Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure x Register Factorial
Researchers investigated how large language models respond to different types of interventions during a state of functional collapse. Using the Qwen3.5-4B model, they found that attention was primarily driven by lexical surprise, with scrambled messages capturing the most attention. However, behavioral responses were significantly influenced by relational interventions, particularly when combined with a first-person register. AI
IMPACT This research offers insights into how LLMs process and respond to different communication styles, potentially informing future AI safety and interaction design.