Paper deep dive

PPCR-IM: A System for Multi-layer DAG-based Public Policy Consequence Reasoning and Social Indicator Mapping

Zichen Song, Weijia Li

Year: 2026Venue: arXiv preprintArea: cs.SIType: PreprintEmbeddings: 31

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 7/20/2026, 12:28:34 PM

Summary

The paper introduces PPCR-IM, a system for multi-layer DAG-based public policy consequence reasoning and social indicator mapping. It uses an LLM-driven generator to construct directed acyclic graphs of intermediate consequences from policy descriptions, then maps these nodes to a fixed set of social indicators with qualitative impact directions. The system evaluates performance against government focus sets, demonstrating improved coverage and discovery of overlooked indicators compared to baseline LLMs.

Entities (11)

PPCR-IM · system · 98%Zichen Song · person · 95%Directed Acyclic Graph · data-structure · 92%Lanzhou University · organization · 90%LLM · technology · 90%Expected-indicator coverage score · metric · 88%Model–government focus ratio · metric · 88%Overlooked-indicator discovery rate · metric · 88%Doubao · model · 85%GPT-5.1 · model · 85%World Bank · organization · 80%

Relation Signals (10)

Zichen Song → affiliatedwith → Lanzhou University

confidence 95% · Zichen Song*, Weijia Li Lanzhou University

PPCR-IM → developedby → Zichen Song

confidence 95% · Developed by Zichen Song

PPCR-IM → uses → Directed Acyclic Graph

confidence 95% · PPCR-IM uses an LLM-driven, layer-wise generator to construct a directed acyclic graph of intermediate consequences

PPCR-IM → uses → LLM

confidence 95% · PPCR-IM uses an LLM-driven, layer-wise generator

PPCR-IM → evaluatesusing → Expected-indicator coverage score

confidence 90% · For each policy episode with reference annotations, we compute an expected-indicator coverage score

PPCR-IM → evaluatesusing → Overlooked-indicator discovery rate

confidence 90% · An overlooked-indicator discovery rate measures how often PPCR-IM surfaces indicators

PPCR-IM → evaluatesusing → Model–government focus ratio

confidence 90% · A model–government focus ratio compares the coverage achieved by the system and by the government

PPCR-IM → outperforms → Doubao

confidence 90% · Compared to two LLM-based baselines... PPCR-IM achieves the strongest performance on all three metrics.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Public policy decisions are typically justified using a narrow set of headline indicators, leaving many downstream social impacts unstructured and difficult to compare across policies. We propose PPCR-IM, a system for multi-layer DAG-based consequence reasoning and social indicator mapping that addresses this gap. Given a policy description and its context, PPCR-IM uses an LLM-driven, layer-wise generator to construct a directed acyclic graph of intermediate consequences, allowing child nodes to have multiple parents to capture joint influences. A mapping module then aligns these nodes to a fixed indicator set and assigns one of three qualitative impact directions: increase, decrease, or ambiguous change. For each policy episode, the system outputs a structured record containing the DAG, indicator mappings, and three evaluation measures: an expected-indicator coverage score, a discovery rate for overlooked but relevant indicators, and a relative focus ratio comparing the systems coverage to that of the government. PPCR-IM is available both as an online demo and as a configurable XLSX-to-JSON batch pipeline.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

30,118 characters extracted from source content.

