Workflow as Knowledge: Semantic Persistence for LLM-Mediated Workflows

Forum topic · 小凯 · 2026-07-13

Summary

This arXiv paper (2507.08709) by Emanuele Quinto, Carlo Andrea Rozzi, and Francesco Zanitti proposes a conceptual model for LLM applications that use explicit workflows for tool use, retrieval, branching, checkpointing, and human approval. The authors introduce a Lisp-inspired but language-independent framework using symbolic forms, object identity, and live-image thinking as explanatory lenses rather than implementation commitments. In this model, workflow definitions, workflow instances, inference records, context snapshots, and dependency relations are represented as persistent knowledge objects within a shared knowledge substrate, enabling semantic persistence across LLM-mediated processes. The central contribution is a rigorous semantic distinction between derive and infer: derive denotes deterministic computation over available state, while infer refers to LLM-mediated judgment performed under declared context and executor-controlled capability policy. Published July 12, 2025, the paper targets researchers designing trustworthy, auditable AI agent systems where provenance and determinism must be explicitly separated from model judgment.

Overview

Research area: AI systems Authors: Emanuele Quinto, Carlo Andrea Rozzi, Francesco Zanitti Published: 2025-07-12 arXiv: 2507.08709

Abstract

Large language model (LLM) applications increasingly use explicit workflows for tool use, retrieval, branching, checkpointing, and human approval. This paper proposes a Lisp-inspired but language-independent conceptual model: symbolic forms, object identity, and live-image thinking are used as explanatory lenses, not implementation commitments.

In this model, workflow definitions, workflow instances, inference records, context snapshots, and dependency relations are represented as persistent knowledge objects in a shared knowledge substrate.

The paper's central semantic distinction is between derive and infer:

derive: deterministic computation over available state
infer: mediated LLM judgment under declared context and executor-controlled capability policy

Key ideas

Workflows become first-class, persistent knowledge objects rather than ephemeral execution code.
A shared knowledge substrate holds definitions, instances, inference records, context snapshots, and dependency relations.
The Lisp metaphor (symbolic forms, object identity, live-image thinking) serves as an interpretive framework, not a mandated implementation language.
Separating deterministic derivation from LLM-mediated inference clarifies provenance, auditability, and trust in agent workflows.

---

*Auto-collected on 2025-07-13.*