (curvature, dark energy, early-time modifications) as well as simulation parameter extraction and ensemble analysis:

• Parameter Spaces: Agents explore parameter sets such as $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$5 for flat $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$6CDM; $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$7 for curved models; and extras $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$8 for CPL dark energy models (Hernández et al., 7 Aug 2025).
• Boundaries and Priors: Priors are typically uniform over physically plausible intervals; e.g., $$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$9, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$0, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$1 (Hernández et al., 7 Aug 2025, Prasad, 2014, Prasad et al., 2011).
• Likelihood Functions: Objective functions reduce to minimizing

$x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$2

where $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$3 is the observed data vector (e.g., SNIa, BAO, CMB), $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$4 is the theoretical prediction, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$5 is the covariance (Hernández et al., 7 Aug 2025, Prasad, 2014).

• Simulation Parameter Extraction: Architecture agent chains parse publication text for $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$6, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$7, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$8, $x_i^{(t+1)} = x_i^{(t)} + v_i^{(t+1)}$9, box sizes, and instantiate code-compliant configuration files according to simulation (e.g., Gadget/MP-Gadget) documentation, check internal consistency (e.g., flatness: $w$0), and resolve units (Zhang et al., 11 Jul 2025).
• Ensemble and Large-Scale Data Handling: For TB-scale simulation sets, agent-based data loaders incrementally subset data into queryable DuckDB or similar, while SQL and Python agents sequentially derive summary statistics (e.g., mass functions, power spectra), always with downstream step-wise validation (Tam et al., 14 Oct 2025).

4. Performance, Benchmarking, and Comparative Evaluation

MAS-CPA implementations consistently demonstrate significant speed-ups, robust exploration, and, under hybrid approaches, statistical completeness comparable to conventional inference methods:

Method	Best-fit Accuracy	Posterior Recovery	Computational Cost
PSO (100 particles, 5D) (Hernández et al., 7 Aug 2025)	$w$1 agreement with MCMC	Fisher error envelope	$w$21.5×10⁴ likelihoods, %%%%33$$v_i^{(t+1)} = w v_i^{(t)} + c_1 r_1 (p_i - x_i^{(t)}) + c_2 r_2 (g - x_i^{(t)}),$$034%%%% faster than MCMC
Snake (12D) (Mikkelsen et al., 2012)	$w$5 mean/σ bias	Full grid, exact $w$6	$w$7 likelihoods, $w$8linear scaling, near-perfect parallel efficiency
SimAgent (parameter extraction) (Zhang et al., 11 Jul 2025)	98.7% micro-F1	Not sampling, exact mapping	$w$92 min/paper, $c_1$0 human time
InferA (terabyte ensemble) (Tam et al., 14 Oct 2025)	76% scientifically valid outputs	MCMC posteriors via agent	Up to 11.2 TB analyzed, sublinear overhead; LLM + code execution
cmbagent (LLM-driven MCMC) (Laverick et al., 2024)	Posterior matches published ACT DR6	Full MCMC posteriors	$c_1$140 min full pipeline, $c_1$2 min MCMC, $c_1$3 typical human

PSO-based and agentic approaches both converge to published parameter constraints (ACT, Planck, SNIa, BAO) with errors or deviations $c_1$4 compared to MCMC chains. Agentic parameter extraction systems surpass single-agent or chain-of-thought baselines by $c_1$54\% F1, with $c_1$6 errors per simulation (Zhang et al., 11 Jul 2025). In TB-scale settings, only multi-agent orchestration achieves feasible interactive analysis (Tam et al., 14 Oct 2025).

5. Extensions, Generalization, and Future Prospects

MAS-CPA architectures are evolving to cover an increasingly broad analytic spectrum:

• Beyond Metropolis Algorithms: Integration of Hamiltonian Monte Carlo, nested sampling, and importance-sampling schemes is straightforward by agentic design (e.g., cobaya "sampler" block selection) (Laverick et al., 2024).
• Automated ML-powered Inference: Agentic systems such as AI Cosmologist automatically design, test, and refine machine-learning pipelines for parameter regression on simulation data (e.g., Quijote), achieving validation RMSE competitive with, or surpassing, previous human-run pipelines (Moss, 4 Apr 2025).
• Automated Provenance and Reproducibility: Documentation agents maintain complete records of code, data subsets, SQL logs, and visualization specs, ensuring all MAS-CPA runs are reproducible and auditable (Tam et al., 14 Oct 2025).
• Domain Adaptability: MAS-CPA frameworks are agnostic to model (flat/cross $c_1$7CDM, CPL, extensions), data source (observational or simulation), and scale (small–TB). Extensions include plug-and-play data loaders, agentic installers, or Web scraping agents to ingest new measurements and simulation codes (Zhang et al., 11 Jul 2025, Laverick et al., 2024, Tam et al., 14 Oct 2025).
• Limitations: Current MAS-CPA systems still require human-in-the-loop oversight at key ambiguity or QA failure points—particularly in deep scientific reasoning, choice of statistical forms, or unit system translation. Automated end-to-end runs in production-scale HPC environments remain an active direction (Tam et al., 14 Oct 2025, Laverick et al., 2024).

6. Impact and Outlook in Cosmological Science

MAS-CPA frameworks markedly reduce human labor and iteration cycles: PSO accelerates maximization by an order of magnitude over traditional MCMC; LLM-driven agents collapse development, documentation location, and error-handling into minutes; agentic ML platforms explore dozens of hypotheses or architectures in hours (Hernández et al., 7 Aug 2025, Laverick et al., 2024, Moss, 4 Apr 2025). These capabilities are enabling a transition from manual, bespoke analyses to scalable, auditable, and increasingly autonomous inference pipelines, reflecting the practical necessity of such systems for next-generation cosmological surveys and simulations.

By integrating multi-agent, optimization, and data-driven reasoning—as well as automated documentation and modular extension—MAS-CPA has become a core methodology in contemporary cosmological parameter analysis, maintaining statistical rigor while meeting the computational and logistical challenges inherent to modern datasets and models.