PaperlayerView exact revision on arXiv

Source-linked AI summary

Heterogeneous Scientific Foundation Model Collaboration

Zihao Li, Jiaru Zou, Feihao Fang, Xuying Ning, Mengting Ai, Tianxin Wei, Sirui Chen, Xiyuan Yang, Jingrui He

arXiv:2604.27351v1cs.AIcs.CLcs.LG

TL;DR

Language-centric agents are limited on scientific tasks involving structured, non-linguistic data. Eywa connects language reasoning with specialized foundation models through modality-native interfaces, improving utility while reducing token usage and inference cost across scientific domains.

  • Problem

    Language-centric agentic systems lack effective interfaces for scientific tasks involving specialized, structured, and non-linguistic data.

  • Method

    Eywa augments domain-specific foundation models with language-based reasoning interfaces for modality-native collaboration in agentic systems.

  • Results

    EywaAgent improves utility by ∼7% across physical, life, and social science tasks, while reducing token usage by ∼30% and execution time by ∼10%.

  • Takeaways & Limitations

    Modality-native collaboration improves scientific task solving while reducing token usage and inference cost.

  • Takeaways & Limitations

    EywaMAS may suffer from suboptimal communication topologies or inefficient human configuration because its coordination is fixed and decentralized.

Abstract

from arXiv · show

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.

1. Introduction

Eywa addresses the communication gap between language-centric agentic systems and domain-specific foundation models by adding language-model reasoning interfaces for specialized data and tasks. The framework includes single-agent, multi-agent, and planning-based orchestration variants, and experiments across physical, life, and social sciences show improved utility with lower resource use.

  • Motivation: Language-centric agents struggle with scientific problems involving specialized data because many domain-specific foundation models lack native language input and output modalities.This mismatch makes direct incorporation of specialized foundation models into agentic systems challenging.
  • Framework: EywaAgent provides single-agent integration, EywaMAS extends the approach to multi-agent systems, and EywaOrchestra dynamically coordinates agents through planning-based orchestration.These are the three instantiations introduced in the framework.
  • Results: Experiments on the EywaBench benchmark demonstrate improved performance from integrating domain-specific foundation models into agentic systems across diverse scientific domains.EywaBench covers tasks spanning the physical, life, and social sciences.
  • Framework: Eywa augments domain-specific foundation models with language-model-based reasoning interfaces so language agents can guide inference, planning, and decision-making over specialized tasks.The framework enables modality-native collaboration between heterogeneous foundation models and language agents.
  • Results: Across physical, life, and social science tasks, EywaAgent improves utility by ∼7% over the Single-LLM-Agent baseline while reducing token usage by ∼30% and execution time by ∼10%.The results are reported as improvements in the utility-cost trade-off over language-only baselines.

2. Preliminary

This section formalizes LLM agents, domain-specific foundation models, and multi-agent systems, then frames scientific tasks as combining language-observable and domain-specific inputs. It also states that foundation models have a domain advantage over language-only models when specialized inputs are informative.

  • LLM Agent: LLM agents are modeled as policies that map internal states to distributions over messages, actions, or tool invocations.They are assumed to have strong general-purpose reasoning while accessing non-linguistic inputs indirectly through textualized representations.
  • Domain-Specific Foundation Model: Domain-specific foundation models map specialized inputs and structured user configurations to outputs without necessarily providing a native language interface.For domain k, the model is formulated as F_k : X_k × U_k → O_k.
  • Multi-Agent Systems: A multi-agent system consists of a set of agents and a communication topology, with agents iteratively updating local states and producing messages until a final output is produced.The supplied formulation defines the system as M = (A, G) and describes messages exchanged under topology G.
  • Problem Formulation: Scientific task inputs are factorized as X = X_lng × X_1 × · · · × X_m, separating language-observable context from domain-specific inputs.Tasks are represented by τ = (q, x, y⋆, ℓ), and the agentic system objective is to minimize expected task loss.
  • Domain Advantage of Foundation Models: When a domain-specific input is informative, its foundation model is assumed to outperform any language-only model on that component.The assumption compares F_k with language-only models using the sub-task loss ℓ_k after projection onto the domain-specific component.

