LogiQAte

This benchmark is introduced in the paper:

“Don’t Judge a Book by its Cover: Testing LLMs’ Robustness Under Logical Obfuscation”
Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026)

Paper Link: https://arxiv.org/abs/2602.01132

LogiQAte is a diagnostic benchmark designed to evaluate whether large language models truly reason or merely rely on surface pattern matching.

The benchmark tests models under logical obfuscation: rewriting problems into logically equivalent but structurally different forms.

If a model genuinely understands the logic, its performance should remain stable.
In practice, we find that performance drops sharply, exposing brittleness in SoTA LLMs.

---

Dataset Organization

Each reasoning task_type is released as a separate configuration in this HuggingFace repository.

All data is provided in the test split for evaluation.

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("abhilekhborah/LogiQAte", "task_type", split="test")
print(ds[0])

---

If you use LogiQAte, please cite:

@inproceedings{borah2026logiqate,
  title     = {Don’t Judge a Book by its Cover: Testing LLMs’ Robustness Under Logical Obfuscation},
  author    = {Borah, Abhilekh and Ghosh, Shubhra and Joshi, Kedar and Guru, Aditya Kumar and Ghosh, Kripabandhu},
  booktitle = {Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
  year      = {2026}
}