代理系統與評測
RSIAgent Freezes Model Weights and Builds Reusable Environment Memory Through Autonomous Exploration
RSIAgent uses curriculum, actor, and verifier agents to first explore tools broadly and then deeply test failure boundaries, saving the results as action–condition–outcome memories. Its partial score on OSWorld 2.0 rose from 71.97 to 78.98, but the comparison mixes different budgets and selectively chosen retries, so it cannot be directly interpreted as surpassing GPT-6 in model capability.
Aether Labs has released RSIAgent, a multi-agent framework that accumulates experience through interaction with its environment instead of updating the underlying model weights. The system consists of three roles: curriculum, actor, and verifier. The curriculum agent generates exploration objectives, the actor operates the desktop or tools, and the verifier checks the resulting artifacts. Experiences that pass verification are then organized into “action–condition–outcome” rules, frozen, and made available for retrieval during formal tasks.
The exploration process is divided into broad and deep stages. The former tries different applications, workflows, and error modes in parallel to expand coverage; the latter repeatedly experiments with difficult cases, hidden constraints, and boundary conditions. For example, in a REAPER audio task, the agent used its exploration results to correct the resampling settings. During the first evaluation using frozen memory, the partial score increased from 68.00 to 94.17. The code includes runners for OSWorld 2.0 and Agents’ Last Exam (ALE), along with VM isolation, memory, and recovery modules, and can be reproduced in an environment with Docker, KVM, and access to model APIs.
The paper reports a score of 78.98 on 82 offline OSWorld 2.0 tasks, compared with 71.97 for the same framework without recursive self-improvement. The full-task success rate increased from 37.80% to 42.68%. The gains on ALE were smaller: the partial score rose from 83.75 to 84.82, while the full success rate increased from 49.25% to 50.75%.
These figures must be interpreted cautiously. In the OSWorld aggregate, RSI results were used for only 41 tasks, while the remaining tasks retained baseline results. ALE likewise replaced results for only 19 tasks and included selectively chosen retries, different budgets, local reevaluation, and protocol variations, rather than retesting every task at the same cost. From an engineering perspective, the issues truly worth tracking are exploration cost, verifier false judgments, and how incorrect rules can be revoked after being written into long-term memory—not merely leaderboard position.