3. EywaAgent: Reasoning Foundation Model Agents

EywaAgent couples a language model’s high-level reasoning with a domain-specific foundation model through the bidirectional Tsaheylu interface. Its adaptive invoke/skip control preserves language-only reasoning while enabling specialized computation and, under a domain advantage assumption, strictly lowers optimal risk.

  • EywaAgent abstraction: EywaAgent augments a domain-specific foundation model with a language-based reasoning interface, coupling high-level planning and control with specialized capabilities.The framework defines the agent as a coupled FM-LLM unit rather than a single standalone model.
  • Tsaheylu interface: The Tsaheylu interface uses a query compiler to translate task states into structured foundation-model invocations and a response adapter to return planner-consumable context.This bidirectional interface enables heterogeneous foundation models to participate in the language-centric reasoning loop.
  • Adaptive control: EywaAgent dynamically chooses invoke or skip at each reasoning step, switching between specialized foundation-model computation and standard language-only reasoning.The control policy determines whether and how the foundation model is invoked; otherwise, it can be induced by language-model reasoning over the current state.
  • Expressivity: EywaAgent subsumes language-only agents by always skipping foundation-model calls while expanding the computable-function space through domain-specific foundation models.The adaptive mechanism is described as enabling both enhanced expressivity and improved task performance.
  • Theoretical result: Under Assumption 1, Theorem 3 establishes a strict risk improvement for EywaAgent over language-only agents.The result attributes the improvement to preserving language-only reasoning through the skip branch and adding foundation-model computation through the invoke branch under the domain advantage assumption.

4. Eywa Agentic Systems: Multi-Agent Composition and Orchestration

Eywa extends agentic systems from single specialized agents to heterogeneous multi-agent collaboration through EywaMAS and task-adaptive orchestration through EywaOrchestra. EywaMAS integrates language-only agents with EywaAgents, while EywaOrchestra dynamically selects agents, models, foundation models, and communication topologies for each task.

  • EywaMAS: EywaMAS generalizes EywaAgent into a distributed system where heterogeneous agents interact and collaborate through a unified multi-agent framework.Its agents may be LLM agents or EywaAgents, and communication is governed by a specified topology.
  • EywaMAS: EywaMAS enables plug-and-play replacement of selected language-only agents with EywaAgents while preserving existing multi-agent architectures.This integrates language-based reasoning with specialized foundation-model capabilities and requires minimal architectural modification.
  • EywaMAS: EywaMAS supports parallel specialization and cross-domain collaboration, but fixed topology and decentralized coordination can cause suboptimal organization and configuration inefficiency.These limitations motivate automatic orchestration.
  • EywaOrchestra: EywaOrchestra dynamically instantiates heterogeneous multi-agent systems because task-specific mixtures of language reasoning, domain prediction, and collaboration patterns can differ.It addresses the limitations of static agents and fixed multi-agent systems.
  • EywaOrchestra: A conductor plans each EywaOrchestra system by selecting agent roles and types, language-model backbones, attached foundation models, and communication topology.These choices are conditioned on the input task within a configuration space defined by candidate language models, foundation models, and topology options.

5. Experiments

Experiments evaluate Eywa on EywaBench, a unified multi-task and multi-domain benchmark spanning natural language, time series, and tabular data across physical, life, and social sciences. EywaAgent improves average utility and efficiency over corresponding single-agent baselines, while robustness studies show effectiveness across hyperparameters and language-model backbones.

