跳转至

文章背景与核心概要

在污水处理厂的日常运营中，操作人员经常需要面对复杂的因果诊断问题（例如“为什么 \(N_2O\) 浓度在上升？”或“如果将曝气量削减20%会发生什么？”）。传统的通用大语言模型由于缺乏特定厂区变量的交互特性和传播速度等物理约束，无法给出准确的专业解答。为了解决这一痛点，本文研究团队评估了将冻结的 Qwen2.5-32B-Instruct 模型与具备架构可解释性的污水处理模拟器（CCSS-IX）相结合的三种不同落地方法。

这三种方法分别为：实时模拟器谕示（Live simulator oracle）、结构化参数注入（Structured parameter injection）以及解耦式检索-推理（DRR）检索器。实验结果表明，这三种方法在198道因果基准测试题上的表现均大幅超越了传统的检索增强基线（最高仅为48%）。其中，方法一达到了99.5%的准确率；而采用解耦式检索-推理（DRR）的方法三不仅在跨厂区迁移时展现出高达88%的准确率，且单厂训练时间仅需约17秒，并在处理复杂干预后的系统行为时展现出100%的准确率，为工业领域的精准因果推理提供了高效、可移植的全新解决方案。

Simulator-Grounded Large Language Models for Industrial Causal Reasoning: Tool-Use, Structured Injection, and Plant-Portable Retrieval for Wastewater Treatment Decision Support

Authors: Gary Simethy, Daniel Ortiz Arroyo, Petar Durdevic
Submitted: 20 May 2026
Primary Subject: Computation and Language (cs.CL)
Additional Subjects: Artificial Intelligence (cs.AI)
arXiv: 2608.05151 [cs.CL] | DOI: 10.48550/arXiv.2608.05151
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0

Authors: Gary Simethy, Daniel Ortiz Arroyo, Petar Durdevic
Submitted: 20 May 2026
Primary Subject: Computation and Language (cs.CL)
Additional Subjects: Artificial Intelligence (cs.AI)
arXiv: 2608.05151 [cs.CL] | DOI: 10.48550/arXiv.2608.05151
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0

Abstract Summary

When wastewater operators ask causal questions—such as "why is \(N_2O\) rising?" or "what happens if I cut aeration by 20%?"—they need answers grounded in the specific interactions and propagation speeds of their plant's variables, rather than generic pretraining text.

This paper evaluates three distinct methods for grounding a frozen Qwen2.5-32B-Instruct model within an architecturally interpretable wastewater simulator (CCSS-IX):

  1. Method 1: Live simulator oracle
  2. Method 2: Structured parameter injection
  3. Method 3: Decoupled Recall-Reasoning (DRR) retriever

Key Findings & Performance

  • Causal Benchmark (198 questions):
  • Method 1: 99.5% accuracy
  • Method 2: 79.0% accuracy
  • Method 3: 75.8% accuracy
    (All three significantly outperform the strongest traditional retrieval-augmented baseline, which caps at 48%).
  • Plant Portability & Training Efficiency:
  • The DRR retriever (Method 3) utilizes 110M parameters and trains per plant in approximately 17 seconds.
  • Following cross-plant transfer to a biologically distinct plant, DRR still achieves 88% accuracy, whereas Method 2 relies on a static table that fails to transfer.
  • Counterfactual Benchmark (60 questions):
  • Only Method 3 successfully handles complex queries regarding system behavior following active interventions (+16.3 percentage points over Method 2; paired 95% CI [+7.1, +26.4] pp), hitting 100% accuracy on timescale and operating-regime categories.
  • Out-of-Domain Replication (AI2 Reasoning Challenge + OpenBookQA):
  • Applying the selective-retrieval mechanism to the ARC dataset achieves 79% accuracy (compared to unconstrained Llama-3.1-8B at 76% and full-injection at 74%), offering a +3 pp out-of-domain boost that confirms the approach is not restricted solely to wastewater treatment tasks.

When wastewater operators ask causal questions—such as "why is \(N_2O\) rising?" or "what happens if I cut aeration by 20%?"—they need answers grounded in the specific interactions and propagation speeds of their plant's variables, rather than generic pretraining text.

This paper evaluates three distinct methods for grounding a frozen Qwen2.5-32B-Instruct model within an architecturally interpretable wastewater simulator (CCSS-IX):

  1. Method 1: Live simulator oracle
  2. Method 2: Structured parameter injection
  3. Method 3: Decoupled Recall-Reasoning (DRR) retriever

Key Findings & Performance

  • Causal Benchmark (198 questions):
  • Method 1: 99.5% accuracy
  • Method 2: 79.0% accuracy
  • Method 3: 75.8% accuracy
    (All three significantly outperform the strongest traditional retrieval-augmented baseline, which caps at 48%).
  • Plant Portability & Training Efficiency:
  • The DRR retriever (Method 3) utilizes 110M parameters and trains per plant in approximately 17 seconds.
  • Following cross-plant transfer to a biologically distinct plant, DRR still achieves 88% accuracy, whereas Method 2 relies on a static table that fails to transfer.
  • Counterfactual Benchmark (60 questions):
  • Only Method 3 successfully handles complex queries regarding system behavior following active interventions (+16.3 percentage points over Method 2; paired 95% CI [+7.1, +26.4] pp), hitting 100% accuracy on timescale and operating-regime categories.
  • Out-of-Domain Replication (AI2 Reasoning Challenge + OpenBookQA):
  • Applying the selective-retrieval mechanism to the ARC dataset achieves 79% accuracy (compared to unconstrained Llama-3.1-8B at 76% and full-injection at 74%), offering a +3 pp out-of-domain boost that confirms the approach is not restricted solely to wastewater treatment tasks.

Metadata & Reference Information

  • ACM Classes: I.2.7; I.2.6; H.3.3; J.2
  • Comments: 20 pages, 2 figures, 8 tables. Preprint submitted to Elsevier.

Full-Text & Access Links

license icon

Metadata & Reference Information

  • ACM Classes: I.2.7; I.2.6; H.3.3; J.2
  • Comments: 20 pages, 2 figures, 8 tables. Preprint submitted to Elsevier.

Full-Text & Access Links

license icon