Don't Take the Premise for Granted: Evaluating the Premise Critique Ability of Large Language Models (2505.23715v1)

Abstract: LLMs have witnessed rapid advancements, demonstrating remarkable capabilities. However, a notable vulnerability persists: LLMs often uncritically accept flawed or contradictory premises, leading to inefficient reasoning and unreliable outputs. This emphasizes the significance of possessing the \textbf{Premise Critique Ability} for LLMs, defined as the capacity to proactively identify and articulate errors in input premises. Most existing studies assess LLMs' reasoning ability in ideal settings, largely ignoring their vulnerabilities when faced with flawed premises. Thus, we introduce the \textbf{Premise Critique Bench (PCBench)}, designed by incorporating four error types across three difficulty levels, paired with multi-faceted evaluation metrics. We conducted systematic evaluations of 15 representative LLMs. Our findings reveal: (1) Most models rely heavily on explicit prompts to detect errors, with limited autonomous critique; (2) Premise critique ability depends on question difficulty and error type, with direct contradictions being easier to detect than complex or procedural errors; (3) Reasoning ability does not consistently correlate with the premise critique ability; (4) Flawed premises trigger overthinking in reasoning models, markedly lengthening responses due to repeated attempts at resolving conflicts. These insights underscore the urgent need to enhance LLMs' proactive evaluation of input validity, positioning premise critique as a foundational capability for developing reliable, human-centric systems. The code is available at https://github.com/MLGroupJLU/Premise_Critique.

Evaluating Premise Critique Ability in LLMs

The paper "Don't Take the Premise for Granted: Evaluating the Premise Critique Ability of LLMs" addresses a critical issue within the domain of LLMs — their proclivity to overlook flawed or contradictory premises during reasoning processes. This tendency can lead to inefficiencies and unreliable outputs, which poses significant concerns over their use in human-centric applications. The authors propose the concept of Premise Critique Ability as an essential competency for LLMs, advocating for proactive identification and articulation of errors in the input premises. To systematically assess this ability, the Premise Critique Bench (PCBench) is introduced, a robust benchmark that incorporates four types of premise errors across three difficulty levels, paired with detailed evaluation metrics.

The paper presents a thorough experimental evaluation of 15 representative LLMs, encompassing both reasoning and non-reasoning models. Their findings suggest that most models rely heavily on explicit prompts to detect errors, exhibiting limited autonomous critique capabilities. Furthermore, the paper reveals a complex interplay between question difficulty, error type, and premise critique ability — models excel in identifying direct contradictions but struggle with intricate inconsistencies or procedural errors. Importantly, the paper highlights that reasoning ability does not consistently correlate with premise critique performance, with some models identifying inconsistencies internally but failing to articulate them outwardly.

Key quantitative results include a significant disparity between the Proactive Premise Critique Rate and Assisted Premise Critique Rate across models, indicating their tendency to more efficiently recognize errors when directed to do so. Additionally, flawed premises trigger overthinking in reasoning models, resulting in markedly elongated responses due to repeated attempts at resolving conflicts, as quantified by the Proactive Premise Critique Cost Ratio (PPCCR) and Assisted Premise Critique Cost Ratio (APCCR).

The paper makes important contributions by demonstrating critical gaps in LLMs' proactive evaluation of input validity and positing premise critique as a foundational capability for developing reliable AI systems. It also underscores the potential for future advancements in AI through the integration of explicit premise critique objectives within LLM training paradigms or by designing datasets with intentionally flawed premises. These advances could enhance LLMs' capacity to convert internal detection of premise flaws into explicit, actionable critiques.

In terms of implications, this research presents practical insights into AI application development, emphasizing the need for models that function as active evaluators rather than passive information providers. Theoretically, it opens avenues for exploring advanced reasoning training techniques that prioritize proactive premise evaluation alongside general accuracy and coherence.

This paper's findings contribute significantly to the comprehensive understanding of LLM vulnerabilities, providing a valuable framework for enhancing the reliability of AI systems in diverse applications. As AI continues to evolve, adopting a paradigm that incorporates premise critique capabilities will be crucial for fostering AI assistants that are truly trustworthy and dependable. Future studies should aim to expand the scope of premise critique evaluation across more models, languages, and domains to generalize findings and further refine AI methodologies.