--- pretty_name: ReproBoundary Execution Traces tags: - code-agents - paper-reproduction - execution-traces - reproducibility - benchmark - research-agents size_categories: - n<1K ---
# ReproBoundary Execution Traces **How Reliably Can Coding Agents Reproduce AI Research Papers?** 🌐 [Project Page](https://reproboundary.github.io/) Β· πŸ’» [GitHub](https://github.com/ReproBoundary/ReproBoundary) Β· πŸ“Š [Leaderboard](https://github.com/ReproBoundary/ReproBoundary.github.io/blob/main/data/leaderboard.json) Β· πŸ“¦ [Submit a Run](https://github.com/ReproBoundary/ReproBoundary.github.io/blob/main/SUBMIT_A_RUN.md)
ReproBoundary construction pipeline from the paper
ReproBoundary comprises 47 resource-bounded paper-reproduction tasks and 423 runs across nine model–agent configurations. This repository releases the execution traces, their verification metadata, and the source-paper files listed for the 47 task IDs. Task definitions and evaluators live in the [GitHub release](https://github.com/ReproBoundary/ReproBoundary); paper results and submission instructions live with the [project website](https://github.com/ReproBoundary/ReproBoundary.github.io). ## Data Structure `TRAJECTORY_INDEX.csv` maps each frozen task and model–agent configuration to one compressed log and its SHA-256 fingerprint. `trajectories/runs/` stores the 423 released `agent.log.gz` files. The authoritative task catalog remains on [GitHub](https://github.com/ReproBoundary/ReproBoundary/blob/main/tasks/TASK_INVENTORY_47.csv) instead of being duplicated here. `task_pool/` contains 43 source-paper PDFs whose downstream redistribution terms were verified. `PAPER_INDEX.csv` covers all 47 tasks and records titles, official sources, license evidence, sizes, page counts, SHA-256 fingerprints, and redistribution status. Four papers remain link-only because an official download exists but downstream redistribution permission was not verified. ```text ReproBoundary-Execution-Trajectories/ β”œβ”€β”€ README.md β”œβ”€β”€ DATASET_MANIFEST.json β”œβ”€β”€ DATASET_SHA256SUMS β”œβ”€β”€ PAPER_INDEX.csv β”œβ”€β”€ REDACTION_POLICY.md β”œβ”€β”€ construction-pipeline.png β”œβ”€β”€ verify_dataset.py β”œβ”€β”€ task_pool/ β”‚ └── .pdf └── trajectories/ β”œβ”€β”€ TRAJECTORY_INDEX.csv └── runs/ └── ///agent.log.gz ``` The 43 bundled PDFs were matched to the frozen project IDs, checked for readable PDF structure and title agreement, and included only where redistribution terms were verified. For the four link-only entries, follow the official source in `PAPER_INDEX.csv`. Hugging Face does not render the gzip logs as table rows. Start with `trajectories/TRAJECTORY_INDEX.csv`, then open only the logs needed for an analysis. ## Download and Verify ```bash hf download SupingSun/ReproBoundary-Execution-Trajectories --repo-type dataset --local-dir ReproBoundary-Execution-Trajectories cd ReproBoundary-Execution-Trajectories && python3 verify_dataset.py ``` ## Practical Uses and Limits - Examine execution records for evidence that required stages ran. - Analyze where and why paper-reproduction attempts fail. - Study elapsed time, token usage, and safety. - Link each released trace to its fixed task and model–agent configuration. The execution traces do not include checkpoints, predictions, model weights, training states, datasets, or complete third-party repositories. Execution records provide evidence that required stages ran, but they cannot prove that every implementation detail matches the paper. Without the retained submitted outputs, the traces alone cannot rerun the official evaluator for most tasks. See the [evaluator provenance](https://github.com/ReproBoundary/ReproBoundary/blob/main/evaluators/EVALUATOR_PROVENANCE.md) and [redaction policy](REDACTION_POLICY.md) for details. ## License Status The project license is under review. Public access does not grant permission to redistribute files that do not carry their own license; source papers, datasets, models, and third-party projects retain their original terms.