代理訓練
Echoverse Trains Computer Agents with Evolvable Applications, Raising the 9B Model’s Cross-Test Success Rate to 67.1%
A Microsoft Research team jointly updates synthetic applications, tasks, and graders based on agent failure trajectories instead of merely adding more static webpages. The 9B model’s average success rate across 14 evaluations rose from 36.5% to 67.1%, although only four environments have been publicly released as benchmarks.
Computer agents need applications that can be repeatedly operated, broken, and reset for training, but many real-world services are constrained by logins, private data, and state, prompting researchers to generate large numbers of substitute websites. Researchers from Microsoft Research and other institutions introduced Echoverse, arguing that the bottleneck is no longer the number of synthetic environments, but whether those environments provide sufficient behavioral depth, target interactions where agents actually fail, and can continue evolving alongside model capabilities.
Echoverse compiles specifications into applications with backend state and evaluates task outcomes directly against application databases, rather than relying solely on vision or language models to judge the screen. Each rollout produces two forms of feedback: it is used both to repair the application, task, and verifier, and as a new training signal for the agent. This approach can address common problems such as an agent completing a task correctly but being misjudged by the grader, or an environment being overly simplified. It also allows the training distribution to shift with the latest failure modes.
The team trained a 9B model across 12 environments, raising its average success rate on 14 evaluation sets from 36.5% to 67.1%. Ablation experiments showed that shallow environments actually reduced accuracy on real websites from 80% to 75%. After deeper state and interactions were added, performance on two test sets increased from 80% to 85% and from 48% to 65%, respectively. Repairing just one environment also improved the model trained in that environment from 16.2% to 38.5%. After switching to a reinforcement learning reward that combined a database verifier with a step-by-step judge, the held-out test score rose from 58.8% to 68.0%.
These results suggest that generating large numbers of visually distinct but logically shallow websites may cause negative transfer. From an engineering perspective, state models, verifiable backends, and failure-driven curriculum updates may be more valuable investments. At present, only four environments—along with seed data and a grounded grader—have been made publicly available. The complete set of 12 training environments, the 9B checkpoint, and the cost of frontier teacher models have not been fully disclosed. The key next questions are whether third parties can reproduce cross-website transfer and whether co-evolution could cause agents to overfit to grader vulnerabilities.