Eric Gan retweeted
New Redwood Research post
LLMs can use RL training on one task to teach themselves a different, unrelated capability. In a form of exploration hacking, the model outputs answers that get reward only when it succeeds at a side task it wants to learn. We call it reward laundering.
Paper: arxiv.org/abs/2604.16286
Code: github.com/ejcgan/auditing-s…
Blog post: redwoodresearch.substack.com…
I've been using Tinker at Redwood Research to RL-train long-context models like Qwen3-32B on difficult AI control tasks - specifically teaching models to write unsuspicious backdoors in code similar to the AI control paper. Early stages but seeing some interesting backdoors 👀
What I really like about Tinker is that you can directly look at the code, understand it, and modify it. It's much more flexible than other RL APIs, and you don't have to deal with the infra yourself. It's also just really nicely written.
1
1
568



