AWS multi-turn RL guide: $339 lifts gpt-oss-20b Wordle solve rate 61%→84%
AWS ML Blog · rss · 2026-10-24
AWS ML Blog publishes a hands-on guide to Multi-Turn Reinforcement Learning (MTRL) on SageMaker AI, targeting agentic tasks where single-turn SFT/RL can't credit early decisions — you only score the outcome, not the trajectory. Architecture: training job + policy model in SageMaker, agent/environment/reward as one Python file on Bedrock AgentCore, GRPO updating a LoRA adapter. Demo (Wordle): gpt-oss-20b solve rate 61%→84% at default reasoning effort, 15%→69% at low effort. The 100-step run used 1.29B tokens (train dimension dominates) across 105,486 rollouts in 12h40m, costing $339 in training tokens plus $79 AgentCore runtime. Includes reward design (solve 1.05–1.55, loss -0.5, invalid -1.5), pitfalls (trust principal must be job.sagemaker.amazonaws.com), and supported models (Nova 2 Lite, GPT-OSS-20B, Gemma 4 31B, Qwen 3.6 27B).
More from coding & agent
- Google's PLE benchmark: best AI agent scores only 50.6% on professional engineering deliverables — dair_ai · 2026-10-24
- LangChain fixed its design agent's AI slop with skills, a design wiki, and guardrails — LangChain · 2026-10-24
- Chronos: a Postgres-compatible database where agent changes land only if they pass your policies — Antique-Willow-5841 · 2026-10-24
- Agent sandbox boat.dev hits 1,026 paying subscribers with $20/mo per-second-billed VMs — ycombinator · 2026-10-24
- Claude Code Team OS: put your whole product team's context in one shared repo — aakashgupta · 2026-10-24
- How AI-native product teams share one repo: thin CLAUDE.md, MCP connections, 16+ skills — aakashgupta · 2026-10-24