RSI Agent
A recursive self-improvement agent system released by Acer AI, built on a multi-agent collaborative architecture. The system improves task completion through experience accumulation and memory freezing mechanisms while keeping the underlying model parameters fixed.
Timeline (last 90 days)
Acer AI recently released an RSI Agent whose core design keeps the underlying model parameters fixed, locks in well-saved memories before formal testing, and reuses those memories for subsequent tasks.
Acer AI's RSI Agent workflow involves multiple agents with divided responsibilities, including an agent for scheduling practice, an execution agent, and a verification agent.
In the Shotcut watermark-removal task, a run without exploration experience scored only 0.40, while a run carrying frozen post-exploration experience scored 1.00.
In the overall results of the 82 offline tasks in OS World 20, the average completion rate rose from approximately 72% to approximately 79%.
The 82 tasks in OS World were not run under identical conditions—half of the tasks retained the original baseline scores, and the invested resources and number of attempts also differed—so the 7-percentage-point improvement cannot be entirely attributed to memory.
The primary risk of the RSI Agent is that memory may amplify errors: if the verification agent makes a misjudgment, it will write the incorrect approach into long-term memory as a successful experience, which will then be reused by subsequent tasks.
In Acer AI's RSI Agent, the improvement regime at the upper layer is still designed by humans, and high-level goals and deployment decisions are also controlled by humans.
All Facts (7)
In the overall results of the 82 offline tasks in OS World 20, the average completion rate rose from approximately 72% to approximately 79%.
50%UnverifiedAcer AI's RSI Agent workflow involves multiple agents with divided responsibilities, including an agent for scheduling practice, an execution agent, and a verification agent.
50%UnverifiedIn the Shotcut watermark-removal task, a run without exploration experience scored only 0.40, while a run carrying frozen post-exploration experience scored 1.00.
50%UnverifiedIn Acer AI's RSI Agent, the improvement regime at the upper layer is still designed by humans, and high-level goals and deployment decisions are also controlled by humans.
50%UnverifiedThe 82 tasks in OS World were not run under identical conditions—half of the tasks retained the original baseline scores, and the invested resources and number of attempts also differed—so the 7-percentage-point improvement cannot be entirely attributed to memory.
50%UnverifiedThe primary risk of the RSI Agent is that memory may amplify errors: if the verification agent makes a misjudgment, it will write the incorrect approach into long-term memory as a successful experience, which will then be reused by subsequent tasks.
50%UnverifiedAcer AI recently released an RSI Agent whose core design keeps the underlying model parameters fixed, locks in well-saved memories before formal testing, and reuses those memories for subsequent tasks.
50%