cd/entity/AObench· home› entities› AObench
grep -l @aobench /news/*.json | wc -l → 1

@AObench

mentions 1 type Organization feed RSS
18:34
2026-06-04
lesswrong.com
artificial-intelligence

Building Better Activation Oracles

Researchers have improved Activation Oracles (AOs)—fine-tuned LLMs that answer natural language questions about a target model's internal activations—by training on on-policy rollouts, using a higher-…

// co-occurs with top 7 entities
Neel Nanda 1 Adam Karvonen 1 Niclas Luick 1 Karvonen et al. 1 MATS 1 Activation Oracle 1 Qwen3-8B-AO-v3 1