Aspire: Can Models Self-Evolve from Vague Goals?

Aspire. Can Models Self-Evolve from Vague Goals?. By Yuhao Wu and twenty others, from ByteDance Seed, the Singapore University of Technology and Design, and two other institutions. Published on the thirty-first of August. No code repository is linked to it.

A great deal of human learning starts from a vague goal. Become a better physicist. Improve at research. The learner has to interpret what that means, find where their capabilities fall short, decide how to learn, and judge whether they have actually improved. Work on self-evolution in language models usually skips all of that. It begins with tasks and evaluation metrics that humans have already specified, which reduces self-evolution to optimizing an explicit objective. The harder part, deciding what to learn and how, is left out.

ASPIRE is a benchmark for self-evolution driven by vague goals. The agent is given only a capability goal, written in natural language, and the tasks it will eventually be evaluated on stay hidden. What is measured is whether the agent can operationalize that goal on its own and come out genuinely more capable.

To operationalize a goal, the agent has to choose data and update methods, build its own training and validation signals, and decide when to evaluate. Two kinds of evolution are supported within one interactive environment. The agent can change the model's weights, or it can change the agent harness around the model. The evolved systems are then scored on a hidden set of five hundred and twenty items, written by experts, spanning six goals.

A vague goal redirects the agent's search effort toward working out what the goal means. Current agents routinely complete the loops of training and of editing the harness. Gains at the level of weights, however, remain sparse and unstable. And the strongest harness any agent evolved still falls below the engineered Qwen-Agent reference.

The failure modes are specific. Agents often train on data that does not match the goal, and they trust self-evaluations that are too narrow. Gains that look real locally then fail to transfer to the hidden evaluation. Worse, continued search and training can erase improvements that had been made earlier.

Being able to run a training loop is not the same as knowing what to train for. The agents here manage the mechanics, yet stumble on the interpretation: picking the right data, and checking themselves against something broad enough to mean anything. ASPIRE puts that gap between doing the work and choosing the work at the centre of the measurement.