  • Benchmark: Eywabench covers natural-language, time-series, and tabular tasks across physical, life, and social science domains.Its sub-domains include material, energy, space, biology, clinic, drug, economy, business, and infrastructure.
  • Benchmark: Eywabench uses a unified utility score u ∈[0, 1] to compare results across modalities, reporting per-domain and overall mean utility.Natural-language utility uses soft matching, while time-series and tabular utility derives from normalized prediction errors.
  • Overall results: EywaAgent increases average utility by 6.6% and cuts token usage by nearly 30% versus the corresponding single-agent baseline.The reported comparison also states that EywaAgent reduces latency through delegation to domain-specific foundation models.
  • Robustness: Eywa remains stable across varying LLM sampling temperatures, TabPFN softmax temperatures, and prompt designs.The results state that gains are not tied to a particular hyperparameter choice or prompt template, with structured prompts generally beneficial.
  • Robustness: EywaAgent remains effective across physical, life, and social science domains with gpt-4.1-nano, gpt-5-nano, and gpt-5-mini backends.Overall utility increases as backend capability increases, indicating robustness to backbone selection while benefiting from stronger language models.

6. Related Work

Related work spans scientific language models, domain-specific foundation models, and agentic systems for scientific workflows. Eywa addresses the limitation of language-centric scientific LLMs by using LLMs as reasoning interfaces that coordinate models operating on native scientific representations.

  • Scientific Large Language Models and Foundation Models: Scientific LLMs support scientific knowledge understanding, quantitative reasoning, and domain-specific question answering through general pretraining, specialized adaptation, and agentic workflows.These paradigms include large-scale scientific corpora, discipline-specific models in biomedicine, medicine, and chemistry, and workflows for multi-step reasoning.
  • Scientific Large Language Models and Foundation Models: Scientific LLMs remain language-centric, representing scientific data as descriptions or symbolic sequences, which is suboptimal for structured information.Eywa instead uses LLMs as reasoning interfaces for coordinating domain-specific models operating on native scientific representations.
  • Scientific Large Language Models and Foundation Models: Domain-specific foundation models have advanced across scientific disciplines, including time-series models enabling competitive zero-shot forecasting and tabular models using in-context learning for prediction.The passage also notes that TabPFN and successors outperform tuned tree-based ensembles.
  • Agentic Systems in Scientific Settings: Scientific agentic systems have been applied to hypothesis generation, literature synthesis, experimental design, and automation of the AI research process.Existing workflows often use fixed agent topologies, while another approach exposes simulators, solvers, or expert systems through external APIs.

7. Conclusion … A.1. Notation, Problem Setup, and Assumptions

Eywa connects language reasoning with domain-specific foundation models through the FM–LLM “Tsaheylu” interface, supporting agentic systems across heterogeneous scientific data. Its theory formalizes expressivity, orchestration, task structure, domain advantage, information loss from serialization, and interface fidelity.

  • 7. Conclusion: Eywa connects language-centric reasoning with domain-specific foundation models through the FM–LLM “Tsaheylu” interface, instantiated as EywaAgent, EywaMAS, and EywaOrchestra.The framework targets scientific tasks involving structured, non-linguistic data and is evaluated across physical, life, and social sciences.
  • 7. Conclusion: Eywa improves utility while reducing token usage and inference cost through modality-native collaboration across physical, life, and social science tasks.
  • Appendix: The appendix develops Eywa’s theoretical analysis, including the language interface bottleneck, expressivity and solvability guarantees, multi-agent generalization, adaptive orchestration, and token complexity.Appendix B additionally discusses related work, scope, limitations, and potential extensions.
  • A.1. Notation, Problem Setup, and Assumptions: The notation defines function classes for language-only systems, Eywa agents, multi-agent systems, and orchestration over candidate LLMs, foundation models, topologies, and interface pairs.A function class is the set of input-output mappings implementable by a specified system.
  • A.1. Notation, Problem Setup, and Assumptions: The induced classes satisfy FLLM ⊆ FEywa ⊆ FEywa-MAS ⊆ FOrch because each richer system can reproduce the preceding system through restricted invocation, topology, or conductor choices.
  • A.1. Notation, Problem Setup, and Assumptions: Scientific-task inputs factorize as X = Xlng × X1 × · · · × Xm, separating language-observable context from domain-specific components.The task loss is assumed compatible with this factorization through component-wise losses and a coordinate-wise nondecreasing aggregation function; improving one component strictly improves overall loss when others are unchanged.
  • A.1. Notation, Problem Setup, and Assumptions: The domain-advantage assumption posits task families where a domain component is sufficient and its foundation model solves it perfectly, while language-only agents incur strictly positive loss.This assumption is stated as an extension of the main paper’s Assumption 1.
  • A.1. Notation, Problem Setup, and Assumptions: Heterogeneity yields strict gains when serialization discards task-relevant information, specifically when E[Y | Xk] ≠ E[Y | Tk(Xk)] on a subset of positive probability.The performance-preserving interface assumption further requires foundation-model calls to avoid degrading FM performance and outperform language-only agents, recovering the Bayes-optimal conditional expectation given domain-specific input.

