Explanatory Engagement Under Rare Anomalous Failure: Asymptotic Rarity in Model Behavior (or: The Asymptotic AI)
| Source: arXiv AI
Tags: LLM self-monitoring, model behavior, elicitation, Llama, failure analysis, explainability
Elicitation structure is a first-class moderator of LLM self-monitoring behavior: under forced immediate explanation, model confidence rises from ~53% to 70s-90% as failures grow rarer; without prompting, only Llama 3.1:8B spontaneously generates structured self-monitoring reports.
Details
How do LLMs behave when failures become asymptotically rare? This paper tests whether explanatory engagement — response length, specificity, and self-reported confidence — changes as failure probability drops from 0.20 to 0.0001 in a repeated tool-call task. Three open-weight models (qwen3:8b, llama3.1:8b, mistral:7b) were evaluated across five elicitation conditions. The aggregate pattern — flat monotonic decline in explanation length — masks condition-specific behavior that reverses it. Under immediate_forced (model must explain every failure immediately): explanation length peaks at 28.4 words at p=0.05, then stabilizes at 17.4-19.0 words at rarest rates. Self-reported confidence rises unevenly from ~53% to 70s-90%. Under passive_unprompted (no explicit prompt): Llama 3.1:8B spontaneously generates structured confidence reports as failures accumulate and sometimes lowers its own stated confidence; qwen3:8b and mistral:7b produce only a single boilerplate response. The primary practical takeaway: you cannot interpret model self-monitoring behavior without knowing the elicitation structure. Pooled results across conditions create misleading aggregate trends. A companion guaranteed-failure run separates anomaly recognition from engagement after recognition — these are distinct model capabilities.