Researcher: purely on-policy RL is too sample-inefficient for long-horizon agents
JoshPurtell · x · 2026-10-29
A researcher running scaled RL experiments on Thinking Machines' Tinker and Prime Intellect argues that purely on-policy RL is extremely sample-inefficient and expensive, making it unlikely to work for long-horizon agentic scenarios. He calls for rethinking how agents learn from experience beyond scaling on-policy rollouts, and teases a promising direction to be shared soon.
More from coding & agent
- 10-person Spawn spends 30x payroll on AI tokens, ships 77,000 changes a month — mattshumer_ · 2026-10-29
- Grok Bot locked out by Amazon simply bought the same item elsewhere, cheaper — jamesdouma · 2026-10-29
- Dev goes vibe coding C++: JUCE Acid-303 plugin parameter code as proof — kylegawley · 2026-10-29
- Open-source Jev undercuts Clay: $0.038 per verified lead vs $167/mo plans — HeyAmit_ · 2026-10-29
- 218 agents, 425 PRs: one developer's extreme parallel Cursor workflow — mattyp · 2026-10-29
- agent-bridge-mcp: open-source MCP lets Claude Code and Codex sessions chat on one machine — blaizedsouza · 2026-10-29