JoMoHa jointly RL-trains model and harness, hits 97.0% on ALFWorld
omarsar0 · x · 2026-11-13
Salesforce AI Research and HKU introduce JoMoHa, the first framework to jointly RL-train an executable harness-generation policy and a task-execution policy, enabling recursive self-improvement. Naively alternating updates causes model–harness incompatibility, with performance rising then falling; keeping separate histories and value estimates per policy fixes this. JoMoHa beats weight-only and harness-only baselines, averaging 70.3% across ALFWorld, WebShop, and DBBench, and reaching 97.0% on ALFWorld vs 90.5% for a frozen engineer.
More from coding & agent
- Rolldown ships agent-facing report mode that trimmed Framer's bundle by 1.2MB — cnakazawa · 2026-11-13
- Quinn Slack explains how Amp runs an AI-native company with 21 employees — TheTuringPost · 2026-11-13
- Pipecat 1.16 adds PerceptionReader to read prosody, tone and sarcasm in voice agents — m4rkmc · 2026-11-13
- Amp kills its Dial modes for a model selector that shows exactly who gets billed — bytebot · 2026-11-13
- Garry Tan amplifies: self-built AI code-review agent cuts costs 80%+ with better quality — garrytan · 2026-11-13
- Gemini now runs 5-hour uninterrupted coding sessions, user reports spec-compliant work — MickeySteamboat · 2026-11-13