Trojsten Benchmark

Evaluating LLM Problem-Solving in Slovak STEM Competition Problems

• Paper: EMNLP 2025
• Authors: Adam Zahradník, Marek Šuppa
• Repository: github.com/gardenerik/trojsten-benchmark
• Contact: adam@zahradnik.xyz

Abstract

Large language models show promising performance on reasoning tasks, yet evaluation methods for low-resource languages remain limited, particularly for complex STEM problem-solving. We introduce Trojsten Benchmark, a Slovak-language dataset of 1,108 high-school competition problems with reference solutions across mathematics, physics, and programming, and a rubric-based LLM grading framework. Using GPT-4 to generate rubrics and grade solutions, we observe 1.05 average absolute deviation from human graders (5-point scale), while benchmarking GPT-3.5-Turbo, GPT-4, GPT-4o, and open-weight models (Llama 3, Phi-3). We quantify multistep reasoning performance by difficulty, show consistent underperformance on harder items, and demonstrate language sensitivity: accuracy drops on English translations of Slovak statements, evidencing challenges beyond translation. Trojsten Benchmark complements English-centric math datasets (e.g., MATH, GSM8K) by targeting open-response, rubric-gradable reasoning under low-resource linguistic framing. We release code and data to enable reproducible evaluation and human-aligned auto-grading for STEM in under-served languages.

Basic Usage

Generate LLM Solutions:

python generate_solutions.py [list] [prefix] [model] [solver]

# Example:
python generate_solutions.py fks_test experiment1 gpt4 zeroshot_cot

Generate Evaluation Rubrics:

python generate_rubrics.py [list] [prefix] [model]

# Example:
python generate_rubrics.py fks_test experiment1 gpt4

Grade Generated Solutions:

python grade_solutions.py [list] [prefix] [grader] [rubricer]

# Example:
python grade_solutions.py fks_test experiment1 zeroshot_v1 zeroshot_v1

Available Solvers:

• zeroshot_v1: Basic zero-shot approach
• zeroshot_cot: Zero-shot with chain-of-thought reasoning
• generated_knowledge: Uses generated knowledge for problem solving
• dualprompt_generated_knowledge: Dual prompt generated knowledge
• least_to_most: Least-to-most problem decomposition

Models Supported:

• OpenAI GPT models (gpt4, gpt3.5-turbo)
• Azure OpenAI deployments
• Custom model configurations

Data Format

Each problem in the dataset is stored in a CSV with the following columns:

id,competition,problem,solution,difficulty

Example:

1b56d285f84b378c,fks,"Sme na vysokom strome...","Zamyslime sa nad tym...",3

Fields:

• id: Unique problem identifier
• competition: Competition type (fks, kms, ksp)
• problem: Problem statement in Slovak
• solution: Detailed solution explanation
• difficulty: Numeric difficulty rating

Citation

@inproceedings{zahradnik-suppa-2025-trojsten,
    title = "Trojsten Benchmark: Evaluating {LLM} Problem-Solving in {S}lovak {STEM} Competition Problems",
    author = "Zahradn{\'i}k, Adam  and
      Suppa, Marek",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1779/",
    pages = "35094--35109",
    ISBN = "979-8-89176-332-6",
    abstract = "Large language models show promising performance on reasoning tasks, yet evaluation methods for low-resource languages remain limited, particularly for complex STEM problem-solving. We introduce Trojsten Benchmark, a Slovak-language dataset of 1,108 high-school competition problems with reference solutions across mathematics, physics, and programming, and a rubric-based LLM grading framework. Using GPT-4 to generate rubrics and grade solutions, we observe 1.05 average absolute deviation from human graders (5-point scale), while benchmarking GPT-3.5-Turbo, GPT-4, GPT-4o, and open-weight models (Llama 3, Phi-3). We quantify multistep reasoning performance by difficulty, show consistent underperformance on harder items, and demonstrate language sensitivity: accuracy drops on English translations of Slovak statements, evidencing challenges beyond translation. Trojsten Benchmark complements English-centric math datasets (e.g., MATH, GSM8K) by targeting open-response, rubric-gradable reasoning under low-resource linguistic framing. We release code and data to enable reproducible evaluation and human-aligned auto-grading for STEM in under-served languages."
}

License

This project is part of academic research. Please contact the authors for usage information.