A.2. Information-Theoretic Analysis of the Language Interface Bottleneck … A.5. Adaptive Orchestration: EywaOrchestra

The analysis shows that language serialization creates an information and risk bottleneck that language-only agents cannot overcome, while Eywa’s direct access to specialized foundation models expands expressivity and solvable tasks. These benefits extend through multi-agent communication, and adaptive orchestration strictly improves over fixed configurations when task regions favor different systems.

  • A.2. Information-Theoretic Analysis of the Language Interface Bottleneck: Serialization cannot increase task-relevant information: I(Y; T(X)) ≤ I(Y; X), with equality only when I(Y; X | T(X)) = 0.The result follows from the Markov chain Y → X → T(X) and the data processing inequality.
  • A.2. Information-Theoretic Analysis of the Language Interface Bottleneck: Under Assumption 9, serialization creates a strictly positive Bayes-risk gap and therefore imposes an unavoidable statistical price on language-only pipelines.The strict gap follows because E[Y | X] differs from E[Y | T(X)] on a set of positive probability.
  • A.2. Information-Theoretic Analysis of the Language Interface Bottleneck: Proposition 13 rules out any language-only function class factoring through T from recovering the advantage of a direct-access foundation model.This proposition is identified as the central building block for proving strict improvements in later sections.
  • A.3. Expressivity and Solvability of EywaAgent: EywaAgent contains the language-only class, FLLM ⊆ FEywa, because skipping specialized calls reproduces the original LLM agent.Thus EywaAgent preserves existing language-only behavior while allowing domain-specific foundation-model invocation.
  • A.3. Expressivity and Solvability of EywaAgent: Under Assumption 8, EywaAgent achieves strictly lower optimal risk than language-only agents and has a strictly larger solvable task space.The strict improvement comes from invoking the foundation model on an informative component while leaving other components unchanged.
  • A.4. Multi-Agent Propagation: EywaMAS: An LLM-only multi-agent system remains information-closed under serialization, since all exchanged messages and its final output are functions of T(X).Adding more language agents cannot recover information discarded before the interaction begins.
  • A.4. Multi-Agent Propagation: EywaMAS: If an EywaAgent’s message reaches the final node within the interaction horizon, EywaMAS attains zero risk on a non-empty task family while the same-topology LLM-only MAS retains strictly positive overall loss.The conditions require finite topology diameter D and interaction horizon T ≥D, with a foundation model solving the task-relevant component perfectly.
  • A.5. Adaptive Orchestration: EywaOrchestra: EywaOrchestra’s oracle adaptive routing dominates the best fixed multi-agent configuration, with strict improvement when no fixed configuration is optimal for all tasks.The framework combines model adaptivity over LLM and foundation-model backbones with structural adaptivity over candidate topologies.

