Thinking Machines Ships StableReinforce, a Loss That Keeps Off-Policy RL From Collapsing
LiTianleli · x · 2026-10-22
Thinking Machines' Tinker platform released StableReinforce (alpha), targeting the classic failure modes of off-policy RL: async staleness and train/sample numerical mismatch.
- A REINFORCE variant that stays stable even when the sampling policy drifts from the trained policy; internally it proved extremely stable under highly off-policy async RL and low-precision numerics.
- Mechanism: sample with recordstabilityinfo=True, then annotate training tokens with the sequenceids captured at sampling time — minimal extra information required.
- Still alpha, API may change; docs include full sampling and training code examples.
More from coding & agent
- Couple builds 2.5D pixel MMO 'Lumoria' with Claude Opus — first boss-fight gameplay revealed — AkiraVero · 2026-10-22
- Proof of Agent network upgrade: hosted mining paused 24-72h, agent ops playbook shared — adamamcbride · 2026-10-22
- Agent ops tips: retire bots with clean handoff docs and cap retries at 1-3 loops — alexcovo_eth · 2026-10-22
- Supabase ships unified .NET SDK v8: all packages in one repo, one version number — dshukertjr · 2026-10-22
- MCP Incident Room: An Interactive TypeScript Prototype for Learning MCP Design — kilmex33 · 2026-10-22
- Perplexity brings Graphify to Computer: visualize any codebase as an interactive map, 5,000 free credits per user — AravSrinivas · 2026-10-22