R-LAM: Reproducibility-Constrained Large Action Models for Scientific Workflow Automation

TL;DR

R-LAM tackles the reproducibility problem in Large Action Models when automating scientific workflows by enforcing structured action schemas, a deterministic execution engine, and a provenance-enabled trace. It decouples intent from execution, enabling replay and controlled forking while guaranteeing auditable, first-class scientific artifacts. The authors implement a lightweight Python framework and open-source PyPI artifact, then demonstrate improved reproducibility and reliability on synthetic and ML workflows without sacrificing adaptive control. The work argues that reproducibility constraints should be integral in scientific automation and provides a concrete reference implementation to foster reproducible research.

Abstract

Large Action Models (LAMs) extend large language models by enabling autonomous decision-making and tool execution, making them promising for automating scientific workflows. However, scientific workflows impose strict requirements on reproducibility, auditability, and deterministic execution, which are not satisfied by generic LLM-based agents. Unconstrained action generation can lead to silent state changes, non-deterministic executions, and irreproducible experimental results, limiting the applicability of LAMs in scientific settings. In this paper, we propose R-LAM, a reproducibility-constrained framework for applying Large Action Models to scientific workflow automation. R-LAM introduces structured action schemas, deterministic execution policies, and explicit provenance tracking to ensure that every action and intermediate artifact is auditable and replayable. The framework supports failure-aware execution loops and controlled workflow forking, enabling iterative experimentation without compromising reproducibility. We implement R-LAM as a lightweight Python framework and release it as an open-source PyPI package to facilitate reproducible research. An experimental evaluation of representative scientific workflows demonstrates that R-LAM improves reproducibility success rates and execution reliability compared to unconstrained LLM-based agents, while retaining adaptive control over workflow execution.

Figures (3)

• Figure 1: High-level architecture of R-LAM. The Large Action Model proposes structured actions that are validated and executed by a deterministic execution engine. All execution effects are recorded in a provenance-aware trace store, which provides observable state for subsequent reasoning. [Note: Final submission should include high-resolution PNG/PDF (>=300 dpi) per IEEE standards]
• Figure 2: Execution trace graph structure. Each node represents a single executed action with its inputs, outputs, execution environment binding, and terminal status. Nodes marked with a failure indicator represent actions that terminated unsuccessfully. Critically, failed actions are retained in the trace (not removed or implicitly retried), enabling subsequent recovery actions to be explicitly linked. This design preserves complete execution history and ensures that all downstream results can be traced to observed events. [Note: Final submission should include high-resolution PNG/PDF (>=300 dpi) per IEEE standards]
• Figure 3: Replay and forking mechanism in R-LAM. A new execution derives from an existing trace by reusing a shared prefix of logged actions without re-execution. Prior actions are replayed by reusing their persisted outputs, ensuring identical inputs across experiments. Forking introduces modifications only at the divergence point, while the original trace remains immutable. This enables controlled exploratory hypothesis testing while preserving auditability, determinism, and complete provenance without naively re-running the entire workflow. [Note: Final submission should include high-resolution PNG/PDF (>=300 dpi) per IEEE standards]