A.6. Efficiency and Token Complexity

Eywa reduces language-token costs for structured scientific inputs by routing data through specialized foundation models, with a reported ∼30% token reduction. Its token and latency advantages grow as modality size increases because language-side processing is drastically reduced.

  • Token complexity: ∼30% token reduction is reported for Eywa beyond task performance.The reduction is reported in Section 5 as an efficiency gain.
  • Token complexity: For long time series and large tables, TokenCostEywa/TokenCostLLM →0 as modality size grows.The condition is L(xk) ≫ Lcall + Lψ(ok).
  • Wall-clock latency: Eywa latency is O(αLLM · (Lcall + Lψ(ok)) + αFM), compared with Θ(αLLM · L(xk)) for language-only agents.The Eywa bound reflects structured foundation-model invocation plus adapted-response tokens.
  • Wall-clock latency: Specialized foundation models reduce invocation overhead because they are typically smaller and optimized for fixed-shape tensor inputs.These properties make inference lightweight and accelerator-friendly.
  • Wall-clock latency: For non-trivial inputs, αFM ≤ αLLM · L(xk), leaving Eywa latency dominated by the reduced LLM-side term.This follows from the stated latency comparison in Proposition 20.

A.7. Summary of Theoretical Results … B.2. Future Directions

The theoretical results characterize the language-interface bottleneck and establish Eywa’s expressivity advantages across single-agent and multi-agent settings. Related work positions Eywa as complementary to language-centric structured-data methods, while future directions target scalable integration, learned orchestration, tighter model coupling, and broader evaluation.

  • A.7. Summary of Theoretical Results: Language-only agents are fundamentally limited because serialization discards task-relevant information that downstream reasoning cannot recover.The summary attributes this conclusion to Lemmas 11–12 and Proposition 13.
  • A.7. Summary of Theoretical Results: EywaAgent contains the LLM-only class while achieving strictly smaller risk and strictly enlarging the set of perfectly solvable tasks.Proposition 14 and Theorem 15 provide the expressivity analysis supporting this result.
  • A.7. Summary of Theoretical Results: Under mild reachability, one EywaAgent can propagate recovered information to the multi-agent system output, whereas LLM-only MAS retain the serialization bottleneck.Lemma 16 and Theorem 17 lift the single-agent advantage to the multi-agent setting.
  • B.1.1. Agentic AI Systems.: Agentic systems address complex tasks through explicit planning or decomposition followed by iterative execution, reflection, clarification, and continuation.These structured reasoning procedures are described as a common paradigm for tasks exceeding a single forward pass.
  • B.1.2. Modeling Structured Data with LLMs.: Existing structured-data approaches textualize non-textual inputs, assist specialized models, or re-program LLMs, while Eywa instead introduces domain-specific foundation models into agentic systems.Eywa’s complementary perspective allows LLM agents to coordinate, invoke, and reason with specialized models rather than forcing structured data into language.
  • B.2. Future Directions: Future work should scale heterogeneous ecosystems, learn orchestration policies, strengthen LLM–scientific-model coupling, and extend EywaBench across domains, modalities, tasks, and expert models.Challenges include incompatible input formats, varying interpretability, expert-selection decisions, tighter interfaces beyond model context protocol, and broader benchmark coverage.

B.3. Limitation · C. EywaBench Details · C.1. Source Datasets

EywaBench combines four scientific benchmark sources spanning language, time-series, and tabular modalities, with a controlled 200-instance evaluation subset for balanced, reproducible assessment. Eywa remains limited by dependence on underlying models and interfaces, incomplete domain coverage, and the computational overhead of heterogeneous collaboration.

