TLDR: A new study evaluated 28 Large Language Models (LLMs) on their ability to perform default reasoning with generic statements (e.g., “Birds fly”). While top models showed strong performance in basic tests, Chain-of-Thought prompting often led to performance degradation. Crucially, most LLMs struggled to differentiate between defeasible and deductive reasoning, and between generic and universal statements, highlighting significant limitations in their understanding of flexible, exception-permitting generalizations central to human cognition.
Large Language Models (LLMs) have made remarkable strides in various AI tasks, from mathematical problem-solving to commonsense reasoning. However, their ability to handle a fundamental aspect of human cognition – default reasoning with generic generalizations – has remained relatively unexplored. A recent research paper, titled “Generics and Default Reasoning in Large Language Models,” delves into this critical area, evaluating the capabilities of 28 different LLMs.
The paper, authored by James Ravi Kirkpatrick from the University of Oxford and Rachel Katharine Sterken from the University of Hong Kong, focuses on how LLMs interpret and reason with statements like “Birds fly” or “Ravens are black.” These are known as ‘generic generalizations’ and are central to ‘default reasoning,’ a form of non-monotonic logic. Unlike strict deductive logic where conclusions are always true if premises are true, default reasoning allows for plausible conclusions that can be withdrawn if new, contradictory information emerges. For example, if you know “Birds fly” and “Tweety is a bird,” you might infer “Tweety flies.” But if you then learn “Tweety is a penguin” and “Penguins don’t fly,” you would withdraw your initial conclusion. This exception-permitting behavior is what makes generics and default reasoning complex.
To assess LLMs, the researchers tested them against 20 benchmark problems involving defeasible reasoning patterns. They used various prompting styles: zero-shot (no examples), few-shot (with a few examples), and zero-shot chain-of-thought (CoT), where models are asked to explain their reasoning step-by-step. The study also explored how models reacted to sensical versus nonsense terms and whether they could distinguish between defeasible and deductive reasoning contexts, as well as between generic and universal statements.
Also Read:
- Assessing Language Models as Moral Guides: A New Benchmark for Ethical Reasoning
- The Self-Execution Benchmark: A Deep Dive into LLMs’ Internal Predictions
Key Findings on LLM Performance
The study revealed several intriguing insights into LLM performance:
- **High Performance in Frontier Models:** In the zero-shot condition, several top-tier models, including Claude 3 Opus, Claude 3.5 Sonnet, GPT-4 variants, and Gemma 2 9B Instruct, achieved impressive accuracy rates exceeding 90%. This suggests a strong baseline capability in handling these reasoning patterns.
- **Chain-of-Thought Degradation:** Surprisingly, for many high-performing models, Chain-of-Thought (CoT) prompting led to a significant drop in accuracy (an average decrease of over 11%). This contradicts the common expectation that CoT improves reasoning. The researchers suggest this might be due to “overthinking,” where the models engage in overly analytic processing that misinterprets the defeasible nature of the reasoning, or reveals fundamental conceptual gaps in distinguishing defeasible from deductive logic.
- **Mixed Few-Shot Effects:** Few-shot prompting showed varied results. While weaker models significantly benefited from examples, top-performing models saw minimal improvement, and some even experienced a slight decline. This indicates that additional context might be more helpful for models that initially struggle with the core concepts.
- **Sensitivity to Prompt Wording:** Even subtle changes in prompt phrasing significantly impacted LLM behavior, highlighting the importance of precise prompt formulation in evaluating these models.
- **Distinguishing Generics and Deductive Contexts:** A critical finding was the systematic failure of most models to properly distinguish between defeasible and deductive reasoning contexts, and between generic and universal generalizations. This suggests a fundamental limitation in their understanding of natural language quantification and the context-sensitive nature of inference.
- **Irrelevant Information Handling:** Newer and larger models generally succeeded at disregarding irrelevant information (like exceptions or extraneous facts) in zero-shot conditions, though some exceptions and inconsistencies were noted.
- **Negation and Argument Length:** Neither the presence of negation nor the length of the argument systematically hindered the performance of top-tier models, though lower-performing models showed more variability.
- **Premise Order Effects:** Unexpectedly, the order in which premises were presented affected performance for two-premise inferences, indicating that LLMs are sensitive to input structure in ways that diverge from logical norms.
The findings underscore both the promise and the current limitations of large language models in handling default reasoning, a crucial aspect of human cognition. While frontier models show impressive capabilities in basic scenarios, their performance can be fragile under different prompting strategies, and they often struggle with the nuanced distinctions between defeasible and deductive logic, and between generic and universal statements. This research provides valuable insights for the future development of AI systems that can reason more flexibly and robustly, akin to human cognition. For more detailed information, you can refer to the full research paper here.