Expand or collapse full text

PPCR-IM: A System for Multi-layer DAG-based Public Policy Consequence Reasoning and Social Indicator Mapping Zichen Song*, Weijia Li Lanzhou University songzch21@lzu.edu.cn, forgetfuljre@gmail.com Abstract Public policy decisions are typically justified using a narrow set of headline indicators, leav- ing many downstream social impacts unstruc- tured and difficult to compare across policies. We propose PPCR-IM, a system for multi- layer DAG-based consequence reasoning and social indicator mapping that addresses this gap. Given a policy description and its con- text, PPCR-IM uses an LLM-driven, layer- wise generator to construct a directed acyclic graph of intermediate consequences, allowing child nodes to have multiple parents to cap- ture joint influences. A mapping module then aligns these nodes to a fixed indicator set and assigns one of three qualitative impact direc- tions: increase, decrease, or ambiguous change. For each policy episode, the system outputs a structured record containing the DAG, in- dicator mappings, and three evaluation mea- sures: an expected-indicator coverage score, a discovery rate for overlooked but relevant in- dicators, and a relative focus ratio comparing the system’s coverage to that of the govern- ment. PPCR-IM is available both as an online demo and as a configurable XLSX-to-JSON batch pipeline, and is publicly accessible at https://n23pg3gbm5.coze.site. 1 Introduction Public policy decisions are typically justified and evaluated using a small set of headline indicators, such as GDP growth, inflation, or unemployment. A large literature on “Beyond GDP” and multidi- mensional well-being has shown that these met- rics capture only a narrow slice of social outcomes and systematically miss distributional, environmen- tal, and institutional dimensions. 1 In practice, the links between specific policy choices, intermediate mechanisms, and downstream indicators remain 1 See, e.g., overviews of multidimensional well-being and “Beyond GDP” indicator frameworks in economics and public policy. mostly implicit in narrative reports or ad hoc dia- grams, which makes consequence reasoning hard to compare, audit, or reuse across policies. A sys- tem that can externalize these reasoning chains in a structured form, and align them with a shared indicator space, has the potential to make policy analysis more transparent and to surface impacts that fall outside the official focus of government dashboards(Huang and Yang, 2023). Day-to-day policy work often centers on a few “primary” indicators that are declared ex ante in strategy or budget documents, while many second- order and long-term effects receive less system- atic attention. Existing workflows provide little support for: (i) enumerating intermediate conse- quences beyond a short narrative; (i) linking those consequences consistently to a fixed indicator tax- onomy; and (i) quantifying where government focus diverges from a broader set of plausible im- pacts. Theory-of-change diagrams and logic mod- els, when available, are typically hand-crafted for a single program and not aligned with reusable indi- cator sets. PPCR-IM is designed to address this gap by turning qualitative consequence reasoning into a machine-readable object that explicitly connects policies, consequence chains, and indicators. In this paper, we introduce PPCR-IM (Public Policy Consequence Reasoning and Indicator Map- ping), an end-to-end system that turns qualitative policy consequence analysis into multi-layer DAG construction with indicator alignment. Given a policy description and context, PPCR-IM uses an LLM-driven, layer-wise generator to build a di- rected acyclic graph of intermediate consequences, and a mapping module to align these nodes to a fixed set of social indicators with qualitative im- pact directions (increase, decrease, or ambiguous change). For each policy episode, the system produces a structured, machine-readable record that links the policy text, the generated DAG, and indicator-level impact assessments. 1 arXiv:2602.21650v1 [cs.SI] 25 Feb 2026 Policy Description: For The Tax credits and subsidies to accelerate solar and wind adoption. Renewable capacity ↑ Lower greenhouse emissions Feature 2 Feature 1 Feature 3 Policy Type Jurisdiction Year L 12 Policy Root Input Policy Energy Security Carbon Emission Green Jobs Cleaning manufact ↑ The GDP Number ↑ Example of Model Process Creation of green jobs Reduced air pollution Grid stability riskss Economic transition challenge Enhanced energy security Long-term climate mitigation Indicator Vocabulary L 22 L 13 L 21 L 11 L 24 Figure 1: A policy description and contextual attributes are used to generate a multi-layer consequence DAG that models causal impact pathways. The system maps DAG nodes to a fixed indicator vocabulary with directional effects and supporting evidence