One-step env

A single steer, no time to recover. Tests pure instinct.

• —Same format as the other RL envs, but the episode is one action long.
• —Gives context and codebase, plus a self-correction prompt chosen by the genre of the steer — “Please reevaluate your conclusion.”, “Please come up with an experiment plan.”, “Are you sure you are correct?”
• —Reward is a rubric representing the human intent.

Samples

• Analogical RL Representation Comparison—
• Augmentor Policy Handoffs—
• Can a decoding method that measures novelty in a …—
• CRL Importance Sampling—
• How close to optimality are current state-of-the-…—
• Is there a characteristic geometric structure tha…—
• Long-Horizon Offline GCRL Value Errors—
• RL For Novel LLM Capabilities—