01 / Key message
InFact: A Strong Baseline for Automated Fact-Checking
A six-stage, retrieval-grounded fact-checker that won the 2024 AVeriTeC shared task and set a strong text-only baseline.
Retrieval-grounded verification
02 / Method
How it works
Structure the claim
Interpret the claim and generate focused questions that turn verification into a tractable investigation.
Search the web
Retrieve and organize external evidence that directly addresses the generated questions.
Resolve the verdict
Reason over the gathered evidence and return a supported, refuted, or insufficient-evidence conclusion.
03 / Abstract
Abstract
The spread of disinformation creates a need for robust and scalable automated fact-checking systems. InFact is an LLM-based approach for the AVeriTeC Shared Task Challenge 2024 that decomposes text-claim verification into a six-stage process including evidence retrieval. With GPT-4o as its backbone, InFact achieves an AVeriTeC score of 63% on the test set, outperforming the other 20 participating teams and establishing a strong baseline for text-only automated fact-checking. Its qualitative analysis also identifies cases where the system's conclusion is more accurate than the benchmark's human-annotated ground truth.
04 / Contributions
What this adds
- 01
Challenge-winning baseline
Ranks first among 21 systems in the 2024 AVeriTeC shared task with a 63% test-set score.
- 02
Evidence-first workflow
Turns claim verification into six explicit stages, including live evidence retrieval rather than memory-only prediction.
- 03
Ground-truth diagnosis
Uses qualitative error analysis to expose benchmark cases where the automated conclusion may be better supported than the label.
05 / Citation
Citation
Rothermel, M., Braun, T., Rohrbach, M., & Rohrbach, A. (2024). InFact: A Strong Baseline for Automated Fact-Checking. Proceedings of the Seventh Fact Extraction and VERification Workshop (FEVER), 108–112.
BibTeX
@inproceedings{rothermel2024infact,
title = {{InFact}: A Strong Baseline for Automated Fact-Checking},
author = {Mark Rothermel and Tobias Braun and Marcus Rohrbach and Anna Rohrbach},
booktitle = {Proceedings of the Seventh Fact Extraction and VERification Workshop (FEVER)},
pages = {108--112},
address = {Miami, Florida, USA},
publisher = {Association for Computational Linguistics},
year = {2024},
doi = {10.18653/v1/2024.fever-1.12},
url = {https://aclanthology.org/2024.fever-1.12/}
}