Inference Stack Degradation Theory (v1.5)
作者发表的研究提案，讨论 LLM 推理中与负载耦合的数值漂移及 Prompt Weather Index 监测方法。该工作区分稳定性与正确性，并提出验证路径。
原始文档语言：英语。切换界面语言不会翻译原始文档。
Inference Stack Degradation Theory
Load-Coupled Numerical Drift and Monitoring in LLM Inference
Version 1.5
Roble Mumin
Independent AI Researcher and AI Consultant
https://roblemumin.com
December 2025
Inference Stack Degradation Theory v1.5
Contents
1 Introduction
2 Scope and Non-Claims
3 Established Findings from Prior Work
3.1 Floating-Point Non-Associativity
3.2 Batch Dependence as Practical Cause
4 Relation to ML Systems Research
5 Inference Stack Degradation Theory
5.1 Formal Framing
5.2 Null Hypothesis
6 Autoregressive Sensitivity Considerations
7 Prompt Weather Index
7.1 Motivation
7.2 Definition
7.3 Stability Versus Correctness
7.4 Falsifiability
8 Minimal Empirical Validation Path
9 Operational and Ethical Considerations
10 Implications if Validated
11 Open Research Questions
12 Conclusion
Roble Mumin Page 2 of
Inference Stack Degradation Theory v1.5
Experiments
13 Experiments
13.1 Experiment 1: Reproducing Batch-Induced Token-Level Divergence
13.1.1 Objective
13.1.2 Setup
13.1.3 Expected Metrics
13.1.4 Null Hypothesis
13.1.5 Agentic Execution Notes
13.2 Experiment 2: Measuring Autoregressive Drift Accumulation
13.2.1 Objective
13.2.2 Setup
13.2.3 Expected ISDT Outcome
13.2.4 Agentic Execution Notes
13.3 Experiment 3: Prompt Weather Index (PWI) Time-Series Monitoring
13.3.1 Objective
13.3.2 Setup
13.3.3 Null Hypothesis
13.3.4 Elaboration: Metric Sensitivity and Weight Robustness
13.3.5 Agentic Execution Notes
13.4 Experiment 4: Task-Level Correctness…
打开文档
引用此文档
Roble Mumin “Inference Stack Degradation Theory (v1.5)”.
2025-12.
访问地址： 文档页面
current summary qualified; independent peer review not verified