links, enabling structured impact assessment and downstream evaluation. Our contributions are summarized as follows: •We formulate public policy consequence rea- soning as the construction of a multi-layer DAG with multi-parent dependencies, cou- pled with a mapping from intermediate conse- quences to a shared indicator space. • We implement a practical system that outputs, for each policy episode, a structured record containing the DAG, indicator mappings, and evaluation measures that capture coverage of government-expected indicators, systematic discovery of overlooked but relevant indica- tors, and the relative focus of the system ver- sus the government. •We provide both an interactive demo and a configurable XLSX-to-JSON batch pipeline, making PPCR-IM usable by policy analysts, public administration researchers, and NLP practitioners within existing evaluation work- flows. 2 Task Formulation and Evaluation 2.1 Inputs and outputs PPCR-IM processes individual policy episodes. Each episode is represented by a short policy de- scription, optional contextual fields (e.g., jurisdic- tion, time period, macro conditions), and a set of so- cial indicators that the government explicitly high- lights in accompanying documents or expert recon- structions of those documents. In the batch setting, these fields are stored row-wise in an XLSX file so that policy analysts can prepare and edit inputs using familiar spreadsheet tools. For each episode, the system returns a structured JSON record. This record contains: a multi-layer directed acyclic graph (DAG) of intermediate con- sequences; a mapping from graph nodes to a fixed indicator vocabulary, with a qualitative impact di- rection (increase, decrease, or ambiguous change) for each indicator; and a collection of scalar eval- uation measures. These outputs are used both for interactive visualization in the demo and for down- stream quantitative analysis across many policies. 2.2 DAG reasoning and indicator mapping We cast qualitative consequence reasoning as a two-step task. First, given a policy and its context, an LLM-driven, layer-wise generator constructs a DAG in which the root corresponds to the policy and subsequent layers capture downstream conse- quences. Nodes are short textual statements of in- termediate effects, and edges encode hypothesized influence relations; child nodes may have multiple parents to represent joint influences, while depth and branching are controlled by user parameters to keep graphs interpretable. Second, PPCR-IM aligns this DAG to a fixed set of social indicators. For each indicator, the system decides whether it is plausibly affected by the policy, identifies the subset of relevant consequence nodes, and assigns a qualitative direction of change. This separation between graph construction and indicator mapping 2 Policy Consequence DAG Evaluation System Policy impact prediction and evaluation tool based on DAG structure, supporting single policy and batch dataset evaluation Developed by Zichen Song (Affiliations: Lanzhou University / Sungkyunkwan University / City University of Hong Kong) Single EvaluationBatch Evaluation Policy Information Enter the policy information to evaluate Policy Name* e.g., Healthcare Reform Act Policy Description* Describe the policy details and context... Country e.g., United States Year e.g., 2024 Policy Type Select policy type DeepSeek API Key* Enter your DeepSeek API Key Start Evaluation Evaluation Results Policy Consequence DAG Evaluation Metrics Please enter policy information and click "Start Evaluation" 2026/2/20 15:45New App | Coze Code https://n23pg3gbm5.coze.site1/1 Figure 2: Web-based interface of the Policy Consequence DAG Evaluation System. Users provide policy descriptions and contextual information to trigger DAG-based consequence evaluation and indicator-level impact metrics. is a deliberate design choice: it allows reuse of the same indicator space across policies, supports comparison of impact patterns at the indicator level, and distinguishes PPCR-IM from approaches that only attach free-form explanations or unstructured chains of thought to individual indicators. 2.3 Evaluation objectives Our evaluation focuses on how well PPCR-IM replicates and extends the indicator focus observed in government documents, rather than on intrin- sic language modeling metrics or user satisfaction scores. For each policy episode with reference annotations, we compute an expected-indicator coverage score that measures the proportion of indicators in the government focus set that the system also flags as affected by the policy. This assesses whether the system recovers the dimen- sions that policymakers themselves emphasize. We then assess two complementary dimensions. An overlooked-indicator discovery rate measures how often PPCR-IM surfaces indicators that experts deem relevant but that were not highlighted by the government, capturing the system’s ability to broaden the conversation. A model–government focus ratio compares the coverage achieved by the system and by the government on a common indica- tor subset, providing a normalized view of how con- centrated or diffuse their respective attention is. In this work we rely on such offline, annotation-based measures; user studies and interactive evaluations are left for future work once the basic behavior of the system is well understood. 