BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models

101+ Models Evaluated | 79 Reasoning Tasks | 138 Variations | >10^15 Unique Instances

Explore Leaderboard | Read Paper | GitHub | Documentation

---

What is BeyondBench?

BeyondBench introduces a revolutionary approach to evaluating reasoning capabilities in language models without relying on traditional static benchmarks. Our system dynamically generates novel problems across 79 distinct reasoning tasks with 138 variations, ensuring that models cannot memorize solutions and must demonstrate true reasoning abilities.

Key Features

• Dynamic Problem Generation — Problem space >10^15 unique instances, zero risk of data contamination
• Three Difficulty Levels — Easy (44 tasks), Medium (15 tasks, 49 variations), Hard (20 tasks, 68 variations)
• Multi-Backend Support — OpenAI, Gemini, Anthropic APIs + vLLM and HuggingFace Transformers
• Contamination-Resistant — No static benchmark memorization, novel problems every run
• Comprehensive Metrics — Accuracy, instruction-following compliance, token efficiency
• 101+ Models Evaluated — Open-source and proprietary, regularly updated

---

Installation

From PyPI

pip install beyondbench

With Optional Dependencies

# All API clients (OpenAI, Gemini, Anthropic)
pip install beyondbench[all-apis]

# vLLM support (requires CUDA)
pip install beyondbench[vllm]

# Everything (all APIs + vLLM + dev tools + visualization)
pip install beyondbench[full]

From Source

git clone https://github.com/ctrl-gaurav/BeyondBench.git
cd BeyondBench
pip install -e .

---

Quick Start

Interactive Wizard

beyondbench

Command Line

# Evaluate GPT-4o on the easy suite
beyondbench evaluate --model-id gpt-4o --api-provider openai --suite easy

# Evaluate a local model with vLLM
beyondbench evaluate --model-id meta-llama/Llama-3.2-3B-Instruct --backend vllm --suite all

# Evaluate Claude on hard tasks
beyondbench evaluate --model-id claude-sonnet-4-20250514 --api-provider anthropic --suite hard

# List all available tasks
beyondbench list-tasks

Python API

from beyondbench import EvaluationEngine, ModelHandler, TaskRegistry

# Initialize model handler
model = ModelHandler(
    model_id="gpt-4o",
    api_provider="openai",
    api_key="your-api-key"
)

# Run evaluation
engine = EvaluationEngine(model_handler=model, output_dir="./results")
results = engine.run_evaluation(suite="easy", datapoints=100)

# Print results
print(f"Average Accuracy: {results['summary']['avg_accuracy']:.2%}")

---

New in v0.2.1 (Apr 17, 2026)

Critical packaging fix — four subpackages (parsers.strategies, configs, eval, prompts) were missing from the v0.2.0 PyPI wheel, causing ImportError on fresh installs. No API changes.

New in v0.2.0 (Apr 16, 2026)

• 79 Reasoning Tasks: 44 easy + 15 medium (49 variations) + 20 hard (68 variations)
• Multi-GPU Parallel Evaluation: Automatic batch auto-tuning and tensor parallelism
• Plugin SDK: Create and share custom tasks with beyondbench plugin scaffold
• Gradio Dashboard: Real-time evaluation monitoring with --dashboard flag
• Response Caching: Skip redundant API calls across runs
• Universal Parser: Unified parsing engine with confidence scoring
• 1000+ Tests: Comprehensive unit, integration, and end-to-end test coverage
• API Server: beyondbench serve - FastAPI REST API with WebSocket support

---

Supported Backends

Backend	Models	Features
OpenAI	GPT-4o, GPT-4o-mini, GPT-5, GPT-5-mini	Reasoning effort control
Gemini	Gemini 2.5 Pro, Gemini 2.5 Flash	Thinking budget configuration
Anthropic	Claude Sonnet 4, Claude Opus 4	Latest Claude models
vLLM	Any HuggingFace model	Batch processing, tensor parallelism
Transformers	Any HuggingFace model	CPU/GPU inference

---

Task Suites

Easy Suite (44 Tasks)

Arithmetic (sum, multiplication, subtraction, division, absolute_difference), Statistics (mean, median, mode), Counting (odd_count, even_count, count_negative, count_unique, and more), Extrema (find_maximum, find_minimum, second_maximum, range, and more), Sequences (sorting, longest_increasing_subsequence, alternating_sum, sum_of_digits), Comparison

Medium Suite (15 Tasks, 49 Variations)

Fibonacci Sequence (6 variations), Algebraic Sequence (10), Geometric Sequence (10), Prime Sequence (11), Complex Pattern (12)

Hard Suite (20 Tasks, 68 Variations)

Tower of Hanoi, N-Queens, Graph Coloring, Boolean SAT, Sudoku, Cryptarithmetic, Matrix Chain, Modular Systems, Constraint Optimization, Logic Grid Puzzles

---

Leaderboard (Top 5)

Rank	Model	Overall	Instruction Following
1	GPT-5*	83.56%	96.15%
2	GPT-5-Nano*	82.04%	93.58%
3	GPT-5-Mini*	81.67%	94.23%
4	o3*	80.36%	94.96%
5	o4-Mini*	79.04%	95.30%

*Models use reasoning/thinking tokens. Full results for 101+ models on the leaderboard.

---

Environment Variables

export OPENAI_API_KEY="sk-..."
export GEMINI_API_KEY="..."
export ANTHROPIC_API_KEY="sk-ant-..."

---

Citation

If you use BeyondBench in your research, please cite our paper (accepted at ICLR 2026):

@misc{srivastava2025beyondbenchbenchmarkfreeevaluationreasoning,
      title={BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models},
      author={Gaurav Srivastava and Aafiya Hussain and Zhenyu Bi and Swastik Roy and Priya Pitre and Meng Lu and Morteza Ziyadi and Xuan Wang},
      year={2025},
      eprint={2509.24210},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.24210},
}

---

Links

• Leaderboard: ctrl-gaurav.github.io/BeyondBench
• GitHub: github.com/ctrl-gaurav/BeyondBench
• Paper: arXiv:2509.24210
• Documentation: Full Docs | Usage Guide
• Issues: GitHub Issues
• Email: gks@vt.edu, xuanw@vt.edu

---

Made with care by the BeyondBench Team | Virginia Tech, Department of Computer Science | Amazon AGI

Advancing the frontier of AI reasoning evaluation, one benchmark at a time.

License: Apache-2.0