Alibaba unveils System-One RSI for agent reflexes
Alibaba Token Foundry and CAS introduced System-One RSI, a dual-loop framework that coaches fast vision-language decision models with slow reasoning agents. By compiling reasoning traces into decision trees and LoRA adapters, it builds sub-second visual reflexes for complex environments without simulators or human demonstrations.
Bridging slow deliberate reasoning and fast reactive execution is the holy grail for interactive and embodied AI. System-One RSI demonstrates that rather than forcing sluggish reasoning models into real-time loops, developers can deploy them as offline coaches that build low-latency muscle memory.
- –Decoupled dual loops: The high-frequency inner loop runs an OODA cycle on a single-pass 27B decision model for sub-200ms latency, while the outer loop uses multimodal reasoning agents to recursively refine prompts, decision trees, and LoRA weights.
- –No simulator required: Unlike typical reinforcement learning setups that require millions of frames in parallel simulated environments, System-One RSI trains directly on raw video feeds and human keyboard inputs with only a few real episodes.
- –Measurable real-time gains: In combat scenarios within Black Myth: Wukong, recursive coaching improved offensive policy win rates from 0% to 66.7% and boosted survival times by 11x against hostile attacks.
- –Practical architecture for action: This coached motor-learning pattern offers a viable template for GUI and robotics agents where autoregressive token generation cannot satisfy sub-second deadlines.
DISCOVERED
26m ago
2026-11-22
PUBLISHED
29m ago
2026-11-22
RELEVANCE
AUTHOR
Discover AI
