AutoLibra

Agent Metric Induction from Open-Ended Human Feedback
ICLR 2026

---

AutoLibra turns open-ended natural-language feedback on agent trajectories into fine-grained, interpretable evaluation metrics. Give it sentences like "the agent did not choose iPhone 14/15" or "this agent has too much autonomy", and it gives back metrics with definitions and concrete examples — ready to drive an LLM-as-a-Judge, diagnose failure modes, or serve as optimization targets for prompt engineering.

The pipeline has three LLM-driven operators:

1. Feedback grounding — break feedback into (behavior, feedback, sign) aspects tied to a specific part of the trajectory.
2. Behavior clustering — group similar aspects into metrics with definitions + positive/negative examples.
3. LLM-as-a-Judge — score trajectories on the induced metrics, producing positive and negative traits.

coverage and redundancy meta-metrics then measure how well the induced metric set covers unseen human feedback.

Install

uv is the supported package manager.

git clone https://github.com/Open-Social-World/autolibra
cd autolibra
uv sync

AutoLibra talks to Azure OpenAI. Create a .env in the repo root:

AZURE_API_KEY=...
AZURE_ENDPOINT=https://<your-resource>.openai.azure.com/
AZURE_OPENAI_4O_MODEL=gpt-4o
AZURE_OPENAI_O3_MODEL=o3-mini
GITHUB_PERSONAL_ACCESS_TOKEN=...   # optional, for some GitHub-backed dataset loaders

Quickstart (Python API)

import asyncio
from openai import AsyncAzureOpenAI
from autolibra_core import (
    feedback_grounding,
    behavior_clustering,
    run_llm_eval,
    MetricTrainingInstance,
)

async def main():
    client = AsyncAzureOpenAI(azure_endpoint=..., api_key=..., api_version=...)

    # 1. Build MetricTrainingInstance objects from your trajectories + feedback.
    #    See `autolibra_core.datasets.*` for loaders; each yields instances like:
    #    MetricTrainingInstance(task=..., agent_id=..., trajectory=..., feedback=...)
    instances: list[MetricTrainingInstance] = load_my_data()

    # 2. Ground feedback into aspects per instance.
    aspects = []
    for inst in instances:
        aspects.extend(await feedback_grounding(inst, client))

    # 3. Cluster aspects into metrics with definitions + behavior examples.
    induced = await behavior_clustering(aspects, client)
    metrics = induced.metrics

    # 4. Run LLM-as-a-Judge on the trajectories with the induced metrics.
    results = await run_llm_eval(instances, metrics, client)
    print(metrics)
    print(results)

asyncio.run(main())

Datasets

Shared trajectories + feedback live on Hugging Face at open-social-world/autolibra. Install git-lfs first, then:

git clone https://huggingface.co/datasets/open-social-world/autolibra .data

Preprocess or convert from the original sources with the dataset loaders in autolibra_core.datasets (one module per dataset):

uv run python -m autolibra_core.datasets.webarena
uv run python -m autolibra_core.datasets.sotopia
uv run python -m autolibra_core.datasets.cogym
uv run python -m autolibra_core.datasets.balrog_babaisai
# ...

To contribute a new dataset, add a loader under packages/autolibra-core/src/autolibra_core/datasets/ that emits MetricTrainingInstance objects, then push the converted artifacts to the shared HF repo.

Collecting feedback

Two annotation frontends ship in src/tty/:

# Terminal UI
uv run python src/tty/tty_annotation.py \
    .data/webarena .data/annotations/webarena \
    --annotator-id <your-name>

# Streamlit UI (browser)
uv run streamlit run src/tty/tty_annotation.py \
    .data/sotopia .data/annotations/sotopia \
    -- --annotator-id <your-name> --use-streamlit

# Review annotations
uv run streamlit run src/tty/view_annotations.py \
    -- .data/annotations/sotopia/annotations

Annotators are shown each trajectory step by step and write a single piece of natural-language feedback per trajectory. See §2.1 of the paper for annotation guidelines.

Training / iterative metric induction

src/training/ contains the end-to-end scripts used in the paper, including the Gemini-based iterative induction pipeline from §4 ("AutoLibra as a Ladder"). See src/training/README.md for the per-script breakdown.

The prompt-optimization harness used on Baba-Is-AI (Fig. 4 in the paper) is in prompt_optimization/; environment submodules (BALROG, etc.) are declared in .gitmodules.

Repository layout

autolibra/
├── packages/
│   ├── autolibra-core/        # Core operators, evaluators, dataset loaders
│   └── osw-data/              # Trajectory / metric primitives
├── src/
│   ├── tty/                   # Annotation UIs (terminal + Streamlit)
│   ├── training/              # Iterative induction / eval pipelines
│   ├── tools/                 # Analysis utilities
│   └── plot/                  # Figure generation
├── prompt_optimization/       # Agent prompt-optimization harness
└── pyproject.toml

Contributing

Issues and PRs welcome. Before opening a PR:

uv run pre-commit run --all-files
uv run pytest

Citation

@inproceedings{zhu2026autolibra,
  title     = {AutoLibra: Agent Metric Induction from Open-Ended Human Feedback},
  author    = {Hao Zhu and Phil Cuvin and Xinkai Yu and
               Charlotte Ka Yee Yan and Jason Zhang and Diyi Yang},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=4BjGVZ7Bxn}
}

Acknowledgments

Supported by ONR N000142412532, NSF IIS-2247357, and DARPA Friction for Accountability in Conversational Transactions. Compute credits from Google Cloud Platform and Modal. Thanks to the Stanford SALT Lab for feedback.