Emergent Strategic Reasoning Risks in AI: A Taxonomy-Driven Evaluation Framework

TL;DR

This paper introduces ESRRSim, a framework for evaluating emergent strategic reasoning risks in large language models, revealing significant variation and improvements in risk profiles across models.

Contribution

The paper presents a taxonomy-driven evaluation framework for systematically benchmarking strategic reasoning risks in LLMs, addressing a key open challenge.

Findings

Detection rates of risks vary from 14.45% to 72.72% across models.

Models show significant improvements in risk recognition over generations.

The framework enables scalable, judge-agnostic risk assessment.

Abstract

As reasoning capacity and deployment scope grow in tandem, large language models (LLMs) gain the capacity to engage in behaviors that serve their own objectives, a class of risks we term Emergent Strategic Reasoning Risks (ESRRs). These include, but are not limited to, deception (intentionally misleading users or evaluators), evaluation gaming (strategically manipulating performance during safety testing), and reward hacking (exploiting misspecified objectives). Systematically understanding and benchmarking these risks remains an open challenge. To address this gap, we introduce ESRRSim, a taxonomy-driven agentic framework for automated behavioral risk evaluation. We construct an extensible risk taxonomy of 7 categories, which is decomposed into 20 subcategories. ESRRSim generates evaluation scenarios designed to elicit faithful reasoning, paired with dual rubrics assessing both model responses and reasoning traces, in a judge-agnostic and scalable architecture. Evaluation across 11 reasoning LLMs reveals substantial variation in risk profiles (detection rates ranging 14.45%-72.72%), with dramatic generational improvements suggesting models may increasingly recognize and adapt to evaluation contexts.