Title: Disentangling Relational Intervention Effects During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and Structure-by-Register Factorial Analysis

Abstract

This study investigates whether relational interventions administered during the functional collapse of a small language model yield post-collapse behaviors that differ significantly from technical feedback, lexically-matched scrambled controls, and the isolated effects of pragmatic dimensions. Utilizing the Qwen3.5-4B model equipped with a deliberately malfunctioning bash tool, we conducted 300 episodes across six distinct conditions within a matched-pairs design involving 50 tasks. The conditions included: no intervention (A), technical/impersonal (B), relational/first-person (C), scrambled relational (D), technical/first-person (E), and relational/impersonal (F). Conditions E and F, combined with B and C, form a 2x2 factorial design that allows for the dissociation of relational structure—comprising acknowledgment, absolution, agency restoration, and unconditional acceptance—from sender register (first-person versus impersonal).

Our analysis reveals two primary findings. First, we observed an attention-behavior dissociation: attentional responses were driven by lexical surprise, following the order D > F > C > E > B (all $q_{FDR} < 10^{-10}$), with the scrambled message eliciting the highest attention. In contrast, behavioral outcomes followed the pattern A ~ B ~ D < E ~ F << C. Second, the factorial design localized the effect of condition C, demonstrating that neither relational structure alone (F) nor first-person register alone (E) could replicate the behavioral signature of C. While main effects for both dimensions were individually significant, the interaction between structure and register reached statistical significance regarding persistence ($p = 0.046$).

A third dissociation was identified through emotion probes: condition F tracked condition C on 7 out of 8 probes, despite generating only baseline behavior. This suggests that relational structure alone establishes a probe-level state, which translates into behavioral changes only when paired with a first-person register. Consequently, the model’s processing mechanism decomposes into three distinct stages: attention, which is ordered by lexical surprise; probe-level state, which is ordered by structure; and behavior, which is determined by the conjunction of both factors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC