Empirical outcome prediction

Would this experiment have been worth the GPUs?

• —A good researcher knows roughly what an experiment returns before paying for it. That prior is why human curation of GPU time beats brute-force search.
• —Prior work measures this only at paper granularity and only on directions that worked. Our transcripts contain the decision points themselves: a result judged insufficient, the frame turned to instead, and the roads declined.
• —Two tasks off the same record — judge whether a direction is useful, and predict the result.

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—