Multi-Turn Agent RL Gains Hinge on Interaction Budgets, CASIA Study Finds
CASIA · hf · 2026-11-05
A CASIA/Tsinghua paper evaluates multi-turn agentic RL via Pass@(K,T) and finds trajectory-level gains depend heavily on interaction budgets, collapsing as budgets tighten. Agentic RL doesn't uniformly improve single-step decisions—its strength lies in error recovery and long-horizon resilience. The authors propose BudgetMix RL, which varies interaction budgets during training for robust performance across deployment settings.
More from coding & agent
- Musk shows off Grok calling a dentist to check insurance coverage — elonmusk · 2026-11-05
- Treat AI bots as employees, not assistants: an onboarding playbook — elonmusk · 2026-11-05
- Codex Voice cuts tool latency by streaming context to the harness before delegation — SIGKITTEN · 2026-11-05
- I split AI short-drama production into 5 agents—here's what broke — LateQuantity4194 · 2026-11-05
- Keynote built entirely with OpenClaw agent via voice notes on a train ride — vincent_koc · 2026-11-05
- DeepSeek's dsh-superagent: a coding agent whose rulebook is 'stop narrating your work' — shoumikgoswami · 2026-11-05