Making Large Language Models More Flexible to Changing Contingencies through Tracking Induction

Abstract

Abstract Recent research has shown that large language models (LLMs) exhibit rule-based insensitivity to contingencies, a functional analog of a well-documented phenomenon in human behavior. The present study extended these findings by examining whether tracking inductions can attenuate this insensitivity. Six frontier LLMs (GPT 5.2, Gemini 3 Flash, Grok 4.1 Fast, DeepSeek V3.2, Qwen3 235B, and Kimi K2) completed a modified rock-paper-scissors paradigm with manipulated outcome probabilities in a 3 (Rule: No Rule, Correct Rule, Incorrect Rule) × 3 (Induction: No Induction, Direct Tracking, Metaphorical Tracking) factorial design across 2,160 sessions. All models demonstrated In-Context Learning and strong rule adherence. Replicating and extending prior findings, rule provision produced significant contingency insensitivity across all six models. Tracking inductions significantly reduced this insensitivity when rules were provided, while having minimal effects in the absence of rules. Direct tracking inductions consistently enhanced contingency sensitivity across all models. Metaphorical tracking showed greater variability, being equally or more effective than direct tracking in some models, but ineffective or paradoxically detrimental in others. These findings demonstrate that prompt-based tracking inductions can modulate rule-based contingency insensitivity in LLMs, with implications for understanding and improving the behavioral flexibility of artificial agents.