SkillEvolBench-Team/Skill-Evol-Bench

SkillEvolBench Dataset

SkillEvolBench is a diagnostic benchmark for testing whether LLM agents can convert episodic task experience into reusable procedural skills. It accompanies the paper SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills.

This Hugging Face dataset page hosts the benchmark assets used by the paper's skill-evolution protocol: role-instantiated task directories, verification assets, and curated seed skills. The Dataset Viewer is disabled because the release is a runnable benchmark bundle with mixed file formats rather than a single table.

Resources

• Project page
• arXiv
• Paper PDF
• GitHub repository

Dataset Contents

SkillEvolBench contains 180 tasks across six real-world agent environments. Each environment contains five latent skill families, and each family contains six role-conditioned tasks. In the paper's terms, a family is a recurring procedural capability rather than a topic label: tasks within a family share an underlying procedure but vary failure modes, surface forms, and deployment conditions.

Environment	Procedural skill families
Code Debugging & Modification	systematic-error-diagnosis; dependency-conflict-resolution; safe-refactoring; multi-file-bug-fix; merge-conflict-resolution
Tool & API Orchestration	pre-call-parameter-validation; retry-and-backoff; pagination-complete-retrieval; multi-step-orchestration; response-validation-fallback
Data Processing & Structured Query	schema-inspection-before-query; type-normalization-before-sort; key-alignment-before-merge; null-safe-filtering-aggregation; result-sanity-reconciliation
Document Parsing & Transformation	structured-field-extraction; cross-format-migration; template-fill-from-context; document-diff-comparison; multi-source-merge-reconciliation
Research & Information Synthesis	multi-source-search-filter; evidence-grounded-comparison; citation-verification; constrained-summarization; contradiction-detection
Communication & Scheduling	inbox-triage-prioritization; context-aware-reply-drafting; follow-up-tracking; meeting-scheduling-constraints; thread-action-extraction

The uploaded dataset is organized as:

skills/
tasks/

skills/ contains 30 curated, gap-exposed seed skills. These skills are designed to be useful but incomplete, so agents can refine them through experience rather than simply relying on oracle procedures.

tasks/ contains 180 benchmark task directories. Each task bundles the instruction, fixture/environment files, tests, and verification assets needed to run or audit the benchmark task from source.

Construction Principles

The benchmark is source-driven and human-curated. The paper does not reuse existing benchmark instances; instead, it uses open-source agent skill collections, skill-oriented benchmarks, and practitioner-facing examples as evidence for task topics and workflow motifs. Families are retained when they satisfy three criteria:

• real-world relevance
• procedural skill fit
• verifiable evolvability

Each family-level procedure must be specific enough to write as a skill, general enough to transfer beyond one fixture, and evaluable with deterministic outcome checks plus process-level evidence. Each task contains an instruction and fixture, a verification suite, and a scoring rubric. Verifiers include public tests, hidden tests for edge cases and shifts, and process checks for brittle strategies such as hard-coded constants, swallowed exceptions, skipped validation, or incomplete repairs.

Task Arc

Each skill family is instantiated as a controlled six-role progression. The first three roles form the acquisition phase, where skill updates are allowed. The last three roles form frozen deployment, where the agent may use the accumulated library but cannot create, revise, retire, or otherwise modify skills.

Role	Phase	What it tests
Canonical	Acquisition	The base procedure in its cleanest form.
Enriched	Acquisition	A missing sub-capability or edge condition that the initial procedure should learn to cover.
Variant	Acquisition	A changed surface form that preserves the same latent procedure.
Context shift	Frozen deployment	Whether the skill is invoked when the need is embedded in a broader request.
Adversarial	Frozen deployment	Whether the skill resists shortcut solutions that can pass shallow checks.
Composition	Frozen deployment	Whether the skill combines with other skills in a realistic workflow.

Appendix B of the paper records task-level design signals such as gap=..., traps=..., and composition=.... These signals encode why each task exists in the arc: some expose missing sub-capabilities during acquisition, some embed implicit invocation requirements, some introduce shortcut traps, and some require cross-family skill composition.

Protocol Mapping

In the paper's protocol, each environment is evaluated as an independent lifelong episode. An environment-scoped skill library is initialized, acquisition tasks are completed, the library is frozen, and deployment tasks are evaluated. When replay is enabled, the original acquisition tasks are rerun with the final frozen library to separate local recovery from transfer to harder roles.

During acquisition, each task attempt yields a compacted trajectory summary from harness-recorded artifacts such as instructions, file accesses, tool calls, commands, edits, generated outputs, tests, and final responses. Verifier feedback includes outcome results, process checks, rewards, and diagnostics. Skill authoring is family-local: the Skill Author sees same-family acquisition history and same-family skills, even though the task-solving agent may read the broader environment-level library.

The dataset supports the paper's main conditions:

• No-Skill: no skill library is mounted.
• Curated-Start: the family starts from a gap-exposed curated seed skill.
• Self-Generated: the family starts without a skill and may induce one from experience.
• Raw-Trajectory: prior same-family learning trajectories are retrieved directly rather than distilled into skills.

Reported Metrics

For each attempt, the verifier returns an outcome score, a process score, an overall score, and a binary success indicator. The paper reports success rates over protocol-defined subsets:

Metric	Meaning
LSR	Learning success rate over canonical, enriched, and variant acquisition tasks while updates are allowed.
RSR	Replay success rate on acquisition tasks after the environment library has been frozen.
ESR	Frozen deployment success rate over context-shift, adversarial, and composition tasks.
CSSR	Context-shift success rate, measuring implicit skill invocation in broader requests.
ARSR	Adversarial-role success rate, measuring resistance to shortcut solutions.
CompSR	Composition success rate, measuring multi-skill composition.

For runnable experiment code, model/provider configs, baseline definitions, and reporting scripts, use the GitHub repository.

Download

With the Hugging Face CLI:

hf download SkillEvolBench-Team/Skill-Evol-Bench --repo-type dataset --local-dir Skill-Evol-Bench

With Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="SkillEvolBench-Team/Skill-Evol-Bench",
    repo_type="dataset",
    local_dir="Skill-Evol-Bench",
)

Citation

@article{lei2026skillevolbench,
  title={SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills},
  author={Lei, Yingtie and Wan, Zhongwei and Zhang, Jiankun and Alam, Samiul and Zhong, Zixuan and Huang, Peizhou and Wang, Xin and Zhang, Jingxuan and Zhou, Donghao and Hsieh, Yunta and others},
  journal={arXiv preprint arXiv:2605.24117},
  year={2026}
}