Researchers have created a benchmark called FM-Bench to evaluate the performance of language model agents in long-term decision-making scenarios. The benchmark simulates running a football club for 20 years and measures six behavioral capabilities behind the final score. The results show that higher-scoring models exhibit more effective managerial behavior, such as reducing slow-payoff investments and keeping cash invested rather than idle. However, no model learns market prices from rejected bids, and self-managed memory fails in two opposite modes. The benchmark is available on GitHub. --- Why it matters: This benchmark matters to AI researchers because it provides a standardized evaluation framework for long-term decision-making capabilities of language models. This can help identify strengths and weaknesses of different models and inform the development of more effective AI systems. Source: https://arxiv.org/abs/2608.18423