System-One RSI: Building Muscle Memory for Interactive Agents through Guided Practice
Organizations: Alibaba Token Foundry - Infra & Model RSI - EconML · Institute of Automation, Chinese Academy of Sciences
Abstract
Reasoning models provide strong multimodal task understanding but struggle with sub-second control because deliberation increases decision latency. Fast decision models can meet the response window, yet prompting and few-shot examples do not readily provide the fine-grained experience required by a new task. Inspired by coached motor learning, this paper presents System-One Recursive Self-Improvement (System-One RSI), which couples two loops. The high-frequency online inner loop is an Observe, Orient, Decide, Act (OODA) cycle: a decision model maps each visual observation and task question to structured judgments, and a decision tree converts them into timed actions without invoking a reasoning model. The offline outer loop is a recursive improvement cycle carried out by a team of reasoning agents. The team revises the decision-model prompt and decision tree, captures interaction records, derives supervision from those records, and trains or deploys updated model parameters. The improved system then returns to the environment and generates the experience for the next round, consolidating learned distinctions as external muscle memory. The framework operates with a small number of real episodes, without expert demonstrations, environment modification, sandboxing, or parallel simulation. Experiments in Snake, Zenless Zone Zero, and Black Myth: Wukong use a 27B decision model. Mean Snake score increases from 1.30 to 7.35. In an evaluated scenario of Zenless Zone Zero, median completion time decreases by 8.4%. In an evaluated combat scenario of Black Myth: Wukong, the evasive policy achieves 11 times the mean survival time of a stationary reference, while the offensive policy win rate rises from 0% to 66.7% across the evaluated configurations.
Figures & tables
| Approach | Advantages | Limitations |
|---|---|---|
| Online RL [ 10 , 11 ] | Learns policies from interaction feedback | Environment replication, reset, and parallel rollouts may be obstructed by external dependencies. |
| Offline RL [ 12 ] | Learns from existing trajectories | Data collection can be costly; coverage limits learning under distribution shift. |
| Reasoning-based computer use [ 1 , 5 ] | Applies task knowledge through deliberation | Repeated reasoning adds latency, and visual grounding can remain unreliable. |
| Reasoning-generated CV/control [ 7 , 8 ] | Converts task knowledge into executable rules | Perceptual conditions require specification and testing across visual edge cases. |
| Unadapted JEV [ 13 ] | Returns structured judgments without a reasoning trace | Pretrained capabilities may not cover the distinctions required by a new task. |
| JEV prompt optimization [ 13 , 9 ] | Revises instructions without parameter updates | Fine distinctions can be difficult to express; longer inputs increase processing cost. |
| Model | Top-1 (%) | Balanced accuracy (%) | NLL |
|---|---|---|---|
| Always none | 86.15 | 33.33 | N/A |
| zzz-rsi-base | 63.85 | 44.52 | 0.6799 |
| zzz-rsi-v1 | 79.23 | 74.87 | 0.4565 |
| zzz-rsi-v2 | 81.15 | 77.98 | 0.4966 |
| Model | Wins | Median completion (s) | Median (s) | p50/p95 (ms) |
|---|---|---|---|---|
| zzz-rsi-base | 2/5 | 261.2 | 167.0 | 231.4/244.2 |
| zzz-rsi-v1 | 5/5 | 236.2 | 26.0 | 229.1/249.4 |
| zzz-rsi-v2 | 5/5 | 239.3 | 29.7 | 230.2/255.0 |
| Opponent | zzz-rsi-v2 mean (s) | zzz-rsi-v3 mean (s) |
|---|---|---|
| YeShiyuan | 242.7 | 159.3 |
| Amalgam | 207.2 | 190.8 |
| Butcher | 214.0 | 163.0 |
| Pompey | 245.0 | 194.0 |
| Priest | 309.0 | 293.0 |
| Sweeper | 278.0 | 242.0 |
| Model | Episodes | Mean (s) | Median (s) |
|---|---|---|---|
| Stationary reference | 6 | 23.6 | N/A |
| wk-rsi-v1 | 10 | 108.1 | 114.9 |
| wk-rsi-v2 | 10 | 160.0 | 170.1 |
| wk-rsi-v3 | 5 | 160.9 | 166.7 |
| wk-rsi-v4 | 5 | 260.5 | 212.2 |
| Model | Episodes | Win rate (%) | Mean duration (s) | Winning mean (s) |
|---|---|---|---|---|
| wk-rsi-base | 10 | 0.0 | 21.9 | N/A |
| wk-rsi-v1 | 10 | 20.0 | 74.1 | 139.7 |
| wk-rsi-v2 | 10 | 30.0 | 89.5 | 122.4 |
| wk-rsi-v3 | 15 | 60.0 | 139.7 | 154 |
| wk-rsi-v4 | 15 | 66.7 | 125.3 | 131 |
Self-Improving Agents
Self-improving agents use experience and feedback to update their models, memories, skills, or workflows for better future task performance.
See all 203 papers