JAY ZENITH

[github] [huggingface] [x]

I build post-training systems for tool-using LLM agents: verifier-based environments, RL training infrastructure, and auxiliary learning objectives.

  • predict

    Training an agent to predict the verified outcome of its own actions before executing them. PREDICT extends GRPO with auxiliary outcome supervision. A coding agent proposes a patch, predicts what will happen if it executes that patch, and commits to KEEP or REVISE before seeing the result. After execution, the environment provides the verified outcome: GRPO trains the agent's actions, auxiliary cross-entropy trains the earlier prediction against that ground truth. Unlike observation-prediction methods such as ECHO, PREDICT's target does not need to appear in the rollout itself. On Qwen3-4B, RL training turned a checkpoint that always predicted PASS into a runtime-error detector reaching 62.5% and 64.1% precision across two seeds; the broader agent-performance advantage is not yet established. [read PREDICT] [code]