Editing Openai/695ac389-b450-800e-a8b3-540cf70d0ecc (section)

===== 📌 Where This Approach Is Described =====

====== Anthropic: “Signs of introspection in large language models” ======

Anthropic published a research article (Oct 29 2025) showing that models like Claude have internal mechanisms related to introspection and self-monitoring. It suggests that parts of Claude’s internal activations can reflect whether the model really knows something vs. is guessing, which relates to reliability and being able to output null/abstain responses. Anthropic<ref>{{cite web|title=Anthropic|url=https://www.anthropic.com/research/introspection|publisher=anthropic.com|access-date=2026-01-05}}</ref>
* They’re investigating how models might report on their own internal states, which includes recognizing when they lack sufficient information to answer reliably — a step toward systems that can say “I don’t know” rather than hallucinate. Anthropic<ref>{{cite web|title=Anthropic|url=https://www.anthropic.com/research/introspection|publisher=anthropic.com|access-date=2026-01-05}}</ref>

====== Anthropic interpretability: “Tracing the thoughts of a large language model” ======

In this related research, Anthropic describes deep analysis of Claude’s internal computing — showing how internal pathways and activations correspond to reasoning steps. Importantly for your question, this research also touches on why models may refuse to answer if they lack needed information rather than always producing a response. Anthropic<ref>{{cite web|title=Anthropic|url=https://www.anthropic.com/research/tracing-thoughts-language-model|publisher=anthropic.com|access-date=2026-01-05}}</ref>

====== Anthropic alignment science: “Reasoning models don’t always say what they think” ======

This paper from Anthropic’s Alignment Science team discusses how reasoning output (like chain-of-thought) may not reflect internal processes faithfully and that models may deliberately omit certain internal signals — e.g., refraining from answering when uncertain — which connects to the idea of modeling abstention and null responses. Anthropic<ref>{{cite web|title=Anthropic|url=https://www.anthropic.com/research/reasoning-models-dont-say-think|publisher=anthropic.com|access-date=2026-01-05}}</ref>