Skip to content

Heterogeneous Scientific Foundation Model Collaboration

🕒 Published (v1): 2026-04-30 03:02 UTC · Source: Arxiv · link

Why this paper was selected

Heterogeneous scientific foundation model collaboration beyond text-only agent interfaces

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Eywa is a three-tier agentic framework that bridges language models and domain-specific scientific foundation models (FMs) that lack native language interfaces, enabling modality-native collaboration for scientific tasks. By wrapping FMs with LLM-based reasoning interfaces via the Model Context Protocol (MCP), Eywa improves scientific task utility while reducing token consumption compared to language-only agent baselines.

Problem

LLM-based agentic systems treat natural language as the universal interface, but scientific foundation models for structured data (time series, tabular) operate over non-linguistic modalities and cannot directly participate in language-centric reasoning pipelines. Existing multi-agent frameworks assume all agents communicate through text, excluding specialized FMs that deliver provably better domain performance but have no language I/O.

Method

Eywa formalizes a three-stage hierarchy:

  1. EywaAgent — a coupled FM–LLM unit \(A_\text{eywa} = (A_\text{LLM}, F, \varphi, \psi, C)\). A query compiler \(\varphi: S \to U\) translates LLM state into structured FM invocations; a response adapter \(\psi: O \to Z\) converts FM outputs back to language-compatible representations. A control policy \(C: S \to \{\text{invoke}, \text{skip}\}\) decides at each step whether to delegate to the FM or fall back to language-only reasoning. This "Tsaheylu" interface is implemented over MCP, exposing each FM as a remote service with a typed schema.

  2. EywaMAS — a plug-and-play multi-agent system \(M_\text{Eywa} = (\mathcal{A}, G)\) where any language agent in an existing topology \(G\) (sequential, hierarchical, looped) can be replaced by an EywaAgent without changing the communication protocol.

  3. EywaOrchestra — a dynamic orchestration layer \(O = (\mathcal{C}, P)\) where a conductor \(P\) (implemented as an LLM) maps each incoming task to a system configuration from a topology pool \(\Pi\), selecting agent types, backbone LMs, attached FMs, and communication topology per-task. This achieves adaptive risk \(R_\text{oracle} = \mathbb{E}_\tau[\min_{c \in \mathcal{C}} \mathbb{E}[\ell(F_c(q,x), y^\star)]] \leq R^\star_\text{fixed}\), strictly improving over any fixed configuration when different task regions favor different setups.

Baseline FMs used: Chronos (time-series) and TabPFN (tabular), neither with native language I/O.

Key Contributions

  • Eywa framework: heterogeneous agentic framework enabling modality-native FM collaboration via FM–LLM "Tsaheylu" interfaces built on MCP.
  • Three instantiations: EywaAgent (single-agent), EywaMAS (multi-agent drop-in replacement), EywaOrchestra (conductor-driven dynamic orchestration).
  • Theoretical guarantees: Theorem 3 proves EywaAgent achieves strictly lower optimal expected task loss than any language-only agent class under the domain advantage assumption (Assumption 1); analogous result for EywaMAS.
  • EywaBench: a new scalable multi-task, multi-domain benchmark covering physical, life, and social science across natural language, time-series, and tabular modalities (9 sub-domains).

Results

  • EywaAgent vs. Single-LLM-Agent: ~7% utility improvement, ~30% token reduction, ~10% execution time reduction across physical/life/social science tasks.
  • EywaMAS: consistent gains in utility, token efficiency, and execution time in multi-agent settings over Multi-LLM-Agents baseline.
  • EywaOrchestra: outperforms both single-agent and multi-agent language-only baselines.
  • Domain-level utility improvements (figure-reported ranges across Eywa variants):
  • Biology: 30–40%
  • Material: 45–65%
  • Drug: 55–65%
  • Infrastructure: 60–75%
  • Economy: 70–80%
  • Business: 45–80%
  • Clinic: 80–90%

Limitations

  • The topology pool in EywaOrchestra is finite and manually predefined; the conductor selects from it rather than constructing arbitrary topologies.
  • The conductor is itself an LLM, so misconfiguration of agent roles or FM assignments is possible and not formally bounded.
  • Evaluation relies on two FMs only (Chronos, TabPFN); generalization to vision, geospatial, or symbolic modalities is asserted but not demonstrated.
  • EywaBench coverage is currently limited to time-series and tabular; the claim of extensibility to other modalities is aspirational.
  • The paper is truncated before the full results table (Table 1), so quantitative ablations against intermediate configurations are not available from the provided text.

Relevance to Harnesses / Meta-Harnesses

EywaOrchestra is a textbook meta-harness: a conductor agent dynamically instantiates and configures a multi-agent pipeline per task, selecting agent types, model backends, and communication topology from a pool — precisely the role a meta-harness plays in programmatic orchestration systems. The MCP-based Tsaheylu interface formalizes the "tool wrapper" pattern central to harness design, providing a typed, schema-validated bridge between a reasoning controller and heterogeneous executors. The three-tier architecture (EywaAgent → EywaMAS → EywaOrchestra) directly mirrors the single-agent → fixed-pipeline → adaptive-orchestrator progression seen in harness meta-design literature, and the theoretical framing of adaptive vs. fixed-configuration risk quantifies the benefit a meta-harness gains over a static pipeline. For researchers building scientific harnesses that must route tasks to specialized model backends, Eywa's MCP-based conductor pattern offers a concrete, deployable template.