Daily AI Paper

Many AI systems can optimize a task once we spell out the goal clearly, but humans often start with something much vaguer, like become better at research or improve my coding. This paper asks whether models can do that too. The authors introduce ASPIRE, a benchmark where the agent only gets a natural-language goal, while the actual evaluation tasks are hidden. The system has to figure out what the goal really means, choose what data to learn from, decide how to update itself, and determine when it has improved. That makes the problem much closer to real self-directed learning than standard benchmark tuning. The key finding is that today’s agents are good at running training loops and modifying their own harnesses, but they struggle to make reliable weight-level improvements. They often learn from the wrong data and overtrust shallow self-checks, so gains do not transfer to hidden tests. This matters because it shows self-evolution is not just about more optimization, but about goal interpretation and choosing the right learning strategy.