Aspire: Can Models Self-Evolve from Vague Goals?

Imagine you're trying to help an AI learn and improve over time, but instead of giving it specific tasks, you just tell it to 'become a better researcher.' The challenge here is that without clear objectives, the AI has to figure out what that means, how to assess its own progress, and what data to use for learning. Current methods often rely on predefined tasks and metrics, which can limit the AI's ability to adapt and grow in a meaningful way. This is what's called a lack of flexibility in learning, where the AI might optimize for the wrong goals or fail to recognize its own shortcomings.

To address this, the authors propose ASPIRE, a benchmark that allows AI agents to evolve based on vague goals. Instead of being given explicit tasks, the agents receive a natural-language capability goal and must determine how to interpret that goal, select appropriate data, and decide when to evaluate their progress. This approach encourages a more exploratory learning process, where agents can adapt their strategies based on their interpretations of the goals. However, the results indicate that while agents can complete training loops, their improvements at the weight level are inconsistent and often do not translate to better performance on hidden evaluations. This means that while ASPIRE offers a new way to think about AI learning, there are still significant challenges to overcome in ensuring that agents can effectively learn and improve in real-world scenarios.

ASPIRE introduces a new approach to vague-goal-driven self-evolution in LLMs.

The evaluation is based on a hidden expert-authored set, but results show instability.