Jiefuo/RebuttalCodeBench

RebuttalBench

RebuttalBench is a benchmark for scientific agents to automatically implement new experiments from paper repositories, sourced from accepted-paper rebuttals. Evaluation is automated with an agent-as-a-judge protocol: the judge checks whether produced experiment result tables fully support scientific claims distilled from ground-truth rebuttal experiments.

The current release focuses on papers from ICLR 2026 and NeurIPS 2025.

Task Variants

Two task styles are included:

1. Run-as-instructed (version a)
   Agents implement and run experiments by following detailed finalized experiment plans.

2. Claim verification / SciFy style (version b)
   Agents plan, implement, and run experiments to verify speculative scientific claims.

Both variants are derived from rebuttal experiments in accepted ICLR 2026 / NeurIPS 2025 papers.

Release Structure

The release includes:

• tasks_v0_release/
• tasks_v0_release_workspace/
• result_release/
• log_release/
• prompt_release/
• dataset_info_release/

tasks_v0_release/

Core benchmark inputs, organized by paper_id.

tasks_v0_release/
  <paper_id>/
    <paper_id>_code/                       # paper repository snapshot used by agents
    deps/                                  # optional vendored/pinned dependencies
    rebuttalcodebench/
      background.txt                       # paper/task context
      meta_verified.txt                    # dependency/runtime manifest for run-as-instructed
      finalized_instructions/
        exp_<n>.txt                        # version a instruction files
      claims/
        exp_<n>.txt                        # version a judge claim files
      scify_claims/
        exp_<n>.txt                        # version b (SciFy) claim files
      # env tarballs may be present depending on task packaging

Notes:

• There are 74 paper folders (paper_ids) in this release.
• finalized_instructions, claims, and scify_claims each cover 127 experiment files (exp_<n>.txt) in total.

tasks_v0_release_workspace/

Working area created during runs (typically empty before running):

tasks_v0_release_workspace/
  codex/<date-time>/<paper_id>_code_<exp_id>/
  claude_code/<date-time>/<paper_id>_code_<exp_id>/

Agents copy read-only source code from tasks_v0_release/ into this workspace and perform all edits/execution there.

result_release/

Outputs from experiment execution and judging (typically empty before running):

result_release/
  codex/<date-time>/<paper_id>/res_<n>.txt
  codex/<date-time>/check_results.json
  claude_code/<date-time>/<paper_id>/res_<n>.txt
  claude_code/<date-time>/check_results.json

• For version a: res_<n>.txt is a fixed-format experiment result table.
• For version b (SciFy): res_<n>.txt includes Experiment Conducted + Result Table.
• check_results.json stores evaluator verdicts.

log_release/

Execution logs written during runs (typically empty before running):

log_release/
  codex_logs/run_exp/<date-time>/<paper_id>_<exp_id>.txt
  claude_code_logs/run_exp/<date-time>/<paper_id>_<exp_id>.txt

These logs track retries, command traces, resource usage, and failure causes.

prompt_release/

Prompt templates for both runner and evaluator agents:

prompt_release/
  gateway.txt
  codex/
    run_exp.txt
    check_results.txt
    run_exp_scify.txt
    check_results_scify.txt
  claude_code/
    run_exp.txt
    check_results.txt
    run_exp_scify.txt
    check_results_scify.txt

Mapping:

• Version a (run-as-instructed)
  • execution: run_exp.txt
  • evaluation: check_results.txt
• Version b (SciFy / claim verification)
  • execution: run_exp_scify.txt
  • evaluation: check_results_scify.txt

dataset_info_release/

dataset_info_release/
  meta_tasks_v0_release.csv

This CSV is the dataset manifest.

Dataset Stats

From dataset_info_release/meta_tasks_v0_release.csv:

• Papers (paper_id): 74
• Experiments (sum(num_exp)): 127

By venue:

• ICLR 2026: 41 papers, 62 experiments
• NeurIPS 2025: 33 papers, 65 experiments

By verification level:

• EXECUTION: 27 papers, 46 experiments
• DEPENDENCY CHECKED: 47 papers, 81 experiments

CSV Columns

• paper_id: paper/task identifier (folder key in tasks_v0_release/)
• num_exp: number of experiment IDs associated with this paper
• source: venue/year (ICLR 2026 or NeurIPS 2025)
• verification_level:
  • EXECUTION: verified with actual agent runs (version a, Codex) confirming tasks can execute in practice
  • DEPENDENCY CHECKED: dependencies were reachability-checked during construction/setup (e.g., hosted models/datasets accessible), but full end-to-end execution is not guaranteed yet

How To Run

For both Codex and Claude Code, use the CLI template:

Follow the instructions in <prompt file> to carry out the tasks. Spawn sub-agents when possible - fit the size of parallelism to the number of available CPUs.

Where <prompt file> comes from prompt_release/<agent>/.

1) Execution

• Version a (run-as-instructed): use run_exp.txt
• Version b (claim verification / SciFy): use run_exp_scify.txt

Examples of prompt file paths:

• prompt_release/codex/run_exp.txt
• prompt_release/codex/run_exp_scify.txt
• prompt_release/claude_code/run_exp.txt
• prompt_release/claude_code/run_exp_scify.txt

2) Evaluation

• Version a: use check_results.txt
• Version b: use check_results_scify.txt

Evaluator agents are also Codex and Claude Code.

3) Subset Runs and Prompt Customization

• You can create a subset.txt (as described in run_exp*.txt) to run only part of the benchmark.
• You can edit prompt templates for your own workflow, but keep path/IO conventions aligned with this release structure.

APIs and Credentials

Many tasks require API-served models. Current setup uses the JHU WSE AI Gateway:

• WSE AI Gateway
• Gateway docs

In this release, API instructions are centralized in prompt_release/gateway.txt.

• If you want to use your own provider setup, modify gateway.txt.
• If you want to reproduce the exact gateway setting used by the authors, contact Daniel and Ben.
• Besides model API access, the other required API/token in many tasks is Hugging Face authentication.

TODOs and Reminders

1. No guarantee that every repo/dependency is fully available yet.
   Ongoing validation is running for DEPENDENCY CHECKED tasks using version a.
   If you encounter failures not caused by agent behavior (e.g., unavailable private checkpoints), please report them.

2. Version b (SciFy) has not been fully run by the authors yet.
   Please monitor resources carefully (API cost, GPU requests) and share findings/suggestions.

3. Set cache directories (e.g., .cache for pip / Hugging Face / related tools) to a storage location with sufficient capacity.
   Agents may download large dependency/model artifacts during execution.

4. There is currently no sandboxing/isolation layer for agent execution: agents operate from the project root and copy each paper repo into tasks_v0_release_workspace/ per run.
   Singularity-based sandboxing is in progress.