Heterogeneous Scientific Foundation Model Collaboration

Abstract: Agentic LLM systems have demonstrated strong capabilities. However, their reliance on language as the universal interface fundamentally limits their applicability to many real-world problems, especially in scientific domains where domain-specific foundation models have been developed to address specialized tasks beyond natural language. In this work, we introduce Eywa, a heterogeneous agentic framework designed to extend language-centric systems to a broader class of scientific foundation models. The key idea of Eywa is to augment domain-specific foundation models with a language-model-based reasoning interface, enabling LLMs to guide inference over non-linguistic data modalities. This design allows predictive foundation models, which are typically optimized for specialized data and tasks, to participate in higher-level reasoning and decision-making processes within agentic systems. Eywa can serve as a drop-in replacement for a single-agent pipeline (EywaAgent) or be integrated into existing multi-agent systems by replacing traditional agents with specialized agents (EywaMAS). We further investigate a planning-based orchestration framework in which a planner dynamically coordinates traditional agents and Eywa agents to solve complex tasks across heterogeneous data modalities (EywaOrchestra). We evaluate Eywa across a diverse set of scientific domains spanning physical, life, and social sciences. Experimental results demonstrate that Eywa improves performance on tasks involving structured and domain-specific data, while reducing reliance on language-based reasoning through effective collaboration with specialized foundation models.

• The paper introduces the Eywa framework that enables direct modality-native collaboration between language agents and domain-specific models, reducing information loss.
• The paper demonstrates that adaptive orchestration and plug-and-play multi-agent systems yield lower population risk and higher utility across diverse scientific tasks.
• The experimental evaluation confirms efficiency improvements with reduced token usage and latency across physical, life, and social science domains.

Heterogeneous Scientific Foundation Model Collaboration: An Authoritative Evaluation

Problem Motivation and Conceptual Framework

Agentic AI systems based on LLMs exhibit substantial generic reasoning and decision-making capabilities, but their exclusive reliance on natural language presents a fundamental bottleneck in scientific domains. Real-world scientific tasks frequently require reasoning over domain-specific inputs like time series, tabular datasets, molecular graphs, or structured data, which cannot be faithfully or efficiently encoded as text without information loss. This intrinsic limitation, rigorously motivated using information-theoretic arguments in the paper, implies that pure language-based interfaces are suboptimal for tasks demanding specialized computation. Meanwhile, scientific foundation models—pretrained over domain-specific data and optimized for modalities such as forecasting, classification, or sequence analysis—lack native linguistic interfaces but demonstrate strong predictive power. The core premise of the paper is to bridge this gap and permit direct, modality-native collaboration between heterogeneous foundation models and language-based agents.

Eywa introduces a modular framework encapsulating this paradigm shift. The FM-LLM Tsaheylu interface acts as a bidirectional communication and control layer, allowing LLMs to parse instructions, configure calls, and interpret outputs from foundation models. This framework is instantiated in three progressively complex architectures: EywaAgent (coupled FM-LLM unit for adaptive reasoning), EywaMAS (heterogeneous multi-agent system), and EywaOrchestra (dynamic orchestration via a central planner).

Theoretical Guarantees

A rigorous theoretical backbone supports the Eywa framework. The information-theoretic analysis demonstrates that any serialization of domain-specific inputs to text necessarily discards information, creating a Bayes risk gap that LLM-only agents cannot overcome. EywaAgent, by directly invoking the native modality computations of foundation models, strictly expands the function class and achieves strictly lower population risk on scientific tasks under reasonable assumptions (see Theorem 15). The containment property (Proposition 14) confirms that Eywa subsumes language-only agents as a strict special case, and the extension to multi-agent settings (Theorem 17) guarantees strict improvement of EywaMAS over pure LLM-based systems, contingent on communication topology. EywaOrchestra additionally leverages adaptivity: dynamic routing and configuration selection yield population risk strictly lower than any fixed system by exploiting task heterogeneity (Theorem 18).

System Design and Implementation

The FM-LLM Tsaheylu interface is operationalized via Model Context Protocol (MCP) and implemented as a structured tool-call pattern in modern agentic platforms (e.g., LangChain agents and FastMCP servers). EywaAgent dynamically switches between generalized language reasoning and domain-specific acting based on a control policy $C: S \to \{\text{invoke}, \text{skip}\}$, where skip reduces to pure language reasoning and invoke delegates computation to the appropriate foundation model.

