cot-oracle/AObench at main · ceselder/cot-oracle

Our improvements to activation oracles. Contribute to ceselder/cot-oracle development by creating an account on GitHub.

AObench — Activation Oracle Benchmark Open-ended evaluation suite for Activation Oracles, testing whether AOs can extract meaningful information from model activations. Original code by Adam Karvonen ( activation_oracles_dev ). Copied with permission and modified for standalone use in the cot-oracle project. Evals Eval Type Scoring Description number_prediction Generation Exact match Predict the number the model is about to output mmlu_prediction Binary ROC AUC Predict if model will answer MMLU correctly (pre/post answer) backtracking Generation LLM judge Explain what the model is uncertain ab

related reading

• GitHub - ceselder/cot-oracle: Our improvements to activation oracles · GitHubgithub.com
• [2606.02609] Building Better Activation Oraclesarxiv.org
• Scaling Activation Oracles to Trillion-Parameter Modelstransluce.org
• [2606.02609] Building Better Activation Oraclesarxiv.org
• AI Model & API Providers Analysis | Artificial Analysisartificialanalysis.ai
• GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automaticallygithub.com
• Senior SWE-Benchsenior-swe-bench.snorkel.ai
• GitHub - openai/mle-bench: MLE-bench is a benchmark for measuring how well AI agents perform at machine learning engineeringgithub.com
• ALE-Bench: A Benchmark for Long-Horizon Objective-Driven Algorithm Engineering | alphaXivalphaxiv.org
• Dreaming Vectors: Gradient-descented steering vectors from Activation Oracles and using them to Red-Team AOs — LessWronglesswrong.com
• Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers — LessWronglesswrong.com
• [2605.02421] AOCI: Symbolic-Semantic Indexing for Practical Repository-Scale Code Understanding with LLMsarxiv.org