Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure × \times Register Factorial
Franco Santana, Horacio Vico · 2026
Read Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure × \times Register Factorial by Franco Santana, Horacio Vico (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00935.txt.
Description
Santana and Vico test kind words as a rescue for a failing AI agent in this 2026 paper. Their setup puts a small language model, Qwen3.5-4B, in a broken-tool trap. Its bash tool is deliberately wrecked, and the model falls into functional collapse. Mid-collapse, a message arrives: technical, relational, scrambled, or none.
They run 300 episodes in six conditions over 50 matched tasks. Two factors are crossed in a grid: relational structure and sender register, first-person or impersonal. Attention and behavior part ways. Attention chases lexical surprise, with the scrambled note drawing the most. Behavior ranks the true relational, first-person note far above the rest. Neither factor alone copies the full effect. Both main effects count, and their interplay is firm on persistence, at p of 0.046. Emotion probes split off in their own pattern. The study treats tone in agent feedback as a variable with real causal weight.
Catalog record
| CREATOR | Franco Santana, Horacio Vico |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.00935