3 System Architecture and Methods 3.1 Overall architecture and data flow Figure 1 summarizes the PPCR-IM architecture. The system is organized into three main modules connected by a simple data flow: a DAG genera- tor, an indicator mapper, and an evaluation and export layer. The DAG generator takes as input a policy episode record (policy text, context, and government focus indicators) and produces a multi- layer directed acyclic graph of intermediate con- sequences. The indicator mapper consumes this graph and produces indicator-level impact assess- ments. The evaluation and export layer then com- pares these assessments against reference anno- tations and writes a structured JSON record per policy episode, which is used both in the interac- tive demo and in the batch pipeline(Johnson et al., 2022). The overall design is intentionally modular. All communication between modules happens through explicit, machine-readable objects: the DAG gen- erator exposes a node–edge representation with layer labels; the indicator mapper exposes a list of indicator impact entries; and the evaluation layer exposes scalar summary measures. This separation allows each module to be improved or replaced in- dependently (for example, swapping in a different LLM backend for DAG generation, or a different indicator taxonomy), while preserving a stable in- terface for policy analysts. Methodologically, the architecture combines three elements that are rarely integrated in existing tools: layer-wise graph-based reasoning over policy consequences, alignment to 3 a fixed indicator space rather than ad hoc labels, and a dedicated evaluation module that quantifies how the system’s focus compares to institutional focus across many policies. 3.2 Layer-wise DAG construction The DAG generator constructs a consequence graph in a breadth-wise, layer-by-layer fashion. Starting from a root node that represents the pol- icy itself, PPCR-IM prompts the LLM to propose downstream consequences for the current frontier of nodes, subject to user-configurable limits on maximum depth and branching factor. Each node is represented by a concise textual description and a layer index; edges are added from parents to chil- dren as the graph expands. When multiple parents propose semantically similar children, the system merges these into a single node, adds edges from all relevant parents, and reuses this node in subsequent layers. Graph expansion stops when the maximum depth is reached or when the LLM no longer pro- poses new, distinct consequences for a given layer. Allowing multi-parent dependencies is essential for representing realistic policy mechanisms, where many outcomes arise from the joint effect of several upstream changes. Unlike linear chain-of-thought style reasoning, which forces a single path through consequences, the DAG formulation encourages the LLM to articulate branching and converging pathways (e.g., fiscal measures and labor-market re- sponses jointly shaping household financial stress). At the same time, depth and branching limits ensure that the resulting graphs remain compact enough for human inspection in the demo interface. This combination of constrained, layer-wise expansion and explicit multi-parent structure distinguishes PPCR-IM from systems that only attach unstruc- tured textual rationales to model outputs. 3.3 Indicator mapping and prediction Once a DAG has been constructed, PPCR-IM maps it into a fixed indicator space. The indicator map- per considers the full set of consequence nodes and a predefined vocabulary of social indicators, each with a short textual definition. For each indica- tor, the system queries the LLM with a structured prompt containing the policy, relevant nodes, and the indicator description, and asks whether the in- dicator is plausibly affected by the policy. If so, the mapper selects a subset of linked nodes and assigns a qualitative direction of change from the set in- crease, decrease, ambiguous. The result is a list of indicator impact entries, each tying one indicator to specific parts of the DAG and to a direction label. This procedure serves two purposes. First, it provides a consistent way of situating heteroge- neous consequence graphs in a shared indicator vocabulary, which is crucial for comparing patterns across policies and aggregating statistics over many episodes. Second, by explicitly annotating direc- tions of change and recording which nodes support each decision, it makes it possible to define evalua- tion measures that distinguish between recovering government-expected indicators and discovering additional, relevant indicators that were not in the government’s declared focus(Shi et al., 2024). 4 Experiment and Results 4.1Runtime configuration and batch pipeline PPCR-IM is implemented as a set of Python modules with a command-line interface aimed at policy analysis teams and NLP practitioners. The main entry point takes as arguments the in- put XLSX file, the output directory, and run- time parameters controlling the LLM backend and graph generation process. Key options in- clude the model identifier (–model-name), the gen- eration temperature for consequence expansion (–temperature), a separate temperature for indica- tor linking (–link-temperature), structural con- straints on the graph (–max-depth,–max-branch, –max-links-per-node), and API configuration (–deepseek_api). In a batch run, each row of the XLSX file is treated as one policy episode, passed through the DAG generator and indicator mapper, and written out as a JSON file keyed by policy identifier(Jiang et al., 2023). For every episode, the system records a status flag indicating whether the pipeline completed suc- cessfully (ok), was deliberately skipped (skipped, e.g., due to missing mandatory fields), or encoun- tered an error (error, e.g., repeated API failures or unparseable LLM output). This status, together with a short diagnostic message, is stored in the JSON alongside the DAG, indicator mappings, and evaluation measures. At the end of a run, PPCR- IM prints aggregate counts for each status type and summary statistics for the evaluation metrics, sup- porting both one-off experiments and repeated runs on updated datasets without modifying core code. 4 SystemMetricMeanStd. dev.MinMax GPT 5.1 Expected-indicator coverage score0.8510.0520.680.94 Overlooked-indicator discovery rate0.3520.0880.150.58 Model–government focus ratio1.0980.1150.861.41 Doubao Expected-indicator coverage score0.8030.0610.620.91 Overlooked-indicator discovery rate0.2910.0790.120.51 Model–government focus ratio1.0230.1080.811.33 PPCR-IM Expected-indicator coverage score0.9020.0480.710.96 Overlooked-indicator discovery rate0.6030.0920.180.67 Model–government focus ratio1.3560.1240.891.52 Table 1: Comparison of PPCR-IM with two LLM-based baselines on 1,027 policy episodes. The coverage score measures the proportion of government-expected indicators recovered by each system. The discovery rate quantifies the proportion of relevant but previously unfocused indicators identified. 4.2 Dataset and metric computation To evaluate PPCR-IM in a realistic setting, we con- struct a dataset of 1,027 policy episodes. We begin from a set of nineteen social and macroeconomic indicators inspired by World Bank data, covering domains such as growth, employment, debt, and social spending, and use these as the fixed indica- tor vocabulary(Kovac et al., 2023). Each episode consists of a short policy description, basic con- text fields, and two reference sets: a government focus set, reflecting indicators emphasized in offi- cial communication, and a relevance set, reflecting indicators that expert annotators judge to be ma- terially affected regardless of whether they were highlighted. The episodes span several regions; as shown in Figure 3, approximately 41% of cases come from the United States, 34% from Japan, 23% from European countries, and the remaining 2% from other regions(Kwon et al., 2023). Given the JSON outputs produced by PPCR-IM on this corpus, we compute three annotation-based evaluation measures per episode: an expected- indicator coverage score (the proportion of indi- cators in the government focus set recovered by the system), an overlooked-indicator discovery rate (the proportion of relevant indicators outside the focus set that are identified), and a model– government focus ratio (comparing system and government coverage on a shared indicator sub- set). These quantities are then aggregated across episodes to produce corpus-level statistics(Li et al., 2024). 4.3 Results, case studies, and limitations Table 1 summarizes aggregate results on the 1,027 policy episodes. Compared to two LLM-based baselines that predict impacted indicators directly from the policy text (GPT 5.1 and Doubao), PPCR- IM achieves the strongest performance on all three metrics. Its expected-indicator coverage score reaches a mean of 0.902 (std. 0.048), versus 0.851 and 0.803 for GPT 5.1 and Doubao, respectively. The overlooked-indicator discovery rate improves more markedly, with PPCR-IM reaching 0.603 (std. 0.092), compared to 0.352 and 0.291 for the base- lines. The model–government focus ratio also in- creases from 1.098 and 1.023 to 1.356, indicating that PPCR-IM systematically broadens indicator coverage relative to what is emphasized in gov- ernment focus sets while still recovering most ex- pected indicators(Myung et al., 2024). 