A framework for evaluating LLM assistants on task alignment — the ability to align with users on their intended tasks.
pip install -e .Requires Python 3.10+. See pyproject.toml for dependencies.
from task_alignment import TaskAlignmentEnv
env = TaskAlignmentEnv(task="shopping", ...) # TODO
obs, info = env.reset()task_alignment/ core package
env.py POMDP env (Gymnasium-compatible)
config.py EnvConfig dataclass
metrics.py CommitReward, TotalIG, CommitAcc@k, ...
llm_utils.py LLM call helpers (retries, key fallback)
adapters/ per-domain adapters (shopping, gdpval, terminal_bench)
user/ user simulator + intent evolver
ranker/ fork-ranker + score-vector MCQ probe
prompts/ per-domain prompt templates
benchmark/ batch runner scripts
data/ dataset placeholders + download scripts
docs/ documentation
https://huggingface.co/datasets/daiandy/task-alignment-dataset
@article{dai2026interactivetaskalignmentpomdp,
title={Interactive Task Alignment as a POMDP},
author={Andy Dai and Zexue He and Zhenyu Zhang and Alex Pentland and Jiaxin Pei},
year={2026},
eprint={2607.16412},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2607.16412},
}
MIT — see LICENSE.