---
license: mit
task_categories:
- question-answering
- text-classification
tags:
- academic-papers
- llm-agents
- benchmarking
- inspect-ai
- react-agents
- citations
- research-trends
size_categories:
- 1K<n<10K
language:
- en
citation: |
  @misc{ye2026prooftimebenchmarkevaluating,
        title={Proof of Time: A Benchmark for Evaluating Scientific Idea Judgments}, 
        author={Bingyang Ye and Shan Chen and Jingxuan Tu and Chen Liu and Zidi Xiong and Samuel Schmidgall and Danielle S. Bitterman},
        year={2026},
        eprint={2601.07606},
        archivePrefix={arXiv},
        primaryClass={cs.CL},
        url={https://arxiv.org/abs/2601.07606}, 
  }
---
# Proof of Time: A Benchmark for Evaluating Scientific Idea Judgments

This dataset contains benchmarks for evaluating LLM agents on academic paper analysis tasks that require understanding research trends, citations, and future directions. All evaluation data uses **post-training-cutoff** (2025) papers to avoid data contamination.

## Dataset Description

**Paper**: *Proof of Time: A Benchmark for Evaluating Scientific Idea Judgments*

**Repository**: [https://github.com/shan23chen/proof_of_time](https://github.com/shan23chen/proof_of_time)

This dataset includes:
1. **Benchmark Tasks** (3.8 MB): JSONL files with multiple-choice questions and evaluation samples
2. **Sandbox Data** (66 MB): Historical paper data, faculty publications, and SOTA metrics for agent evaluation

### Why "Proof of Time"?

The benchmark suite focuses on temporal reasoning: agents must analyze historical patterns to make predictions about future research directions, award recipients, and citation impact. Tasks require genuine understanding of research trends rather than memorization.

## Dataset Structure

### Benchmarks Directory (3.8 MB)

Contains 4 task families with 10 evaluation datasets:

#### Award Prediction (641 KB)
Predict which papers will win best paper awards at top NLP conferences.

- `pre-cutoff_mcq.jsonl` (421 KB): Pre-2025 conference awards (ACL/EMNLP/NAACL 2018-2024)
- `post-cutoff_emnlp.jsonl` (29 KB): Post-2025 EMNLP awards
- `post-cutoff_acl_naacl.jsonl` (191 KB): Post-2025 ACL/NAACL awards

#### Citation Forecasting (2.7 MB)
Predict future citation counts for recently published papers.

- `multiple_choice.jsonl` (1.1 MB): Predict highest-cited paper among choices
- `ranking.jsonl` (1.2 MB): Rank papers by predicted citation counts
- `bucket_prediction.jsonl` (368 KB): Classify papers into citation ranges (0-1, 1-5, 5-10, 10-50, 50+)

#### Faculty Future Work (469 KB)
Predict research directions of AI faculty members based on publication history.

- `professor_field_mcq.jsonl` (49 KB): Predict research field for professor's future work
- `professor_article_mcq.jsonl` (404 KB): Predict which article professor would author
- `field_focus_mcq.jsonl` (16 KB): Classify research focus by field

#### SOTA Forecasting (26 KB)
Predict state-of-the-art performance ranges on ML benchmarks.

- `mcq_dataset.jsonl` (26 KB): Predict benchmark performance buckets (0-20, 20-40, 40-60, 60-80, 80-100)

### Sandbox Data Directory (66 MB)

Reference data for ReAct agents to query during evaluation:

- `citation/historical_papers_2021_2024.jsonl` (21 MB): Historical papers with citation counts

## Usage

### With Inspect AI

```python
from datasets import load_dataset
from inspect_ai import eval
from inspect_ai.dataset import json_dataset

# Load dataset from HuggingFace
ds = load_dataset("AIM-Harvard/proof-of-time")

# Run evaluation with Inspect AI
eval(
    task="your_benchmark.py@task_name",
    model="openai/gpt-5-mini-2025-08-07",
    limit=5
)
```

### Quick Start

```bash
# Install dependencies
pip install inspect-ai datasets

# Clone repository with benchmark implementations
git clone https://github.com/shan23chen/proof_of_time.git
cd proof_of_time

# Download dataset
from datasets import load_dataset
ds = load_dataset("AIM-Harvard/proof-of-time")

# Run evaluation
inspect eval benchmarks/award_react/benchmark.py@pre_cutoff_task \
    --model openai/gpt-5-mini-2025-08-07 \
    --limit 5
```

## Dataset Fields

Each benchmark JSONL file contains samples with:

- **`question`**: Task prompt for the agent
- **`answer`**: Correct answer (for evaluation)
- **`choices`**: Multiple choice options (if applicable)
- **`metadata`**: Additional context (paper titles, years, venues, authors, etc.)

Example from award prediction:

```json
{
  "question": "Which recognition tier (Findings/Main/Outstanding/Best) best fits the paper?",
  "context": "{title}+{abstract}+{author}"
  "answer": "A",
  "choices": ["Best", "Outstanding", "Main", "Findings"],
}
```

## Benchmark Design

- **ReAct Agents**: Agents use tools (bash, Python, text editor) to explore sandboxed paper datasets
- **Sandboxed Environments**: Docker containers with read-only paper data (no internet access)
- **Offline Prompt**: Custom "Antigravity" prompt inspired by principles of focused exploration
- **Multiple Variants**: Each task has standard (agent), simple (zero-shot), and no-offline-prompt versions

## Supported Models

The benchmark suite has been tested with:

- **OpenAI**: gpt-5.2, gpt-5.1, gpt-5-mini, gpt-5-nano
- **Google**: gemini-3-pro, gemini-3-flash, vertex/gemini-2.5-pro, vertex/gemini-2.5-flash
- **Anthropic**: vertex/claude-opus-4-5, vertex/claude-sonnet-4-5, vertex/claude-haiku-4-5

## Data Sources

- **Award Predictions**: ACL Anthology, EMNLP/ACL/NAACL conference proceedings
- **Citation Forecasting**: Google Scholar citation counts
- **Faculty Predictions**: AI faculty CVs and publication records
- **SOTA Forecasting**: Papers with Code leaderboards

## License

- **Code**: MIT License
- **Data**: Derived from publicly available academic papers and conference proceedings

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{ye2026prooftimebenchmarkevaluating,
      title={Proof of Time: A Benchmark for Evaluating Scientific Idea Judgments}, 
      author={Bingyang Ye and Shan Chen and Jingxuan Tu and Chen Liu and Zidi Xiong and Samuel Schmidgall and Danielle S. Bitterman},
      year={2026},
      eprint={2601.07606},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.07606}, 
}
```


## Additional Resources

- **GitHub Repository**: [https://github.com/shan23chen/proof_of_time](https://github.com/shan23chen/proof_of_time)
- **Documentation**: See repository README for detailed usage
- **Setup Guide**: [SETUP.md](https://github.com/shan23chen/proof_of_time/blob/main/SETUP.md)
- **Paper**: [arXiv](https://www.arxiv.org/abs/2601.07606)

## Contact

- **Issues**: [https://github.com/shan23chen/proof_of_time/issues](https://github.com/shan23chen/proof_of_time/issues)
- **Email**: aim@seas.harvard.edu
- **Organization**: AIM Harvard

## Updates

- **2026-01-08**: Initial release (Tiers 1-2: benchmarks + sandbox data, 69.8 MB total)