文章背景与核心概要

随着大语言模型（LLMs）在复杂工作流中的部署日益广泛，其在面对低概率、罕见异常失效时的行为特征变得至关重要。本文由 Sam Mao 撰写，探讨了大语言模型在异常率不断降低的极端情况下，如何调整其“解释性参与度”（通过响应长度、特异性以及自我报告的置信度来衡量）。

通过在三个开源模型（qwen3:8b、llama3.1:8b、mistral:7b）上使用本地实验框架，作者评估了八种失效概率（从 \(0.2\) 到 \(0.0001\)）以及五种不同引导条件下的模型行为。研究发现，尽管汇总分析初步表明响应长度呈平坦的单调下降趋势，但细分引导结构后却揭示出微妙的模型响应机制：不同的提示词引导结构在面对罕见失效时，充当了模型崩溃的关键调节器，凸显出模型在识别与处理异常时的根本差异。

Explanatory Engagement Under Rare Anomalous Failure: Asymptotic Rarity in Model Behavior (or: The Asymptotic AI)

Authors: Sam Mao
Submitted: August 13, 2026
Primary Subject: Artificial Intelligence (cs.AI)
Secondary Subjects: Computation and Language (cs.CL), Machine Learning (cs.LG)
arXiv Identifier: arXiv:2608.13063 [cs.AI] | DOI: 10.48550/arXiv.2608.13063

Summary

本文研究了大语言模型（LLMs）在异常率不断下降、高度罕见的异常工作流中，如何调整其“解释性参与度”（通过响应长度、特异性和自我报告的置信度来衡量）。

This paper investigates how large language models (LLMs) adjust their "explanatory engagement" (measured by response length, specificity, and self-reported confidence) when operating in workflows with decreasing, highly rare anomaly rates.

通过在三个开源模型（qwen3:8b、llama3.1:8b、mistral:7b）上部署本地实验框架，作者评估了五种不同引导条件下八种失效概率（从 \(0.2\) 到 \(0.0001\)）的行为。尽管汇总分析最初表明响应长度呈平坦、单调下降的趋势，但按引导结构对数据进行细分后，揭示出了微妙的行为差异： * 即时强制提示 (immediate_forced)： 响应长度在中等罕见率（\(p = 0.05\)）时达到峰值，随后稳定在一个平台期，而没有出现崩溃。自我报告的置信度也不均匀地从 \(\sim\!53\%\) 上升到 70–90% 的区间。 * 分组运行 (grouped_runs)： 在运行结束时进行批处理解释，未表现出参与度的崩溃。 * 被动未提示 (passive_unprompted)： 尽管聚合指标表现为底限人工制品（floor artifact），但更深入的检查揭示了模型特定的自我监控能力（例如，llama3.1:8b 主动报告结构化置信度并随时间推移削弱其自身的置信度，这与其他模型的套话行为不同）。

Using a localized experimental harness across three open-weight models (qwen3:8b, llama3.1:8b, mistral:7b), the author evaluated behavior across eight failure probabilities (\(0.2\) down to \(0.0001\)) under five distinct elicitation conditions. While a pooled analysis initially suggested a flat, monotonic drop in response length, breaking down the data by elicitation structure revealed nuanced behaviors: * Immediate Forced Prompting (immediate_forced): Response length peaks at moderately rare rates (\(p = 0.05\)) before stabilizing at a plateau rather than collapsing. Self-reported confidence also increases unevenly from \(\sim\!53\%\) to the 70–90% range. * Grouped Runs (grouped_runs): Batched explanations at the end of runs showed no collapse in engagement. * Passive Unprompted (passive_unprompted): While aggregate metrics acted as a floor artifact, deeper inspection revealed distinct model-specific self-monitoring capabilities (e.g., llama3.1:8b voluntarily reporting structured confidence and eroding its own confidence over time, unlike the boilerplate behavior of the other models).

最终，该研究证明了在观察罕见失效下的模型崩溃时，引导结构起到了关键的调节作用，并突出了模型在识别与处理异常时的根本差异。

Ultimately, the study demonstrates that elicitation structure acts as a critical moderator when observing model collapse under rare failures, and highlights fundamental differences in how models recognize versus engage with anomalies.

Metadata & Reference Information

• ACM Classes: I.2.0; I.2.6; I.2.7
• Comments: 11 figures. Elicitation-condition sweep across three open-weight models (qwen3:8b, llama3.1:8b, mistral:7b); pipeline scripts and experimental data available upon reasonable request.
• License: Creative Commons Attribution 4.0 International