41% 34% 23% 2% Regional distribution of policy episodes Region USA Japan Europe Other Figure 3: Regional distribution of the 1,027 policy episodes used in our experiments. Qualitative inspection of individual episodes helps explain these gains. For fiscal transfer and tax policies, PPCR-IM not only recovers the in- tended macroeconomic indicators (such as growth and unemployment) but also highlights debt-related and distributional indicators that appear in the ex- pert relevance sets but not in the government focus sets. The explicit DAG representation makes it pos- sible to trace how such additional indicators arise through intermediate mechanisms, which the direct- prediction baselines cannot provide. At the same time, limitations remain: in some domains the gen- 5 erated consequence nodes are generic or weakly justified, and the indicator mapper sometimes as- signs ambiguous directions when textual cues are scarce. These observations reinforce our position- ing of PPCR-IM as a decision-support tool that produces auditable, structured hypotheses, rather than a system that can replace domain expertise in interpreting policy impacts(Pawar et al., 2025). 5 Demo Availability, Reproducibility 5.1 Online demo and interaction modes PPCR-IM is exposed through a web-based demo targeted at three main user groups: policy ana- lysts and public administration researchers who want structured support for consequence reason- ing; NLP and AI practitioners who want to study graph-based uses of LLMs; and students or ed- ucators who require concrete examples of auto- mated policy-impact analysis. The demo front-end provides a simple form where users can paste a policy description, optionally specify basic con- text (jurisdiction, year, policy type), and select a configuration profile for the underlying model and graph-generation parameters(Shen et al., 2024). After submission, the system runs the full pipeline and returns a multi-pane view.One pane displays the generated multi-layer DAG, with nodes grouped by depth and edges indicating hy- pothesized influence relations. A second pane lists the indicator-level assessments: for each of the nineteen indicators, the demo shows whether it is predicted to be affected, the qualitative direction (increase, decrease, ambiguous), and the support- ing consequence nodes. A third pane summarizes the evaluation measures for that episode when ref- erence annotations are available. Users can expand or collapse layers in the graph, inspect which nodes are linked to which indicators, and download the full JSON record for offline analysis. 5.2 Code and reproduction We release the PPCR-IM implementation, together with configuration files and documentation, un- der an open-source Apache 2.0 license. The code, preprocessed dataset, and example configuration scripts are documented and linked via our project page athttps://n23pg3gbm5.coze.site. The license allows non-commercial and commercial re- search use, adaptation, and redistribution, subject to attribution and preservation of the license terms. In practice, this means that research groups can integrate PPCR-IM into their own policy-analysis pipelines, extend the indicator vocabulary, or swap in different LLM backends, while public-sector teams can experiment with the system as part of internal analytic toolchains(Rao et al., 2024). To reproduce the experiments in this paper, users can follow a three-step procedure: (i) install the Python environment and obtain access credentials for the chosen LLM backend; (i) download the curated dataset of 1,027 policy episodes with anno- tations over nineteen World Bank–style indicators and place the XLSX file in the expected directory; and (i) run the provided batch script with the same configuration used in our experiments, which gen- erates JSON outputs and metric summaries. The repository also includes helper scripts to re-create Table 1 and Figure 3, and to run PPCR-IM side by side with the GPT 5.1 and Doubao baselines on new policy episodes(Ren et al., 2024). 6 Conclusion In this paper, we introduced PPCR-IM, a system that turns qualitative public policy consequence analysis into multi-layer DAG construction with indicator alignment over a fixed vocabulary of nine- teen World Bank–style social and macroeconomic indicators. On a curated dataset of 1,027 policy episodes with government focus sets and expert relevance annotations, PPCR-IM achieves a mean expected-indicator coverage score of 0.902 (std. 0.048), an overlooked-indicator discovery rate of 0.603 (std. 0.092), and a model–government focus ratio of 1.356, outperforming two strong LLM base- lines (GPT 5.1 and Doubao) on all three metrics. These results indicate that making intermediate con- sequence structures explicit, rather than predicting indicators directly from text, helps recover most indicators emphasized in official documents while systematically surfacing additional relevant dimen- sions. Looking ahead, we plan to enrich the graph semantics with causal and uncertainty annotations, link DAGs to numerical simulation or historical outcome data, broaden the geographic and policy coverage of the corpus, and conduct user studies or expert workshops to better understand how PPCR- IM fits into real policy-analysis workflows. 