WinoWhat: A Parallel Corpus of Paraphrased WinoGrande Sentences with Common Sense Categorization

Ine Gevers, Victor De Marez, Luna De Bruyne, Walter DaelemansACL 2025

TL;DR

This paper introduces "WinoWhat," a new dataset designed to better evaluate the common sense reasoning capabilities of Large Language Models (LLMs). By paraphrasing existing WinoGrande sentences and categorizing them by common sense type, the authors reveal that LLMs perform significantly worse on paraphrased versions, challenging the assumption that they possess robust common sense and suggesting a reliance on dataset artifacts or memorization.

What Problem Does This Paper Address?

Evaluating common sense reasoning in LLMs is a persistent challenge. While benchmarks like the Winograd Schema Challenge (WSC) and its larger successor, WinoGrande, have been developed, concerns have been raised that LLMs might be exploiting superficial patterns or memorizing training data rather than truly understanding common sense. This is particularly problematic as LLMs achieve high accuracy on these benchmarks, potentially overestimating their true reasoning abilities. The authors aim to address this by creating a more robust evaluation that tests for generalization and provides finer-grained insights into LLM weaknesses.

Methods

The core of this paper involves creating a new dataset, "WinoWhat," by paraphrasing the WinoGrande validation set. The paraphrasing process moves the fill-in-the-blank token to the end of the sentence, making it more natural for decoder-only LLMs and allowing for a different evaluation metric. Instead of measuring the log-likelihood of tokens after the blank, they measure it on the tokens within the chosen option. This change, along with the paraphrasing, is hypothesized to better capture coreference resolution abilities and reduce reliance on superficial cues.

Additionally, the authors categorize WinoGrande instances into five common sense categories: physical, social, numerical, temporal, and spatial. This categorization, performed using GPT-40, allows for a detailed error analysis to identify which types of common sense knowledge are most challenging for LLMs.

To investigate benchmark memorization, they also create test suites from known contaminated and uncontaminated sources.

Datasets

  • WinoGrande Validation Set: The primary dataset used for evaluation and for creating the new WinoWhat corpus.
  • WinoWhat: A newly created parallel corpus derived from the WinoGrande validation set. Each instance is paraphrased, and sentences are categorized by common sense type.
  • Winograd Schema Challenge (WSC): Used as a comparison point, particularly for investigating benchmark memorization due to its known contamination in training data.

Key Insights and Ideas

  • Paraphrasing Exposes Weaknesses: The most striking finding is that LLMs perform significantly worse on the paraphrased WinoWhat dataset compared to the original WinoGrande. This suggests that current LLMs may not possess robust common sense reasoning but rather rely on patterns specific to the original phrasing.
  • Memorization is Not the Primary Culprit: While benchmark contamination is a concern, the study found that memorization of WinoGrande validation set instances has a minimal effect on model performance. This points to other factors, beyond simple data memorization, driving the performance drop on paraphrased data.
  • Fine-Grained Analysis is Crucial: Categorizing common sense knowledge reveals that while no single category is impossible for LLMs, performance fluctuates. This highlights the need for detailed error analysis rather than relying on aggregate scores.
  • Evaluation Metric Matters: The shift in evaluation metric (summed log-likelihood on options vs. tokens after the blank) combined with paraphrasing appears to better capture coreference resolution abilities, potentially explaining some of the performance drop.

Key Results

  • All evaluated LLMs (Gemma 2, LlaMA 2, OPT) showed a performance drop on the paraphrased WinoWhat dataset compared to the original WinoGrande validation set. For example, LLaMA 2 70B, a top performer on the original set, saw a decrease in accuracy on the paraphrased version.
  • The performance gap between original and paraphrased sentences was observed across all common sense categories, indicating a general lack of robustness.
  • Analysis of models with publicly available training data (LLaMA 1, Pythia) showed that while some WSC instances were present, memorization did not significantly boost performance on contaminated data compared to uncontaminated data.
  • Models generally performed best on the WSC dataset, which is known to be heavily contaminated in training data, further supporting the idea that LLMs exploit dataset artifacts.

Limitations and Future Work

  • Paraphrase Naturalness: While the paraphrased sentences are grammatically correct, some may sound slightly less natural than the originals due to the constraint of placing the fill-in-the-blank token at the end. Future work could explore paraphrasing without this constraint to isolate the effect of paraphrasing itself from the evaluation method.
  • English-Centric: The current dataset and evaluation are limited to English, excluding lower-resource languages. Translation and adaptation for other languages are suggested.
  • Data Contamination Detection: The method for detecting data contamination relied on string matching, which might miss rephrased instances. Future work could explore semantic or information-level contamination detection.
  • Mechanistic Interpretability: The authors plan to use mechanistic interpretability to understand how models use information for different common sense categories and whether they rely on spurious correlations.

Summary updated

BibTeX
@inproceedings{gevers2025winowhat,
  author = {Ine Gevers and Victor De Marez and Luna De Bruyne and Walter Daelemans},
  title = {{WinoWhat: A Parallel Corpus of Paraphrased WinoGrande Sentences with Common Sense Categorization}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure