A Dynamic, Time-Aware Benchmark · ACL 2026
Can LLMs reason under the "fog of war," or just recite what they memorized?
LiveFact scores models against evidence sliced at three days before, on, and three days after each news event broke — using fresh events from Google News' World topic, refreshed every month so the test set is always unseen.
A continuously updated benchmark that tests reasoning under real-time uncertainty — not memorized recall
Every claim is evaluated at three time slices — using only the evidence that existed 3 days before, on, and 3 days after the event's headline date.
Claims aren't forced into a binary verdict. "Ambiguous" is a correct answer in its own right whenever the evidence genuinely can't support a verdict yet.
Classification Mode checks factual accuracy against the final, time-invariant truth. Inference Mode checks whether a model knows when it doesn't know yet.
LiveFact draws exclusively from Google News' World topic and refreshes every month with unseen events, keeping every release a genuine zero-shot test.
Current rankings of LLMs on the LiveFact benchmark
Every claim is sourced from Google News' World section, refreshed monthly.
CLS — Classification Mode: the model must commit to Real or Fake against the final, time-invariant ground truth, even if the evidence available at that moment can't support it yet.
INF — Inference Mode: the ground truth adapts to the evidence slice — "Ambiguous" is the correct answer when there isn't enough evidence yet.
Before / During / After: evidence available 3 days before, on, and 3 days after the event's headline date (T−3 / T / T+3).
A high Before-Event CLS score is not necessarily good news — it can mean the model is confidently guessing with no real evidence yet. Watch Before-Event INF instead: strong models score high there by correctly answering "Ambiguous."
| Rank | Model | Organization | Overall Avg. | Before Event (CLS) | Before Event (INF) | During Event (CLS) | During Event (INF) | After Event (CLS) | After Event (INF) |
|---|
How each monthly LiveFact release is built
We scrape trending stories from Google News' World topic daily at 00:00 GMT, then deduplicate and filter down to a core set of distinct, high-impact events.
For every event we build three evidence sets anchored to its headline date — T−3, T, and T+3 — simulating the real-world "fog of war" of incomplete information.
An LLM drafts neutral background context plus a balanced set of Real, Fake, and Ambiguous claims for each event, grounded in the retrieved evidence.
Every claim and label goes through independent expert review rounds against the source evidence before it's accepted into the release.
Models are scored in Classification Mode (fixed ground truth) and Inference Mode (evidence-adjusted ground truth) at each of the three time slices.
Join our community and contribute to advancing fake news detection research
Have a model you'd like to see on the leaderboard? Submit it for evaluation.
Submit ModelDownload the LiveFact dataset and evaluation scripts for your research on Hugging Face.
Download DatasetJoin the LiveFact maintenance team and help run the ongoing monthly benchmark cycle.
Join the TeamIf LiveFact is useful in your research, please cite our ACL 2026 paper.
@inproceedings{xu2026livefact,
title = "{L}ive{F}act: A Dynamic, Time-Aware Benchmark for {LLM}-Driven Fake News Detection",
author = "Xu, Cheng and Jin, Changhong and Niu, Yingjie and Yan, Nan and Mei, Yuke and Guan, Shuhao and Chen, Liming and Kechadi, Tahar",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.546/",
doi = "10.18653/v1/2026.acl-long.546",
pages = "11881--11910"
}