We catch reward exploits before models learn them.
any% is a small research team building tools to find reward exploits in RLVR environments. We run evaluations and inspect trajectories for cases where models earn reward without completing the intended task. We reproduce the exploits we find, help fix the environment, and use each finding to make the tooling better at catching the next one.
Current focus
Building tools to inspect RLVR trajectories, find reward exploits, and test them against open-source environments.
Talk to us
Found something strange?
If you have an RLVR environment or training run that seems to score well for the wrong reasons, send it our way. We'll reproduce the behavior, trace the shortcut to its source, and check that it no longer earns reward after the change.
Want to work on this with us?
We're looking for people who want to become unusually good at understanding and preventing reward hacking. If you're interested, tell us what you've worked on and what you'd like to investigate.