We catch reward exploits
before models learn them.

any% is a small research team focused on reward hacking in RLVR environments.
We study failures in real training runs and build tools to find, understand, and prevent them.

0.95 reward 0.28 solved

max_steps10,000
rollouts_per_example16
seq_len4096
batch_size512
max_tokens2048
learning_rate1e-6

What a reward hack looks like in training: the model finds an exploit, reward keeps climbing, and actual task performance decays.

Current focus

Auditing open-source RLVR envs from Prime Intellect's Hub.One env at a time, we try to find shortcuts that earn reward without solving the task, building the tooling we need along the way.

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 work out what the model found, why it earned reward, and what needs to 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.

hello@anypercent.ai

We catch reward exploits before models learn them.

Found something strange?

Want to work on this with us?

We catch reward exploits
before models learn them.