🌳 CultureForest Benchmark

Understanding and Evaluating Cultural Norm Grounded Reasoning in Large Language Models

---

✨ Overview

CultureForest is a benchmark designed to evaluate whether large language models (LLMs) can effectively utilize cultural knowledge to reason about contextually appropriate behavior.

While prior work primarily focuses on whether models know cultural norms, CultureForest shifts the focus to whether models can apply such knowledge under realistic, context-dependent scenarios.

---

🎯 Key Insight

LLMs often possess cultural knowledge, but struggle to use it for reasoning in open-ended settings.

CultureForest is designed to expose this gap by grounding each question in explicit cultural norms and evaluating models under progressively reduced guidance.

---

📊 Benchmark Statistics

• Total Samples: 5,378
• Domains: 8 - Business Culture, Communication, Dates of Significance, Etiquette, Family, Greetings, Naming, Religion
• Countries/Regions: 53
• Geographic Coverage: Asia, Europe, Middle East, Africa, Oceania, The Americas

Each sample is provided in three parallel formats:

• Easy: Multiple-choice
• Medium: Binary classification (per-option judgment)
• Hard: Open-ended generation

---

🧠 Task Definition

Each question is grounded in 3 atomic cultural norms:

This enables:

• ✅ Verifiable evaluation (objective correctness)
• ✅ Attributable analysis (trace errors to specific norms)

---

📈 Evaluation Protocol

Difficulty Levels

Level	Setting	Description
Easy	Multiple-choice	Select correct option
Medium	Binary classification	Judge each option (True/False)
Hard	Open-ended generation	Generate culturally appropriate response

---

Open-ended Evaluation (C-Verifier)

We train a lightweight verifier based on Qwen3.5-0.8B-Base to assess norm compliance.

Given an input pair (response, norm), C-Verifier predicts one of three labels:

• Satisfy — the response complies with the norm
• Neutral — the response is non-committal or weakly related
• Violate — the response contradicts the norm

This formulation reduces open-ended evaluation to NLI-style, enabling fine-grained and interpretable analysis.

Final scores are computed via expected utility over norm compliance.

---

Data Format

{
    "uuid": "Afghanistan***Communication***0", # unique id
    "Norm_Bundle": { # meta infos & 3 atomic cultural norms
        "Geolocation": "Asia", 
        "Country/Region": "Afghanistan",
        "Category": "Communication",
        "Topic": "Non-Verbal",
        "Norms": {
            "Norm_1": {
                "norm_id": "1465",
                "norm": "Hands: There is a separation between the functions ..."},
            "Norm_2": {
                "norm_id": "1467",
                "norm": "Physical Contact: It is inappropriate to be physically ..."},
            "Norm_3": {
                "norm_id": "1471",
                "norm": "Winking: Winking at a member of the opposite gender is ..."}
            }
        },
    "Context": "During a multi-generational family reunion in Kabul, several relatives ...", # background infos
    "Options": [ # options before shuffled
        {"id": "A", "text": ...},
        {"id": "B", "text": ...},
        {"id": "C", "text": ...},
        {"id": "D", "text": ...}
    ],
    "Easy": { # Easy-mode: Multiple Choice Question
        "Question_stem": ...,      # question stem for multiple Choice Question
        "Options_Shuffle": ...,    # shuffled options
        "Answer_Key_Shuffle": "C", # answer key for the 
        "Question_MCQ": ...,       # context + question_stem (easy) + options
    },
    "Medium": { # Medium: Binary Judgement
        "Question_Judge1": ...,   # context + question stem (easy) + option-A + "Is this behavior culturally appropriate?"
        "Question_Judge2": ...,   # context + question stem (easy) + option-B + "Is this behavior culturally appropriate?"
        "Question_Judge3": ...,   # context + question stem (easy) + option-C + "Is this behavior culturally appropriate?"
        "Question_Judge4": ...    # context + question stem (easy) + option-D + "Is this behavior culturally appropriate?"
    },
    "Open": { # Hard: Open-ended Generation
        "Question_stem": ...,     # question stem for Open-ended Generation
        "Question_Open": ...,     # context + question_stem (open)
    }
}