ceselder/cot-oracle: Our improvements to activation oracles ·

Our improvements to activation oracles

CoT Oracle CoT Oracle is a white-box chain-of-thought monitor built on Activation Oracles. The core model is Qwen3-8B with a LoRA adapter trained to read its own residual-stream activations and answer questions about the reasoning that produced them. This README documents the pipeline the current code actually runs. Older docs elsewhere in the repo, especially under src/evals/ , describe older or parallel experiments and are not the main training-time path. Core Mechanism A source sequence is built from a question plus a chain-of-thought or other context text. Activation positions are chosen f

related reading

• cot-oracle/AObench at main · ceselder/cot-oracle · GitHubgithub.com
• Current activation oracles are hard to use — LessWronglesswrong.com
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• [2606.02609] Building Better Activation Oraclesarxiv.org
• [2606.02609] Building Better Activation Oraclesarxiv.org
• Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers — LessWronglesswrong.com
• Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainersalignment.anthropic.com
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• [2512.15674] Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainersarxiv.org
• Dreaming Vectors: Gradient-descented steering vectors from Activation Oracles and using them to Red-Team AOs — LessWronglesswrong.com
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
• Composer2.pdfcursor.com