Technical Summary: The CitizenQuery Benchmark

Problem Statement

"Citizen queries"—questions from the public regarding government policies, services, and guidance (e.g., benefits, taxes, immigration)—represent a high-stakes use case for Large Language Models (LLMs). While LLMs offer the potential to provide context-aware, accessible, and rapid information services, their deployment in the public sector is hindered by the risk of hallucination and misinformation. Inaccurate responses to citizen queries can lead to severe, often invisible negative ramifications for individuals relying on such advice. Currently, there is a lack of standardized benchmark datasets or evaluation suites specifically designed to measure LLM performance in the context of citizen queries, making it difficult for governments and stakeholders to assess the trustworthiness of AI integration in public services.

Methodology

Dataset Curation: CitizenQuery-UK

The authors introduce CitizenQuery-UK, a novel benchmark dataset comprising 22,066 synthetic citizen query-response pairs.

• Source Material: The dataset is derived entirely from textual information on gov.uk, the official UK government website, which serves as the single source of authoritative truth for the benchmark.
• Generation Process:
  • Corpus Construction: 2,781 gov.uk pages were scraped, organized into a 3-layer taxonomy (Home page $\rightarrow$ Domains $\rightarrow$ Sub-domains).
  • Query Synthesis: To ensure naturalistic phrasing, query generation was grounded in real-world data from online public forums (primarily Reddit) to mimic colloquial, context-rich user inputs, avoiding the sterile phrasing typical of synthetic generation.
  • Model: The Qwen 2.5 72B-Instruct model was used to generate the Q&A pairs, conditioned on specific personas (varying by age, education, digital literacy, profession, income, and region) and metadata.
  • Quality Assurance: A rigorous human annotation protocol involving three annotators reviewed a stratified sample to refine generation prompts, ensuring responses were factually grounded and accessible. Toxicity and bias scores were generated via Cleanlab Studio but deemed largely irrelevant to the core objectives, with toxicity remaining low.
• Dataset Structure: The dataset includes 139 subdomains across 16 main domains. Prompts are categorized as procedural, instructional, informational, or comparative. Each entry includes the query, the expected answer (ground truth), and extensive metadata.

Evaluation Pipeline

The authors propose an evaluation methodology adapted from FActScore and the SAFE architecture, consisting of three main components:

1. Abstention Detection (AB):

   • Utilizes a fine-tuned classifier (LibrAI/longformer-action-ro) to identify if a model refuses or is unable to answer a query.
   • The authors adapt the standard annotation scheme to treat "disclaimers" (label 3) as pass rather than abstention, acknowledging that disclaimers are standard and encouraged in government advice.
   • A confidence threshold ($AB_{th} = 0.925$) is applied to minimize false positives.

2. Atomised Fact Generation (AFG):

   • Decomposes both the expected response ($a_i$) and the model's generated response ($\hat{a}_i$) into atomic facts/claims.
   • Uses a retrieval-based few-shot prompting strategy with an instruct-tuned LLM ($M_{AFG}$) to break down sentences into granular, verifiable components.
   • Atomic facts for the ground truth are pre-computed to ensure a fixed reference point.

3. Atomised Fact Verification (AFV):

   • Employs an LLM-as-a-judge ($M_{AFV}$) to verify if the model's claims are supported by the ground truth facts.
   • Metrics:
     • F1@K: Calculates the harmonic mean of precision and recall relative to the ground truth facts ($K$). This penalizes models for redundancy (repeating facts) and excessive verbosity, rewarding information density.
     • Verbosity ($\Delta K$): Measures the difference between the number of claims generated by the model ($\hat{K}$) and the ground truth ($K$). Positive values indicate over-generation; negative values indicate omission.

Key Results

The benchmark evaluated 11 LLMs (including Llama, Claude, Gemini, Qwen, and OpenAI families) under both Zero-Shot and Few-Shot conditions.

• Performance Profiles:

  • Factuality: Median F1@K scores were generally high (often >0.90), indicating that models often answer correctly. However, mean scores were lower than medians, revealing a "long tail" of poor performance and high variance (standard deviations ranging from 0.18 to 0.36).
  • Model Families: Closed-source models like Claude-4.5-Haiku showed high accuracy but also high verbosity. Open-weight models (e.g., Qwen3-32B, Llama-3.3-70B) performed competitively against closed models. Kimi-K2-Instruct showed lower factuality but higher abstention rates.
  • Few-Shot Impact: Contrary to expectations, providing few-shot examples generally decreased average F1@K scores and increased variance, while slightly reducing verbosity. In some cases, models generated less information but that information was less factually supported.

• Key Behavioral Findings:

  • Low Abstention: Models rarely abstain from answering (rates <1.5%), even when uncertain.
  • High Verbosity: Models consistently over-generate information ($\Delta K$ was positive and significant, e.g., +16.63 for Claude-4.5-Haiku in Zero-Shot).
  • Variance: High standard deviations suggest that while models are often correct, their failure modes are unpredictable, posing a reliability risk.

Significance and Claims

The paper claims that CitizenQuery-UK provides the first evidence-based framework for assessing LLM trustworthiness in citizen query tasks. Its significance lies in:

1. Evidence for Public Sector Adoption: It offers a mechanism for governments to evaluate AI accuracy before integration, moving beyond theoretical assumptions to measured performance.
2. Highlighting the "Verbosity vs. Reliability" Trade-off: The results suggest that current LLMs prioritize generating information over admitting uncertainty. In high-stakes citizen queries, this "eagerness to talk" combined with high variance creates a risk where users may over-trust inaccurate or outdated information.
3. Call for Acknowledged Fallibility: The authors argue that trustworthy AI in the public sector requires models that can acknowledge their limitations (fallibility) rather than over-generating potentially incorrect claims.
4. Foundation for Future Governance: Built entirely on open data, the benchmark lays the groundwork for centralized evaluation suites that can guide policy and procurement decisions regarding AI in the public sector.

The authors conclude that while LLMs show promise, their current tendency toward high variance and over-generation makes them unsuitable for direct deployment without significant improvements in reliability and mechanisms for acknowledging uncertainty.