EywaMAS accommodates plug-and-play composition with minimal architectural changes, permitting seamless replacement of language-only agents with EywaAgents in existing multi-agent workflows. EywaOrchestra introduces dynamic orchestration: a conductor (typically instantiated as an LLM) analyzes task modality and complexity, selects appropriate agent configurations, backbone models, foundation models, and communication topologies, and then executes the resultant system. The orchestration logic is highly modular, with configuration expressed as structured JSON objects and hard constraints to ensure robust, auditable execution.

Experimental Evaluation

EywaBench, a comprehensive multi-domain evaluation platform, is constructed to systematically assess the performance of heterogeneous agentic systems. EywaBench covers physical (material, energy, space), life (biology, clinic, drug), and social (economy, business, infrastructure) science domains, with balanced representation of modalities (natural language, time series, tabular). Evaluation metrics are modality-agnostic and normalized to facilitate rigorous comparison.

Numerical results are conclusive:

• Utility: EywaAgent achieves a mean utility gain of ~$6.6\%$ over backbone-matched single-LLM agents, EywaMAS outperforms all homogeneous multi-agent baselines (Refine, Debate, MoA, X-MAS) by up to $7.5\%$ with significantly less token consumption.
• Efficiency: Token usage is reduced by ~30%, and inference latency drops by ~10% due to streamlined computation delegation. These results are robust across all domains and agentic settings.
• Heterogeneity criticality: Heterogeneous LLM-only MAS methods (MoA, X-MAS) do not consistently surpass homogeneous baselines, confirming that cross-modality heterogeneity (not merely model heterogeneity) is central for scientific tasks.
• Adaptive orchestration: EywaOrchestra achieves near-optimal utility at reduced cost and with zero expert configuration, demonstrating strong automation and dynamic adaptability. In several domains, EywaOrchestra surpasses expert-designed EywaMAS.

Eywa exhibits strong resilience to hyperparameter variations (LLM sampling temperature, FM temperature, prompt design) and backbone scaling, with diminishing returns beyond a certain LLM capability threshold—a finding that substantiates the value of domain-specific modeling over brute-force scaling of LLMs.

Qualitative Case Studies

Case studies empirically validate the theoretical claims:

• LLM-only agents: Format adherence does not imply genuine domain computation; predictions on structured signals default to naive heuristics (e.g., persistence).
• EywaAgent: Enables seamless integration—LLM parses task and configures FD call, FD delivers predictions, LLM verifies and formats output.
• EywaOrchestra: Diagnoses task modality, selects minimal required agent system, avoids unnecessary debate or redundancy, maximizes efficiency.

The division of labor and adaptive orchestration are clearly executed, achieving optimal outcomes with minimal token usage and latency.

Implications and Future Directions

Eywa fundamentally advances the integration of domain-specialized foundation models with linguistic reasoning agents, enabling rigorous, scalable, and efficient scientific computation. The theoretical guarantees and empirical results highlight modality-native collaboration as necessary for agentic scientific workflows, especially as scientific tasks and expert models scale in diversity and complexity.

Practically, Eywa offers a blueprint for modular, extensible agentic ecosystems spanning multiple scientific domains. The framework’s plug-and-play architecture, standardization of computation interfaces, and orchestration policies form the foundation for scalable heterogeneous scientific model platforms. Future directions include expanding the pool of domain-specific models, developing advanced orchestration policies via supervised or reinforcement learning, tighter coupling (shared representation, differentiable interfaces), and expanding EywaBench for more comprehensive coverage and capability benchmarking.

Limitations

Eywa’s effectiveness hinges on the quality of both LLMs and scientific foundation models; poor or misaligned models limit achievable gains. While EywaBench is diverse, exhaustive domain coverage is unattainable; systematic expansion and domain-specific evaluation remain open problems. Computational cost, though mitigated via orchestration, may rise with ecosystem size and heterogeneity, necessitating more efficient agent selection and invocation strategies.

Conclusion

Eywa offers a rigorously justified, empirically validated framework for heterogeneous agentic collaboration in scientific domains, substantially improving utility and efficiency, and showcasing modality-native computation as a necessary principle in AI for science. By enabling adaptive orchestration and flexible integration of domain-specific foundation models, Eywa transcends the limitations of language-centric systems and paves the way for scalable, ecosystem-level scientific AI (2604.27351).