Can AI Agents Predict the Future?
mlj.solutions
mlj.solutions Agent, který projde jen happy path, není připravený na provoz. Chybějící data, procesní výjimky i špatná oprávnění ukážou, jestli dokáže bezpečně změnit plán. 📄 FutureSim: https://arxiv.org/abs/2605.15188
View on InstagramWhat the study is about
FutureSim replays real news in the order it was published, and agents are asked to predict events beyond the limits of their knowledge. The authors tested top agents on the period from January to March 2026. The study measures how well agents can adapt to new information.
Statistics
25%
accuracy of the best agent
The best of the tested agents correctly predicted only a quarter of events.
3 months
of simulated time
Agents followed a chronological replay of real news, and questions were scored as soon as they closed.
What this means
- Static tests aren't enough, the world changes, and agents must be able to adapt to information that arrives continuously.
- There are clear differences between agents, the benchmark revealed a distinct spread of capabilities among the tested systems.
- Uncertainty is a weak spot, open-weight models had negative forecast quality, with estimates worse than making no prediction at all.
- Replaying reality is a more realistic test, real events in the order they happened measure adaptation better than isolated tasks.
- The study opens up further research, FutureSim is well suited to studying long-term adaptation, retrieval, memory and reasoning under uncertainty.
Our take
An agent shouldn't work with a frozen world. As soon as data, prices or rules change, it has to adapt.
Michal Dobrovolný
Founder & CTO
Zdroj
Goel, S., Chandak, N., Arun, A., Prabhu, A., Staab, S., Hardt, M., Andriushchenko, M., & Geiping, J. (2026). FutureSim: Replaying World Events to Evaluate Adaptive Agents. arXiv:2605.15188. https://arxiv.org/abs/2605.15188
Don't miss these posts
When More Agents Pay Off and When They Don't
More AI agents don't necessarily mean a better result. A study shows that what matters is how much information is lost when it is passed between agents.
AI Remembers a Claim Even When You Tell It It's False
When a model is fine-tuned on documents that deny a claim, it often learns the claim as true anyway. The authors named this effect Negation Neglect.
The Skill That Explains Anything to Anyone
ELI5 is an open-source skill for Claude Code that adapts explanations to the audience. With the skill it passed 10 of 12 criteria, without it only 5.
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
