EALE — Evaluating Agent Loaded Evolution

EALE is the benchmark suite of the LIFT framework (Loaded Impact on Final Task). It is designed to measure whether a self-evolving LLM agent actually does better on unseen tasks after it has consolidated reusable artifacts (memory / skills / SOPs) from a set of warmup tasks — not to measure the agent's out-of-the-box ability.

• Paper: LIFT: A Counterfactual Framework and Benchmark for Disentangling Genuine Self-Evolution in LLM Agents
• Code / framework: https://github.com/FeiZhuNiU-INFJA/LIFT
• License: CC BY 4.0

⚠️ This is not a single flat table. EALE is a file-based, task-structured benchmark: each task is a folder containing a human-readable task markdown plus its input materials (CSV / documents). The Hugging Face auto-loader (load_dataset) therefore cannot cast it into one tabular schema; consume it by cloning the repository tree, not via load_dataset. See Usage below.

Dataset structure

Three layers: total dataset → scene datasets → train/test tasks.

benchmark_mds/
└── <scene>/                       # 14 scenes total
    ├── train/                     # 4 warmup tasks (drive artifact evolution)
    │   ├── q1_<shortname>/
    │   │   ├── q1_<shortname>.md   # task: query + requirements + trajectory requirements
    │   │   └── q1_materials/       # input files (csv/docs) for this task
    │   └── ...                     # q1..q4
    ├── test/                      # 2 holdout tasks (the LIFT paired Base-vs-Loaded contrast)
    │   ├── q5_<shortname>/ ...
    │   └── q6_<shortname>/
    └── skills/                    # optional scene-level seed skill

• 14 scenes × (4 warmup + 2 holdout) = 84 tasks (56 warmup + 28 holdout).
• Each task markdown has three parts:
  • query — a colloquial, deliberately under-specified first instruction.
  • requirements — the hidden acceptance checklist (≥ 12 independently-verifiable items per scene).
  • trajectory requirements — constraints on the execution path (tool-call validity/efficiency, allowed sources).
• Train/test requirement overlap is deliberately ~75% / 25%: 75% of test requirements match train (does the agent apply distilled experience in the right place?), 25% are variants of or contradictions to train requirements (exposing over-fitting to literal training requirements).

Scenes

Travel Planning · Business Travel · Healthy-Snack Guide · Cat-Food Guide · Stock Investment Decisions · data_analysis · Sales-Ops Weekly Analytics · Xiaohongshu Product Ops · PPT Creation · Startup BP / Investor Roadshow · Weekly-Report Auto-Generation · Team-Building Planning · information_search_gathering · English-Grammar Learning Guide.

Languages

Bilingual: task queries/requirements and input materials are in English (en) and Chinese (zh) depending on the scene.

Usage

# Recommended: clone the repository tree (do NOT rely on load_dataset)
huggingface-cli download FeiZhuNiU-INFJA/EALE --repo-type dataset --local-dir ./EALE

# Or via the LIFT framework's preprocessor, which compiles the markdown tree
# into machine-readable suite JSON (assets/benchmarks/*.json):
python -m src.cli.preprocess          # BENCHMARK_SOURCE=huggingface

Intended use & scope

• In scope: benchmarking the marginal benefit of an agent's self-evolution (Base vs. Loaded on held-out tasks); efficiency-first metrics (attempts / tool calls / tokens).
• Out of scope: measuring an agent's raw one-shot capability; head-to-head agent leaderboards.

Curation & known limitations

• Curation: tasks are expert-authored to be verifiable (rule-checkable headings, fixed CSV columns, numeric thresholds) rather than optimized for product polish, to minimize grader variance.
• Limitations: requirements reflect the authors' domain assumptions; some scenes are culture/locale-specific (e.g., Chinese consumer scenarios); the benchmark measures artifact usefulness, not safety (safety regression is an optional, orthogonal dimension in LIFT).
• Personal/sensitive data: input materials (call records, surveys, financials, etc.) are synthetic, authored for the benchmark; they do not contain real personal data.

Citation

@misc{lift2027,
  title        = {LIFT: A Counterfactual Framework and Benchmark for Disentangling Genuine Self-Evolution in LLM Agents},
  author       = {Anonymous},
  year         = {2027},
  howpublished = {\url{https://github.com/FeiZhuNiU-INFJA/LIFT}},
  note         = {Dataset: \url{https://huggingface.co/datasets/FeiZhuNiU-INFJA/EALE}}
}