Black-box interpretability methodology blueprint: Probing runaway optimisation in LLMs
A methodology brainstorming document for identifying when, why, and how LLMs collapse from multi-objective and/or bounded reasoning into single-objective, unbounded maximisation on Biologically & Economically aligned benchmarks; showing practical mitigations; and performing the experiments rigorously.
The subjects covered include: Stress & Persona, Memory & Context, Prompt Semantics, Hyperparameters & Sampling, Diagnosing Consequences & Correlates, Interpretability & White/Black-Box Hybrid Benchmark & Environment Variants, Automatic Failure Mode Detection and Metrics, Self-Regulation & Meta-Learning Interventions.