Evaluating Defeasible Reasoning in LLMs with DEFREASING

Emily Allaway, Kathleen McKeownNAACL 2025 · DOI

Abstract

While Large Language Models (LLMs) excel at many tasks, they struggle with defeasible reasoning—the ability to update a conclusion when presented with new, potentially conflicting information. This paper introduces DEFREASING, a massive benchmark of ~95,000 questions, revealing that even state-of-the-art models fail to consistently reason about how new evidence strengthens, weakens, or simply has no impact on a conclusion.

The Challenge of Non-Monotonic Reasoning

In everyday life, most of our inferences are "defeasible." If a person is told "birds fly," they reasonably infer that a specific sparrow will fly. However, if they are then told "that sparrow has a broken wing," their confidence in the conclusion should drop. This is non-monotonic reasoning: new information can overturn previous conclusions.

An illustration of defeasible reasoning: how new information can strengthen, weaken, or contradict a conclusion drawn from a generic rule.

Current LLM evaluations often focus on formal deductive logic (where conclusions are absolute) or commonsense benchmarks that conflate a model's reasoning ability with its memorized world knowledge. The authors identify a gap in evaluating how LLMs handle property inheritance from generics (e.g., "Seagulls steal food"). They aim to determine if LLMs can truly reason through the impact of new evidence or if they are merely relying on shallow heuristics and pre-trained associations.

The DEFREASING Benchmark

The authors formulate defeasible reasoning as a task adjacent to Natural Language Inference (NLI). Given an initial premise (a generic rule) and a conclusion, the model must predict how "additional information" impacts the belief in that conclusion: does it strengthen, weaken, or have no impact?

To build the DEFREASING dataset, five distinct reasoning patterns were utilized:

  1. S-case (Strengthening): Providing evidence that other subtypes share the property.
  2. S-alt (Strengthening - Alternate): Showing that diverse, related concepts also have the property.
  3. W-case (Weakening): Providing counter-examples (exceptions) to the general rule.
  4. N-alt (No Impact - Alternate): Providing information about an unrelated concept.
  5. N-prop (No Impact - Property): Providing information about a different, irrelevant property.

Crucially, the authors introduced nonsense types (e.g., "A Wumox is a cat") alongside real types. This allows them to "stress test" the models: if a model succeeds on real animals but fails on "Wumoxes," it is likely relying on memorized facts rather than actual logical reasoning.

Key Insights and Model Performance

The researchers evaluated 12 instruction-tuned models, including Llama 3, Mixtral, and GPT-4o. The results highlighted several critical weaknesses in current AI architectures:

  • Monotonicity Bias: Models are generally better at recognizing information that strengthens a conclusion (monotonic) than information that weakens it (non-monotonic).
  • The "Irrelevant" Trap: Most models struggle significantly with "No Impact" cases. They seem to operate under a bias that any new information provided in a prompt must be relevant, leading them to incorrectly guess that irrelevant facts strengthen or weaken the conclusion.
  • Shallow Heuristics: A correlation was found between the presence of the word "not" in premises and models predicting a "weakens" label, suggesting models use keywords as shortcuts rather than processing semantic meaning.
  • World-Knowledge Over-reliance: Models like Zephyr showed a 16-point performance drop when switching from real animals to nonsense types, proving they are "anchored" to training data and struggle with abstract logic.
Performance of various LLMs across the DEFREASING benchmark categories, showing accuracy and F1 scores.

Results Summary

As shown in the evaluation table, no model performed well across all categories. The best overall F1 score was only ~0.64, achieved by Llama 3. Models that were nearly perfect at "Strengthening" cases (like Starling) were often abysmal at "No Impact" cases. Conversely, while GPT-4 and GPT-4o were highly accurate on "No Impact" examples, they exhibited a massive bias, predicting "no impact" for approximately 60% of all cases, which severely degraded their performance on strengthening and weakening tasks.

Conclusion and Future Work

The DEFREASING benchmark reveals that LLMs still lack a robust capacity for defeasible reasoning. While the dataset currently focuses on property inheritance in English, it provides a vital tool for identifying where models rely on memorized facts versus logical processing. Future work should investigate defeasible reasoning in causal or temporal contexts and seek architectures that better disentangle world knowledge from the reasoning process.

Summary updated

BibTeX
@inproceedings{allaway2025evaluating,
  author = {Emily Allaway and Kathleen McKeown},
  title = {{Evaluating Defeasible Reasoning in LLMs with DEFREASING}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.naacl-long.529},
}
Figure