Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating LLMs for Police Decision-Making: A Framework Based on Police Action Scenarios

Sangyub Lee, Heedou Kim, Hyeoncheol Kim

TL;DR

This paper introduces PAS, a Police Action Scenarios framework for evaluating LLMs in police operations, addressing the gap where general evaluation methods fall short for domain-specific tasks. PAS formalizes a five-stage process (Scenario definition, Reference responses, LLM generation, metric extraction, and expert-informed evaluation) and represents outputs as $E_{ ext{police}} = f(S, R, G, M, P)$ with $G = \text{LLM}(S)$ and $P = h(G, R, M)$. The authors built a police-manual–based QA dataset from 8,348 curated entries across 75 questions and validated five core metrics (Logical Correctness, Completeness, Factuality, Logical Efficiency, Logical Robustness) as strong predictors of response quality. Experiments with GPT-4, Gemini, and Claude show that commercial LLMs underperform relative to expert-derived references, especially on specialized police knowledge, underscoring the need for expert-in-the-loop validation and jurisdiction-tailored references. The work provides a scalable benchmark and prompts for replicability, with implications for safer, more reliable AI-assisted policing and potential generalization to other high-stakes professions.

Abstract

The use of Large Language Models (LLMs) in police operations is growing, yet an evaluation framework tailored to police operations remains absent. While LLM's responses may not always be legally incorrect, their unverified use still can lead to severe issues such as unlawful arrests and improper evidence collection. To address this, we propose PAS (Police Action Scenarios), a systematic framework covering the entire evaluation process. Applying this framework, we constructed a novel QA dataset from over 8,000 official documents and established key metrics validated through statistical analysis with police expert judgements. Experimental results show that commercial LLMs struggle with our new police-related tasks, particularly in providing fact-based recommendations. This study highlights the necessity of an expandable evaluation framework to ensure reliable AI-driven police operations. We release our data and prompt template.

Evaluating LLMs for Police Decision-Making: A Framework Based on Police Action Scenarios

TL;DR

This paper introduces PAS, a Police Action Scenarios framework for evaluating LLMs in police operations, addressing the gap where general evaluation methods fall short for domain-specific tasks. PAS formalizes a five-stage process (Scenario definition, Reference responses, LLM generation, metric extraction, and expert-informed evaluation) and represents outputs as with and . The authors built a police-manual–based QA dataset from 8,348 curated entries across 75 questions and validated five core metrics (Logical Correctness, Completeness, Factuality, Logical Efficiency, Logical Robustness) as strong predictors of response quality. Experiments with GPT-4, Gemini, and Claude show that commercial LLMs underperform relative to expert-derived references, especially on specialized police knowledge, underscoring the need for expert-in-the-loop validation and jurisdiction-tailored references. The work provides a scalable benchmark and prompts for replicability, with implications for safer, more reliable AI-assisted policing and potential generalization to other high-stakes professions.

Abstract

The use of Large Language Models (LLMs) in police operations is growing, yet an evaluation framework tailored to police operations remains absent. While LLM's responses may not always be legally incorrect, their unverified use still can lead to severe issues such as unlawful arrests and improper evidence collection. To address this, we propose PAS (Police Action Scenarios), a systematic framework covering the entire evaluation process. Applying this framework, we constructed a novel QA dataset from over 8,000 official documents and established key metrics validated through statistical analysis with police expert judgements. Experimental results show that commercial LLMs struggle with our new police-related tasks, particularly in providing fact-based recommendations. This study highlights the necessity of an expandable evaluation framework to ensure reliable AI-driven police operations. We release our data and prompt template.
Paper Structure (27 sections, 1 equation, 3 figures, 6 tables)

This paper contains 27 sections, 1 equation, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. LLM Evaluation Methods
  4. Challenges in Domain-Specific LLM Evaluation
  5. Framework for Evaluating LLMs in Policing
  6. Step 1: Policing Scenario Definition($S$)
  7. Step 2: Reference Responses ($R$)
  8. Step 3: LLM Response Generation ($G$)
  9. Experiment
  10. Applying PAS to Real Police Data
  11. Simulating Operational Readiness
  12. Reference Answers from Police Manuals
  13. LLM Response Generation
  14. Core Metric Alignment
  15. Experiment Results
  16. ...and 12 more sections

Figures (3)

  • Figure 1: LLMs are increasingly used to support police. However, existing evaluation focus only on information accuracy heedouFei:Lawbenchhwang2022multi, which risks indiscriminate use in real-world scenarios by overlooking key police-specific considerations. Our framework offers guidance for improving LLM from a policing perspective.
  • Figure 2: An overview of the PAS. Unlike previous benchmark studies that primarily focused on accuracy in legal matching or crime classification heedoukwon2024aihwang2022multibaek2021smart, this study adopts a Police Action Scenario-based framework to evaluate LLMs. By simulating real-world police situations, the framework generates both LLM responses and expert reference answers. Furthermore, it designs domain-specific evaluation metrics to comprehensively assess the LLM’s situational applicability in policing contexts.
  • Figure 3: PAS Application in Real Police Work Sequences