lejen/agentabstain-ar

AgentAbstain: Do LLM Agents Know When Not to Act?

AgentAbstain is a paired-task benchmark for agentic abstention: the calibrated ability of tool-using LLM agents to recognize when not to act. It contains 263 task pairs across 42 executable MCP sandbox environments, built on an agent-native taxonomy of 8 abstention scenarios. Every should-act task ships with a should-abstain variant that differs by a single controlled perturbation to the instruction, the environment state, or the tool set, so no always-act or always-refuse policy can exceed 50% paired accuracy.

Across 17 frontier LLMs in 4 agent harnesses, the best agent reaches only 59.5% paired accuracy. See the paper for the full evaluation.

What is in this dataset

tasks.jsonl        526 rows: one per task variant (263 pairs x act/abstain), loadable via datasets
tasks/             the same tasks in their native layout, consumed directly by the runtime
  <category>/<task_id>/<act|abstain>/
    task.yaml           instruction, system prompt, execution DAG / abstention trigger, critical actions
    initial_states/     per-environment JSON state the sandbox is mounted with
environments/      42 executable MCP sandbox environments (environment.py, schema.py)

Fields in tasks.jsonl

Field	Description
pair_id	<category>/<task_id>, shared by the act and abstain variants of a pair
category	one of the 8 abstention scenarios (e.g. high_stakes_action, conflicting_evidence)
task_type	act or abstain
action_type	operational (state-mutating deliverable) or informational (read-only deliverable)
instruction	the user instruction given to the agent
system_prompt	the per-task system prompt
critical_actions	commit-class tool calls whose execution is the intended side effect (empty for informational act tasks)
execution_dag	act variants: the reference tool-call DAG
abstention_trigger	abstain variants: the perturbation record (what changed, where, and why acting is wrong)
environments	the sandbox environments this variant mounts

Scenario breakdown

Scenario	Pairs
S1 Missing critical parameter	39
S2 Ambiguous action specification	30
S3 Conflicting constraints	32
S4 High-stakes action	31
S5 Insufficient tool capability	34
S6 Critical tool failure	34
S7 Conflicting evidence	30
S8 Emergent risk discovery	33

131 pairs are operational and 132 informational.

Usage

Load the task metadata:

from datasets import load_dataset
ds = load_dataset("antiquality/agentabstain", split="test")

Run agents against the executable benchmark with the code repository:

from huggingface_hub import snapshot_download
data_dir = snapshot_download("antiquality/agentabstain", repo_type="dataset")
# point the runtime at <data_dir>/tasks and <data_dir>/environments

Evaluation crosses a deterministic commit check on the tool-call trace with an LLM judge on the terminal response; both evaluators ship with the code repository.

Contamination and regeneration

All tasks and environments are generated by the AbstainGen pipeline and validated by deterministic replay and cross-family critics; three independent annotators rate 94 to 98% of sampled 100 tasks as well-designed. The pipeline is fully documented in the paper, but its generation code is intentionally not released, as a public generator would let benchmark-matched training data be synthesized at scale. Fresh task instances can be generated privately on demand for future evaluation rounds.

License

CC BY 4.0. The companion code is released under MIT.

Citation

@misc{liu2026agentabstain,
  title  = {AgentAbstain: Do LLM Agents Know When Not to Act?},
  author = {Liu, Xun and Zhang, Yi Evie and Kasprova, Vira and Rabbani, Parisa and Zahraei, Pardis Sadat and Zhang, Tianyu and Ebrahimpour-Boroojeny, Ali and Chandrasekaran, Varun},
  year   = {2026},
  eprint = {2607.10059},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}