GISTBench — Groundedness & Interest Specificity Test Bench

GISTBench evaluates how well LLMs understand users from their engagement history. Given a user's interactions with content (videos, articles, books, etc.), the benchmark measures whether an LLM can extract meaningful interests, ground them in evidence, and cite specific relevant items.

Dataset Details

• Rows: 4,214,059 engagements
• Users: ~1,000 anonymized users
• Format: Parquet

Schema

Column	Type	Description
user_id	int64	Anonymized user identifier
object_id	int64	Anonymized content item identifier
object_text	string	Text description of the content item
interaction_type	string	One of: explicit_positive, implicit_positive, implicit_negative, explicit_negative
interaction_time	string	Anonymized interaction timestamp

Interaction Types

Type	Meaning	Examples
explicit_positive	User actively expressed positive interest	Liked, favorited, rated highly
implicit_positive	Passive positive signal	Watched fully, clicked through
implicit_negative	Passive negative signal	Scrolled past, skipped
explicit_negative	User actively expressed dislike	Downvoted, reported, rated low

Usage

from datasets import load_dataset

ds = load_dataset("facebook/gistbench")
print(ds)
# DatasetDict({
#     train: Dataset({
#         features: ['interaction_type', 'user_id', 'object_id', 'interaction_time', 'object_text'],
#         num_rows: 4214059
#     })
# })

# Access rows
print(ds["train"][0])

Citation

If you use GISTBench in your research, please cite:

@misc{fostiropoulos2026gistbench,
  title={GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification},
  author={Iordanis Fostiropoulos and Muhammad Rafay Azhar and Abdalaziz Sawwan and Boyu Fang and Yuchen Liu and Jiayi Liu and Hanchao Yu and Qi Guo and Jianyu Wang and Fei Liu and Xiangjun Fan},
  year={2026},
  eprint={2603.29112},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2603.29112}
}

License

This dataset is released under CC BY-NC 4.0.