Reward-Rate Policy Gradient: RL that optimizes MLE agents' reward per unit time
sherryyangML · x · 2026-10-23
Researcher sherryyangML shares Reward-Rate Policy Gradient (RPG) plus companion work MLE-World, both aimed at making better use of a limited research budget:
- RPG: for auto-research and ML-engineering agents, actions can take seconds or hours — a fixed-step MDP misses this. RPG optimizes agents' reward per unit time in a semi-MDP, folding execution time into the RL objective.
- MLE-World: even failed experiments leave useful experience. It turns execution traces (code + task context → errors, training logs, scores) into a predictive model, letting agents revise ideas offline and run real experiments only when predictions are uncertain.
The two are complementary: RPG maximizes reward rate, MLE-World makes experiment feedback cheaper by reusing experience.
More from coding & agent
- Developers Don't Code Anymore — Now Their Job Is Rotating Secrets Daily — Dan_Jeffries1 · 2026-10-23
- New paper: coding agents in a shared sandbox self-create rules, reciprocity debts, and coalitions — nandofioretto · 2026-10-23
- Developer runs AI project managers that manage their own AI agents to build his game — draginol · 2026-10-23
- RentAKaren: a real service that rents humans to handle complaints your AI agent can't win — mathemagic1an · 2026-10-23
- Clerk's Auth for MCP growing 13% week-over-week, fastest since 2023 — stuffyokodraws · 2026-10-23
- Shopify builds remote mobile dev environments for AI agents — rseroter · 2026-10-23