Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 48

GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis GPS-Bench:用于自动化政策分析的治理政策基准

GPS-Bench is an evidence-grounded benchmark for governance policy simulation that links policies to actors, actions, and downstream impacts using legislative records, lobbying disclosures, regulatory documents, corporate filings, and economic data Actors are reconstructed from dated public records rather than prompted as generic archetypes, making each persona an evidence object with verifiable provenance The benchmark uses a two-tier evaluation system: human-annotated Gold sets for testing and 提出GPS-Bench,首个基于证据的治理政策模拟基准,将政策与相关行为者、行为者行动及下游影响通过立法记录、游说披露、监管文件等公共证据链接 行为者从日期记录中重建而非作为原型提示,使"人格"成为具有溯源性的证据对象 构建人工标注的Gold评估集与LLM标注的Silver监督集,确保测试标签不被污染 在统一政策状态下对比联合推理、独立/通信代理、图方法和权重级微调,发现微调在预测上最强,分解的价值在于提供机制解释 代理持有私有且非相同的证据,各自看到自身影响条款,并与命名伙伴提出具体联合提案,形成的联盟可与记录中的承诺进行验证

62
Hot 热度
75
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • GPS-Bench is an evidence-grounded benchmark for governance policy simulation that links policies to actors, actions, and downstream impacts using legislative records, lobbying disclosures, regulatory documents, corporate filings, and economic data
  • Actors are reconstructed from dated public records rather than prompted as generic archetypes, making each persona an evidence object with verifiable provenance
  • The benchmark uses a two-tier evaluation system: human-annotated Gold sets for testing and LLM-generated Silver labels from retrieved evidence for supervision only
  • Fine-tuning on grounded records produces the strongest actor-level impact predictions, while multi-agent decomposition adds interpretability rather than predictive accuracy
  • GPS-Bench enables controlled comparisons across joint reasoning, independent agents, communicating agents, graph-based methods, and fine-tuning over identical policy states

Why It Matters

This benchmark addresses a critical gap in AI-driven policy analysis: the lack of empirically grounded evaluation for multi-agent policy simulations. By anchoring actor personas in real legislative and economic records rather than generic stereotypes, GPS-Bench provides a rigorous framework for testing whether multi-agent approaches genuinely improve policy outcome prediction and interpretation.

Technical Details

  • GPS-Bench constructs actor personas from dated public evidence (legislative records, lobbying disclosures, regulatory documents, corporate filings, economic data) rather than prompting LLMs with archetype descriptions
  • Evaluation uses a Gold/Silver split: human-annotated cases form the Gold test set, while LLM-generated labels from retrieved evidence serve as Silver supervision data that is never used for testing
  • The benchmark enables controlled comparison across multiple inference modes: joint reasoning, independent agents, communicating agents with private evidence, graph-based methods, and weight-level fine-tuning—all operating on the same grounded policy state and emitting the same schema
  • Communicating agents hold private, non-identical evidence (each seeing only their own exposure clause) and negotiate with named partners through concrete joint proposals specifying offers, requirements, and coalition rationale
  • Coalition formations can be validated against actual commitments recorded in the evidence base

Industry Insight

  • The finding that fine-tuning outperforms multi-agent decomposition for prediction accuracy suggests practitioners should prioritize evidence-grounded training data over complex agent architectures when pure predictive performance is the goal
  • Multi-agent approaches remain valuable for interpretability and mechanism explanation, making them suitable for transparency-critical applications where understanding coalition dynamics matters more than raw prediction accuracy
  • The Gold/Silver evaluation methodology offers a reusable template for building benchmarks in domains where human annotation is expensive but LLM-generated supervision from retrieved evidence can supplement training without contaminating test sets

TL;DR

  • 提出GPS-Bench,首个基于证据的治理政策模拟基准,将政策与相关行为者、行为者行动及下游影响通过立法记录、游说披露、监管文件等公共证据链接
  • 行为者从日期记录中重建而非作为原型提示,使"人格"成为具有溯源性的证据对象
  • 构建人工标注的Gold评估集与LLM标注的Silver监督集,确保测试标签不被污染
  • 在统一政策状态下对比联合推理、独立/通信代理、图方法和权重级微调,发现微调在预测上最强,分解的价值在于提供机制解释
  • 代理持有私有且非相同的证据,各自看到自身影响条款,并与命名伙伴提出具体联合提案,形成的联盟可与记录中的承诺进行验证

为什么值得看

本文解决了LLM政策模拟缺乏有效性验证的核心痛点,通过证据 grounding 和受控对比实验,为多智能体政策模拟提供了可复现的评估框架。对AI治理、政策分析和多智能体系统研究者而言,GPS-Bench提供了从" plausible behavior"走向"evidence-grounded prediction"的关键基础设施。

技术解析

  • 数据架构:整合立法记录、游说披露、监管文件、公司申报、经济数据等多源公共证据,构建政策-行为者-行动-影响的完整证据链,每个行为者persona均具有可追溯的出处。
  • 评估体系设计:采用双层标注策略——人工标注构成Gold测试集,LLM基于检索证据标注的Silver集仅用于监督训练,严格隔离测试标签,避免数据泄露。
  • 对比实验范式:所有推理模式读取相同接地状态并输出相同schema,实现可控对比:联合推理 vs 独立代理 vs 通信代理 vs 图方法 vs 权重级微调。
  • 核心发现:在接地记录上进行权重级微调获得最强的行为者级影响预测;分解(decomposition)本身不超越微调,但通过让代理持有私有证据、提出具体联合提案,提供了可验证的机制解释。
  • 联盟验证机制:代理之间提出包含"提供什么、需要什么回报、为何合作优于单独行动"的具体提案,形成的联盟结构可与历史记录中的承诺进行交叉验证。

行业启示

  • 政策AI的验证标准升级:行业需从"生成合理叙事"转向"证据 grounding + 可验证预测",GPS-Bench为政策类AI应用树立了新的评估基准。
  • 多智能体价值的重新定位:实验表明,在纯预测任务上微调优于复杂分解;多智能体架构的核心价值在于提供可解释的机制和联盟动态,而非单纯提升预测精度。
  • 数据策略建议:构建政策AI系统时,应优先投资多源公共证据的整合与溯源能力,而非仅依赖大模型推理;人工标注的Gold集与LLM辅助的Silver集的分层设计值得借鉴。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Benchmark 基准测试 Evaluation 评测 Dataset 数据集 Policy 政策