Aspire: Can Models Self-Evolve from Vague Goals?

ByteDance Seed (with SUTD, M-A-P and TokenWave.AI) reframes self-evolution evaluation: existing work hands the model human-specified tasks and metrics, reducing self-improvement to optimizing an explicit objective. Aspire gives the agent only a vague natural-language capability goal ("become a better physicist") while the downstream evaluation tasks stay hidden — the agent must operationalize the goal itself: choose data and update methods, construct its own training and validation signals, and decide when it has actually improved.

The resulting systems are judged on a hidden, expert-authored set of 520 items spanning six goals, supporting both model-weight and agent-harness evolution in one interactive environment. Experiments show vague goals redirect search effort toward goal interpretation: agents routinely complete training and harness-editing loops, but weight-level gains remain elusive. The goal-formation leg of the Seed team's Self-Developing Agents trilogy, with S³Gym (experience integration) and HarnessDev (system integration).