Teaching Models to Dream of Better Monitors through Evaluation Conditioned Training — LessWrong
lesswrong.com · 4,288 words · saved by 1 readers
Edit: we renamed the technique from Monitor Sensitive Training (MST) to Evaluation Conditioned Training (ECT)[1] …
x Teaching Models to Dream of Better Monitors through Evaluation Conditioned Training — LessWrong Outer Alignment Scalable Oversight AI Frontpage 49 Teaching Models to Dream of Better Monitors through Evaluation Conditioned Training by Alec Harris , Kasey C , Archie Chaudhury , yix 19th Mar 2026 12 min read 2 49 Edit: we renamed the technique from Monitor Sensitive Training (MST) to Evaluation Conditioned Training (ECT) [1] TL;DR We introduce Evaluation Conditioned Training (ECT), a new post-training technique where we augment training data with evaluation labels that describe how evaluation i
saved by
related reading
- Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimesarxiv.org
- Should We Train Against (CoT) Monitors? — LessWronglesswrong.com
- Constitutional AI: Harmlessness from AI Feedbackarxiv.org
- [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
- Model Spec Midtraining: Improving How Alignment Training Generalizesalignment.anthropic.com
- Training on Documents about Reward Hacking Induces Reward Hackingalignment.anthropic.com
- The Most Forbidden Technique — LessWronglesswrong.com
- Alignment Faking Mitigationsalignment.anthropic.com
- [2203.02155] Training language models to follow instructions with human feedbackarxiv.org
- How hard is it to inoculate against misalignment generalization? — LessWronglesswrong.com
- How far does alignment midtraining generalize?alignment.openai.com
- (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL — LessWronglesswrong.com