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
This paper presents Agentics, a novel AI framework that leverages transduction between structured data types within a data-centric paradigm, enhancing various data workflow tasks with empirical validation.
Contribution
It introduces a new data-centric, transduction-based paradigm for structured data workflows, embedding agents within data types for logical data transformations.
Findings
Effective in data wrangling and semantic parsing tasks
Demonstrates improved performance in domain-specific question answering
Supports scientific discovery workflows
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.
Peer Reviews
Decision·ICLR 2026 Conference Desk Rejected Submission
1. This work is significant for addressing the critical challenge of building robust, composable, and scalable data workflows with LLMs, offering a principled alternative to fragile, chat-based agent systems. 2. The proposed "logical transduction" paradigm is well executed, which elegantly re-frames LLM agents as stateless transducers over typed data. It combines a sound theoretical foundation (provides formal definitions with proven properties) with a practical Python implementation, demonstra
1. This paper's motivation is not convinced and clear. It never truly clarifies why existing agent frameworks are fundamentally ill-suited for structured data, offering no tangible examples of where they fail and why a new paradigm is needed. 2. The experiments are not sufficient to verify the proposed framework's advantage, since the baselines are really weak (GPT-3.5), missing the comparisons with recent frameworks like DSPy or AutoGen on the same tasks. 3. The paper frames LTA as a major c
- The introduction of Logical Transduction Algebra (LTA) provides a principled, algebraic framework for composing LLM-based workflows. This formalism is a significant contribution, offering reproducibility, statelessness, and composability. - The asynchronous MapReduce-style programming model (aMap/aReduce) enables efficient parallel execution, which is crucial for large-scale data workflows.
- The framework is highly optimized for structured data workflows but may not generalize well to open-ended, context-heavy, or creative tasks that require rich contextual understanding or instruction-following. - For some tasks (e.g., schema matching), the baselines are limited to GPT-3.5. Comparing against more recent or task-specific models could strengthen the claims.
Rather than offering an incremental improvement on existing frameworks, the paper introduces a genuine paradigm shift, moving from chat-centric models to a principled, data-centric functional approach. The introduction of a formal Logical Transduction Algebra is a profound and novel contribution, as it grounds the behavior of LLM agents in a mathematically rigorous foundation—a rarity in this empirically-driven field. This theoretical rigor directly translates into the work's high quality. The
1. Lack of Direct Comparative Evaluation with Existing Frameworks: The most significant weakness is the absence of a direct, quantitative benchmark against the very frameworks it critiques, such as LangChain. The paper makes strong claims about overcoming their brittleness and lack of formal semantics. While the internal experiments and ablation studies are valuable, they only compare variations of the Agentics approach against a baseline. To truly substantiate the superiority of this new paradi
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsScientific Computing and Data Management