Aspire: Can Models Self-Evolve from Vague Goals?
2026-08-31 • Computation and Language
Computation and Language
AI summaryⓘ
The authors created ASPIRE, a system that challenges AI agents to improve themselves starting from vague goals like "get better at research" rather than clear tasks. These agents must figure out how to learn, what data to use, and how to check their progress without being told exactly what to do. The authors found that while agents can run training loops, real improvement is limited and unstable because they often pick the wrong data and rely on narrow self-checks that don't generalize well. Overall, ASPIRE tests how well AI can handle messy, unclear goals that humans often have.
Self-evolutionLarge Language ModelsVague GoalsBenchmarkTraining Data SelectionModel Weight UpdateSelf-evaluationAgent HarnessHidden EvaluationCapability Gap
Authors
Yuhao Wu, Jingyuan Zhang, Jiajun Shi, Yuxuan Zhang, Xinping Lei, Junting Zhou, Zexuan Wang, Yuchen Wu, Huan Zhou, Duo Wang, Yinzhu Piao, Yongchang Peng, Yunfeng Shi, Jin Chen, Zuo Wang, Jinkai Liu, Jiaheng Liu, Wenxuan Zhang, Shen Yan, Wenhao Huang, Ge Zhang
Abstract
Many important forms of human learning begin with a vague goal, such as "become a better physicist" or "improve at research." Learners must interpret the goal, identify capability gaps, decide how to learn, and determine whether they have actually improved. In contrast, existing work on LLM self-evolution typically begins with tasks and evaluation metrics specified by humans, reducing self-evolution to optimizing an explicit objective rather than deciding what and how to learn. We introduce ASPIRE, a benchmark for vague-goal-driven self-evolution. ASPIRE provides only a natural-language capability goal while downstream evaluation tasks remain hidden. The agent must operationalize the goal by choosing data and update methods, constructing training and validation signals, and deciding when to evaluate. ASPIRE supports both model-weight and agent-harness evolution in a unified interactive environment and evaluates the resulting systems on a hidden, expert-authored set of 520 items spanning six goals. Our experiments show that vague goals redirect search effort toward goal interpretation. Current agents routinely complete training and harness-editing loops, but weight-level gains remain sparse and unstable, and the strongest evolved harness remains below the engineered Qwen-Agent reference. Agents often train on mismatched data and trust narrow self-evaluations, so local gains fail to transfer to hidden evaluation and continued search and training can erase earlier improvements.