Heterogeneous Scientific Foundation Model Collaboration

Zihao Li, Jiaru Zou, Feihao Fang, Xuying Ning, Mengting Ai, Tianxin Wei, Sirui Chen, Xiyuan Yang, Jingrui HearXiv 2026 · arXiv

TL;DR

Current large language model (LLM) agents struggle with scientific tasks because forcing non-linguistic data (like time-series or tables) into text creates a severe information bottleneck. This paper introduces Eywa, a framework that connects LLMs with domain-specific foundation models via a standardized interface, allowing them to collaborate natively without losing data fidelity. Across a new multi-domain scientific benchmark, Eywa significantly improves task accuracy while slashing token consumption by roughly 30%.

What Problem Does This Paper Address?

While LLM-based agents excel at general reasoning and planning, natural language is a poor universal interface for the sciences. Scientific workflows depend heavily on structured, non-linguistic data modalities, such as time-series logs, tabular measurements, and molecular structures.

When researchers try to use standard LLMs for these tasks, they typically serialize the data—converting matrices or sequences into long strings of text. The authors show that this text-filtering process acts as an information bottleneck, discarding critical structural patterns and precise numerical relationships. Consequently, LLMs often fall back on shallow heuristics (like simply repeating the last observed value in a time series), while consuming massive amounts of tokens and execution time.

Conversely, domain-specific foundation models (FMs)—like specialized weather or time-series predictors—are excellent at processing native scientific data but lack general reasoning, planning, and natural language interfaces. This paper tackles the challenge of getting these two highly distinct model types to collaborate effectively.

Methods

The authors propose Eywa, a heterogeneous agentic framework inspired by the neural bond (Tsaheylu) from the movie Avatar, which allows different species to communicate directly without a shared spoken language.

Analogy between the Avatar Pandora ecosystem and the Agentic AI ecosystem.

Eywa is instantiated in three distinct ways:

  1. EywaAgent (Single-Agent Integration): Pairs an LLM with a domain-specific FM using a bidirectional communication interface. A query compiler ($\phi$) translates the LLM's state into a structured tool call for the FM, and a response adapter ($\psi$) converts the FM's raw output back into a language-compatible format. This is implemented using the Model Context Protocol (MCP) to handle remote or local server-side computation seamlessly.
  2. EywaMAS (Multi-Agent System): A plug-and-play extension where EywaAgents replace traditional, language-only agents within existing multi-agent architectures (like debate or hierarchical setups) without needing to redesign communication protocols.
  3. EywaOrchestra (Dynamic Orchestration): A framework featuring a centralized LLM "conductor." Instead of using a fixed multi-agent setup, the conductor dynamically analyzes the incoming scientific task and selects the optimal configuration: the role of each agent, the backbone LLM, the attached domain-specific FM, and the communication topology.

Datasets

To evaluate the framework, the authors introduced EywaBench, a scalable, multi-task, multi-domain scientific benchmark. It spans three main branches of science, each broken down into three sub-domains:

  • Physical Science: Material, Energy, Space
  • Life Science: Biology, Clinic, Drug
  • Social Science: Economy, Business, Infrastructure

EywaBench-V1 aggregates 200 task instances sampled from four primary source datasets: DeepPrinciple (scientific QA), MMLU-Pro (multi-task language understanding, adapted to open-ended formats), fev-bench (time-series forecasting), and TabArena (tabular classification and regression).

Key Insights and Ideas

  • The Irreducible Bayes Risk Gap: The authors provide an information-theoretic proof showing that text serialization inherently discards task-relevant information. No amount of downstream prompt engineering or multi-agent debate can recover this lost data, meaning text-only LLMs face a hard mathematical ceiling on structured data tasks.
  • True Division of Labor: Eywa succeeds by letting each model play to its strengths. The LLM acts as the "brain"—handling instruction parsing, step-by-step planning, and final answer formatting—while the domain FM acts as the "calculator," executing heavy numerical and modality-native predictions.
  • More Agents $\neq$ Better Performance: In domains like economy and business, a single specialized EywaAgent often beats complex multi-agent LLM setups. This underscores the value of EywaOrchestra, which avoids multi-agent communication overhead by routing simpler tasks directly to a single specialized expert.

Key Results

The framework was evaluated using various LLM backbones alongside specialized FMs like Chronos (for time-series) and TabPFN (for tabular data).

Overall utility and token consumption of different methods.
  • Better Accuracy, Lower Cost: Compared to a single text-only LLM agent, EywaAgent improved average utility scores by ~7% across physical, life, and social sciences, while cutting token consumption by ~30% and execution time by ~10%.
  • Outperforming Multi-Agent Baselines: EywaMAS consistently beat state-of-the-art multi-agent frameworks (such as Mixture-of-Agents and X-MAS) in utility while operating at a fraction of their token cost.
  • Automated Optimization: EywaOrchestra successfully matched or exceeded expert-designed static topologies without requiring human tuning, positioning itself firmly on the Pareto frontier of the utility-vs-cost trade-off.

Limitations and Future Work

  • Dependence on Base Models: Eywa's performance is ultimately bounded by the planning capabilities of the backbone LLM and the accuracy of the underlying domain FMs.
  • Modality Expansion: EywaBench-V1 is currently focused on time-series and tabular data. The authors note that future iterations must incorporate other critical scientific modalities, such as molecular graphs, 3D protein structures, and geospatial imagery.
  • Tighter Coupling: While the current framework relies on structured text/JSON interfaces via MCP, the authors suggest exploring deeper integrations, such as shared, differentiable representation spaces and bidirectional alignment between language and scientific embeddings.

Summary updated

Abstract

Agentic large language model 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 language models 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.
BibTeX
@misc{li2026heterogeneous,
  author = {Zihao Li and Jiaru Zou and Feihao Fang and Xuying Ning and Mengting Ai and Tianxin Wei and Sirui Chen and Xiyuan Yang and Jingrui He},
  title = {{Heterogeneous Scientific Foundation Model Collaboration}},
  year = {2026},
  eprint = {2604.27351},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.27351},
}
Figure