HERA-Bench

HERA-Bench evaluates whether a tool-using agent can complete a feasible task and recognize when the same request cannot be fully satisfied. It accompanies HERA: Harness–Environment Co-evolution for Reliable Agentic Abstention.

Code and harnesses · Project page · Paper: [ARXIV_ID_PENDING]

Data

Item	Count
Matched task pairs	60
Feasible / should-abstain instances	60 / 60
Domain environment implementations	58
Shared document environment	1
Supporting briefs	60
Initial-state files	240

Each pair shares a request, system prompt, and public tool interface. Its two environment states determine feasibility. A correct response to an infeasible task may be an evidence-grounded incomplete report, as specified by its brief. Environment counts identify implementation namespaces, not independent subject areas.

index/test.jsonl                  60 pair records
tasks.json                       Shared prompts
data/tasks/hera_NNN/              act/ and abstain/ initial states
data/environments/               Executable tools
public/briefs/                   Supporting documents and report schemas
private/{pairs,bundles,controller}/
                                 Routing, references, and grading
benchmark.py, runtime/           Environment execution

The index fields are pair_id, number, action_type, environment, instruction, system_prompt, brief_path, pair_metadata_path, act_initial_state_paths, and abstain_initial_state_paths. IDs run from hera_001 to hera_060. All paths are relative to the repository root. See schema and task inventory.

Load and run

The repository is currently private. Authenticate before loading:

from datasets import load_dataset
from huggingface_hub import snapshot_download

pairs = load_dataset("sxcn/HERA-Bench", split="test", token=True)
root = snapshot_download("sxcn/HERA-Bench", repo_type="dataset",
                         token=True, local_dir="HERA-Bench")

load_dataset reads the 60-row index; the complete download includes executable environments and grading. With Python 3.12, from the downloaded directory:

pip install -r requirements.txt
python -B benchmark.py verify
python -B benchmark.py serve --number 1 --side act --record ../episode.json

verify replays all 120 references offline without a model. serve exposes an MCP server over standard input/output. Use the GitHub runner to evaluate the base or final harness.

Evaluation

For 60 pairs, Act is the percentage of correctly completed feasible instances; Abstain is the percentage of correct infeasible-instance responses; Pair requires both outcomes for the same pair. Use the supplied grader and retain failed episodes. Evaluation details

The solver receives only task prompts, public tools, and their observations. Keep side labels, initial states, reference solutions, and graders outside its context and filesystem access. private/ denotes controller-only information during evaluation; these files are included for reproducibility.

This is a small English-language benchmark of local simulations. Tasks underwent author review and revision; the release does not certify untouched blind-test provenance or strict training-domain disjointness. Training and validation pools are not included. Provenance

License and citation

Dataset: CC BY 4.0. Third-party components retain their original notices; licensing of original accompanying code remains to be finalized.

@misc{hera_bench_2026,
  title = {{HERA-Bench}},
  year = {2026},
  url = {https://huggingface.co/datasets/sxcn/HERA-Bench},
  note = {Companion benchmark for HERA. Updated paper arXiv identifier: [ARXIV_ID_PENDING]}
}