flâneur — a map of the web's best reading

Preliminary Thoughts on Reward Hacking

beren.io · 3,535 words · saved by 1 readers

Epistemic status: Early thoughts. Some ideas but no empirical testing or validation as yet. I’ve started thinking a fair bit about reward hacking recently. This is because frontier models are reportedly beginning to show signs of reward hacking especially for coding tasks. Thus, the era of easy-to-align pretraining-only models appears...

Epistemic status : Early thoughts. Some ideas but no empirical testing or validation as yet. I’ve started thinking a fair bit about reward hacking recently. This is because frontier models are reportedly beginning to show signs of reward hacking especially for coding tasks. Thus, the era of easy-to-align pretraining-only models appears to be coming to a close. Also in discussions and from my own experience, it seems to be that model’s reward hacking is one of the key bottlenecks that hinder recent reasoning RL models from continuing to improve is that they start to be able to hack the environm

Explore this link on the map →

related reading