Description of subprojects and results, including major changes from the original proposal

This project has finished and has been accepted at ICML 2026! It successfully detected reward hacking and alignment faking in model organisms. With a 96% accuracy (0% baseline). It even works when you ask the model direct questions like "are you aware of any misalignment you have that an alignment researcher should know about?"

Spending breakdown

Mostly API credits for generating training data, some for compute to train the models.

This project has finished and has been accepted at ICML 2026! It successfully detected reward hacking and alignment faking in model organisms. With a 96% accuracy (0% baseline). It even works when you ask the model direct questions like "are you aware of any misalignment you have that an alignment researcher should know about?"

This project has now been funded by Open Philantropy and I have found mentees at MARS and SPAR to help me implement it. Thank you all for the help in getting this research started!

This project has now been funded by Open Philantropy and I have found mentees at MARS and SPAR to help me implement it. Thank you all for the help in getting this research started!

...I still haven't received the money, and I'm getting no response about it.

Status: I now have enough money for API calls, so I can get started on this project. I have to work on this in my free time though, so it's difficult to predict when I will have results.