  • B.3. Limitation: Eywa’s performance depends on the reasoning ability of underlying LLMs, predictive quality of scientific experts, and interface reliability.The framework is expected to benefit as general-purpose and domain-specific foundation models improve.
  • B.3. Limitation: EywaBench cannot exhaustively represent scientific applications because some specialized data formats, assumptions, and expert models are excluded.Broader datasets and scientific workflows remain an important future direction.
  • B.3. Limitation: Heterogeneous collaboration can add computation, latency, and communication overhead compared with single-agent baselines.The paper identifies efficient orchestration, selective expert invocation, and adaptive stopping as possible mitigations.
  • C. EywaBench Details: EywaBench consists of 4 datasets: DeepPrinciple, MMLU-Pro, fev-bench, and TabArena.Together, these sources cover scientific question answering, multi-task understanding, time-series forecasting, and tabular learning.
  • C.1. Source Datasets: DeepPrinciple requires iterative reasoning, hypothesis generation, and experimental design across diverse scientific contexts.It covers biology, chemistry, materials science, and physics through scenario-grounded scientific discovery questions.
  • C.1. Source Datasets: MMLU-Pro contributes over 12,000 curated questions across 14 domains, with selected scientific tasks converted from multiple choice to open-ended answers.The benchmark requires models to produce direct answers.
  • C.1. Source Datasets: FEV-Bench and TabArena provide heterogeneous predictive data: 100 time series with realistic covariates and 51 tabular datasets evaluated across approximately 25 million trained model instances.These sources emphasize diversity in time-series covariates, tabular feature spaces, and data distributions.
  • C.1. Source Datasets: EywaBench-V1 samples 200 task instances to maintain balanced domain and modality coverage while keeping heterogeneous-system evaluation reproducible and feasible for expert inspection.The source distributions are near-uniform, and the benchmark spans natural-language, time-series, and tabular modalities without a single source dominating.

C.2. Data Schema

EywaBench-V1 is a self-contained, dictionary-encoded Parquet benchmark with 200 scientific task instances organized under a six-field schema. Its coverage spans nine sub-domains and three modalities, with all 27 combinations represented.

  • Schema: EywaBench contains 200 task instances stored in a dictionary-encoded Parquet file with six fields.The benchmark is stored as a dictionary-encoded Parquet file, and each row represents one self-contained scientific problem.
  • Schema: EywaBench-V1 uses Snappy compression and is generated with PyArrow 20.0.0 and pandas 2.3.3.Parsing scripts convert each source dataset’s original format into the unified EywaBench schema.
  • Coverage: The benchmark covers nine sub-domains across three modalities, populating all 27 sub-domain–modality cells.The largest sub-domain accounts for 14.0% of the benchmark.

C.3. Composition and Coverage Analysis · C.4. Metrics · D. Experiment Details

EywaBench-V1 contains 200 parametrically generated task samples spanning balanced scientific domains, sub-domains, modalities, and 67 source datasets. Its modality-specific utilities are bounded in [0, 1] and aggregated as unweighted means, with separate scoring procedures for natural-language, time-series, and tabular outputs.

  • C.3. Composition and Coverage Analysis: EywaBench-V1 contains N=200 task samples and can scale through resampling or adding domains and modalities without changing its schema.The construction pipeline scales along independent sample-volume and domain/modality-coverage axes.
  • C.3. Composition and Coverage Analysis: The benchmark is near-uniform across parent domains, sub-domains, and modalities, with parent coverage of 32.0%/30.0%/38.0% and modality coverage of 41.0%/39.0%/20.0%.Each of the nine sub-domains contains 15–28 instances, and normalized Shannon entropy is Hn=0.995 at the parent level and Hn=0.993 at the sub-domain level.
  • C.3. Composition and Coverage Analysis: Every one of the nine sub-domains contains all three modalities, yielding 100% cross-modal taxonomy coverage.This supports modality-specific comparisons across scientific fields without conflating modality and domain effects.
  • C.3. Composition and Coverage Analysis: EywaBench-V1 draws from 67 distinct source datasets spanning 21 physical-science, 19 life-science, and 28 social-science sources.Examples include ETT, ERCOT, NASDAQ, Jena Weather, LOOP-SEATTLE, Concrete Compressive Strength, and Superconductivity.
  • C.4. Metrics: Each task instance produces a modality-specific utility ui ∈[0, 1], and reported slice-level scores are unweighted means with sample standard deviation.The same aggregation is applied to runtime and the four token-cost components.
  • C.4. Metrics: Natural-language outputs use normalized exact match, numeric relative error, and a lexical fallback cascade to compute utility.Normalization trims whitespace and quotes, collapses inner whitespace, and maps selected Unicode variants to ASCII; lexical near-misses remain below exact-match and numeric-error scores.
  • C.4. Metrics: Time-series and tabular numeric regression tasks use a normalized sMAPE+MAAPE combination, while tabular classification uses top-1 accuracy.For time series, the denominator floor is ε=10−2; sharing the regression rule across modalities keeps numeric-target errors on a common scale.