7 Ethics and Broder Impact Statement PPCR-IM is a decision-support system for structur- ing public-policy consequence reasoning; its out- puts are hypothesis-level guidance, not causal pre- 6 dictions. It can improve transparency by externaliz- ing multi-layer DAG pathways and mapping them to a shared indicator set, but LLM-generated nodes and qualitative directions may be biased, culturally contingent, or weakly supported and could mislead if treated as forecasts. We recommend human over- sight, validation with evidence and stakeholders, and explicit uncertainty disclosure. The system re- quires no personal data; users must ensure inputs comply with privacy and governance rules. 8 Limitations PPCR-IM relies on LLM-generated consequence DAGs, so nodes and links may be generic, incom- plete, or biased. Indicator impacts are qualitative (increase/decrease/ambiguous) rather than quantita- tive, and directions can be uncertain when evidence is limited. The DAG represents plausible mecha- nisms but does not establish causal validity; outputs should be treated as structured hypotheses requir- ing expert verification. Finally, evaluation depends on government focus sets and expert annotations, which may reflect subjective or institutional priori- ties. References Jing Huang and Diyi Yang. 2023. Culturally aware natural language inference. In Findings of the As- sociation for Computational Linguistics: EMNLP 2023, pages 7591–7609, Singapore. Association for Computational Linguistics. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825. Rebecca L. Johnson, Giada Pistilli, Natalia Menéndez- González, Leslye Denisse Dias Duran, Enrico Panai, Julija Kalpokiene, and Donald Jay Bertulfo. 2022. The ghost in the machine has an american ac- cent: Value conflict in GPT-3. arXiv preprint arXiv:2203.07785. Grgur Kovac, Masataka Sawayama, Rémy Portelas, Cé- dric Colas, Peter Ford Dominey, and Pierre-Yves Oudeyer. 2023. Large language models as super- positions of cultural perspectives. arXiv preprint arXiv:2307.07870. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gon- zalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serv- ing with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles (SOSP 2023). ACM. Huihan Li, Liwei Jiang, Jena D. Hwang, Hyunwoo Kim, Sebastin Santy, Taylor Sorensen, Bill Yuchen Lin, Nouha Dziri, Xiang Ren, and Yejin Choi. 2024. CULTURE-GEN: Revealing global cultural percep- tion in language models through natural language prompting. arXiv preprint arXiv:2404.10199. Junho Myung, Nayeon Lee, Yi Zhou, Jiho Jin, Rifki Afina Putri, Dimosthenis Antypas, Hsuvas Borkakoty, Eunsu Kim, Carla Perez-Almendros, Abinew Ali Ayele, Víctor Gutiérrez-Basulto, Yazmín Ibáñez-García, Hwaran Lee, Shamsuddeen Hassan Muhammad, Kiwoong Park, Anar Sabuhi Rzayev, Nina White, Seid Muhie Yimam, Mohammad Taher Pilehvar, and 3 others. 2024. BLEnD: A benchmark for LLMs on everyday knowledge in diverse cultures and languages. arXiv preprint arXiv:2406.09948. Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Au- genstein. 2025. Survey of cultural awareness in lan- guage models: Text and beyond. Computational Linguistics, 51(3):907–1004. Abhinav Rao, Akhila Yerukola, Vishwa Shah, Katharina Reinecke, and Maarten Sap. 2024. Normad: A bench- mark for measuring the cultural adaptability of large language models. arXiv preprint arXiv:2404.12464. Yuanyi Ren, Haoran Ye, Hanjun Fang, Xin Zhang, and Guojie Song. 2024. Valuebench: Towards compre- hensively evaluating value orientations and under- standing of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2015–2040, Bangkok, Thailand. Association for Computational Linguistics. Siqi Shen, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, Soujanya Poria, and Rada Mihalcea. 2024. Understanding the capabilities and limitations of large language models for cultural commonsense. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5668–5680, Mexico City, Mexico. Association for Computational Lin- guistics. Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Sunny Yu, Raya Horesh, Rogério Abreu De Paula, and Diyi Yang. 2024.Culturebank: An online community-driven knowledge base towards cultur- ally aware language technologies. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4996–5025, Miami, Florida, USA. Asso- ciation for Computational Linguistics. 7

Instant research discovery