MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset

2024-06-04

arXiv.org

AI Summary

This paper addresses the weak metaphysical reasoning capability of large language models (LLMs) when confronting contextual shifts—i.e., distributional dynamics induced by environmental changes or agent behaviors. To this end, it proposes a novel “metaphysical reasoning” framework, formalizing a three-level discriminative reasoning paradigm: behavior change → state change → contextual shift. Building upon this, the authors introduce MARS, the first multi-task benchmark specifically designed to evaluate this capability. MARS comprises a human-annotated and controllably generated dataset with fine-grained contextual-shift labels, structured via a conceptual taxonomy. Through systematic evaluation across model scales and training paradigms, empirical analysis reveals significant deficiencies in this reasoning chain across 20 mainstream (L)LMs. Further experiments demonstrate that integrating large-scale conceptual classification knowledge during pretraining substantially improves performance.

Technology Category

Cognitive Modeling & Cognitive Systems: Conceptual Inference and ReasoningSearch and Optimization: Metareasoning and MetaheuristicsMachine Learning: Large Multimodal Models (LMMs)

Application Category

Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactionsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

To enable Large Language Models (LLMs) to function as conscious agents with generalizable reasoning capabilities, it is crucial that they possess the reasoning ability to comprehend situational changes (transitions) in distribution triggered by environmental factors or actions from other agents. Despite its fundamental significance, this ability remains underexplored due to the complexity of modeling infinite possible changes in an event and their associated distributions, coupled with the lack of benchmark data with situational transitions. Addressing these gaps, we propose a novel formulation of reasoning with distributional changes as a three-step discriminative process, termed as MetAphysical ReaSoning. We then introduce the first-ever benchmark, MARS, comprising three tasks corresponding to each step. These tasks systematically assess LLMs' capabilities in reasoning the plausibility of (i) changes in actions, (ii) states caused by changed actions, and (iii) situational transitions driven by changes in action. Extensive evaluations with 20 (L)LMs of varying sizes and methods indicate that all three tasks in this process pose significant challenges, even for state-of-the-art LLMs and LMs after fine-tuning. Further analyses reveal potential causes for the underperformance of LLMs and demonstrate that pre-training them on large-scale conceptualization taxonomies can potentially enhance their metaphysical reasoning capabilities. Our data and models are publicly accessible at https://github.com/HKUST-KnowComp/MARS.

Problem

• Assessing LLMs' ability to reason about situational distribution changes
• Evaluating comprehension of action and state transition plausibility
• Addressing lack of benchmarks for metaphysical reasoning in AI

Innovation

• Three-step discriminative process for reasoning
• MARS benchmark with three assessment tasks
• Pre-training on conceptualization taxonomies enhances reasoning