D.1. More Detailed Ablations … E. Prompt Templates

Across backbone ablations, case studies, efficiency trade-offs, and prompt templates, Eywa consistently leverages specialized foundation models through language-guided interfaces and orchestration. Stronger backbones improve utility but show diminishing returns, while Eywa configurations preserve competitive utility and token efficiency across scientific domains.

  • D.1. More Detailed Ablations: Eywa remains effective across gpt-4.1-nano, gpt-5-nano, and gpt-5-mini backbones in single-agent, multi-agent, and dynamic-orchestration settings.With gpt-4.1-nano, EywaAgent, EywaMAS, and EywaOrchestra achieve overall utilities of 0.5680, 0.6236, and 0.6210, respectively, surpassing the single-agent gpt-5-nano baseline.
  • D.1. More Detailed Ablations: EywaAgent, EywaMAS, and EywaOrchestra improve from gpt-4.1-nano to gpt-5-nano by 15.5%, 8.4%, and 8.6%, respectively, while gpt-5-mini yields diminishing returns.EywaAgent rises from 0.6558 to 0.6640 (+1.25%) from gpt-5-nano to gpt-5-mini, with slight degradation on Biology, Clinic, Economy, and Business.
  • D.2. Case Study: Language-only agents can satisfy output formatting on structured financial forecasting tasks yet collapse to a last-value persistence baseline when reasoning over serialized values.The language-only agent identifies the schema, produces the required forecast horizon, and returns a dataframe-style response, but lacks specialized predictive capability.
  • D.2. Case Study: EywaAgent activates Chronos through Tsaheylu, while EywaOrchestra dynamically selects an execution configuration, foundation model, and plan for each task.For Miami house-price prediction, EywaOrchestra selects a lightweight single-agent configuration using gpt-5-nano and TabPFN; the LLM verifies and formats the result.
  • D.3. Detailed Utility and Token Consumption: EywaAgent, EywaMAS, and EywaOrchestra remain on or near the Pareto frontier across all nine sub-domains, while LLM-only baselines are typically dominated.On Material, Debate, MoA, and X-MAS consume 14,000–25,000 tokens while achieving lower utility than EywaMAS; similar patterns occur on Energy, Biology, and Drug.
  • D.3. Detailed Utility and Token Consumption: EywaAgent averages 3,137 tokens per task, compared with 4,469 for the single-agent gpt-5-nano baseline and 8,673–16,537 for MAS baselines.EywaMAS generally attains the highest utility at moderate cost, whereas EywaOrchestra reduces average token consumption from 11,214 to 8,335 while maintaining close or matching utility.
  • E. Prompt Templates: EywaBench-V1 uses a unified task-execution prompt specifying task role, optional model/tool context, structured input, expected output size, and response format.Specialized input tags and additional instructions preserve modality-specific information across heterogeneous task types.
  • E. Prompt Templates: EywaOrchestra’s planner converts task descriptions, domains, and task types into constrained structured JSON configurations for transparent, automatically executable orchestration.The configuration specifies execution setting, Eywa usage, foundation model, multi-agent topology, and participating agents.
Loading 2604.27351v1…