Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Transduction is All You Need for Structured Data Workflows

Alfio Gliozzo, Naweed Khan, Christodoulos Constantinides, Nandana Mihindukulasooriya, Nahuel Defosse, Gaetano Rossiello, Junkyu Lee

TL;DR

Agentics introduces a data-centric paradigm for LLM-based structured workflows built on Logical Transduction Algebra (LTA), where agents are stateless transducers operating over typed data via the transduction operator $\ll$. By embedding schemas (via Pydantic models) and asynchronous MapReduce-style primitives (amap, areduce), the framework enables robust, scalable, and auditable pipelines that outperform traditional prompt-centric approaches on data wrangling, text-to-SQL, and domain-specific reasoning tasks. The paper provides formal definitions, a Python implementation around an AG meta-class and PydanticTransducer, and extensive experiments (schema matching, T2SQL, data imputation, domain MCQA, and DiscoveryBench) illustrating improved accuracy, robustness to perturbations, and favorable runtime characteristics. Collectively, Agentics offers a principled, modular approach to generative structured data workflows with direct implications for enterprise data engineering, reproducibility, and scalable AI-assisted data tasks.

Abstract

This paper introduces Agentics, a functional agentic AI framework for building LLM-based structured data workflow pipelines. Designed for both research and practical applications, Agentics offers a new data-centric paradigm in which agents are embedded within data types, enabling logical transduction between structured states. This design shifts the focus toward principled data modeling, providing a declarative language where data types are directly exposed to large language models and the data values are composed through transductions between input and output types. We present a range of structured data workflow tasks and empirical evidence demonstrating the effectiveness of this approach, including data wrangling, text-to-SQL semantic parsing, and domain-specific multiple-choice question answering, and data-driven scientific discovery tasks.

Transduction is All You Need for Structured Data Workflows

TL;DR

Agentics introduces a data-centric paradigm for LLM-based structured workflows built on Logical Transduction Algebra (LTA), where agents are stateless transducers operating over typed data via the transduction operator . By embedding schemas (via Pydantic models) and asynchronous MapReduce-style primitives (amap, areduce), the framework enables robust, scalable, and auditable pipelines that outperform traditional prompt-centric approaches on data wrangling, text-to-SQL, and domain-specific reasoning tasks. The paper provides formal definitions, a Python implementation around an AG meta-class and PydanticTransducer, and extensive experiments (schema matching, T2SQL, data imputation, domain MCQA, and DiscoveryBench) illustrating improved accuracy, robustness to perturbations, and favorable runtime characteristics. Collectively, Agentics offers a principled, modular approach to generative structured data workflows with direct implications for enterprise data engineering, reproducibility, and scalable AI-assisted data tasks.

Abstract

This paper introduces Agentics, a functional agentic AI framework for building LLM-based structured data workflow pipelines. Designed for both research and practical applications, Agentics offers a new data-centric paradigm in which agents are embedded within data types, enabling logical transduction between structured states. This design shifts the focus toward principled data modeling, providing a declarative language where data types are directly exposed to large language models and the data values are composed through transductions between input and output types. We present a range of structured data workflow tasks and empirical evidence demonstrating the effectiveness of this approach, including data wrangling, text-to-SQL semantic parsing, and domain-specific multiple-choice question answering, and data-driven scientific discovery tasks.

Paper Structure

This paper contains 78 sections, 6 theorems, 38 equations, 5 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Logical Transduction Algebra
  4. Definition: Agentics (AG)
  5. Definition: Transduction operator ($\ll$)
  6. Definition: Prompt function (P)
  7. Lemma: Properties of LTA
  8. AType Operators
  9. Asynchronous MapReduce.
  10. Technical Implementation
  11. Example Usage of Metaclass AG
  12. Metaclass AG
  13. Pydantic Models as Schemas
  14. Logical Transduction ($\ll$)
  15. Map/Reduce Paradigm
  16. ...and 63 more sections

Key Result

Proposition 1

Let $AG[X]$ be an Agentic structure and let $\xi$ be the set of all instances of $AG[X]$. Define a binary operation $\circ$ on $\xi$ such that for any $\mathbf{x}_1, \mathbf{x}_2 \in \xi$, their composition $\mathbf{x} = \mathbf{x}_1 \circ \mathbf{x}_2$ is an Agentic instance whose state list is the

Figures (5)

  • Figure 1: Logical Transduction Applied to Sentiment Summary
  • Figure 2: Average time (sec) per question
  • Figure 3: Domain Specific MCQA Perturbation Results.
  • Figure 4: Improvement of test score over iterations: The x-axis represents the number of iterations, and the y-axis shows the test score evaluated using the best prompt template found up to that iteration.
  • Figure 5: Average running time per iteration: The x-axis represents the batch size of the asynchronous execution, and the y-axis shows the average running time in seconds.

Theorems & Definitions (27)

  • Definition 1: Types
  • Definition 2
  • Definition 3: Agentic Structure $AG$
  • Proposition 1: Monoid of Agentic Instances
  • proof
  • Definition 4: Product of Agentic Structures
  • Proposition 2: Monoid of Agentic Product
  • proof
  • Definition 5: Equivalence Relation on Agentic Instances
  • Definition 6: Statewise Equivalence of Agentic Instances
  • ...and 17 more