Muyu He@HeMuyu0327
We've had a lot of fun building this benchmark (asking LLMS to run a startup), which gives the clearest signal on LLMs' "long-term coherence" ability. We observe that frontier models have significant variance on this benchmark, showing that long-term execution is still under-optimized.
The benchmark is easily runnable on HF and OpenReward, which we give links below. The evals will give these very interesting leaderboards for all models (p1) and open source models (p2).
Major takeaways from analyzing their performances:
- Most LLMs have long-term commitment issues. To run a company, it is very beneficial to maintain a good relationship with target clients, since that means more rewards and less work. Most models never follow suit. Only very few of them dedicate to 1-2 clients and yield huge returns. This is alarming because committing to specific clients is kind of a "free ride", yet most models never think of it.
- Most LLMs also do not check on their failure modes well enough. Some clients are designed to be bad, giving models extra work at no benefit. Models need to spot and blacklist them, and they have perfect access to this information after a few task failures (or even at their first interaction with the client). Again, only very few models correctly notice subtle abnormality and act preemptively.
In the near future, we want autonomous agents to handle intensive long-term management work, acting like product manager, tech leads, and even founders. Our benchmark shows the concrete axis of optimization we need to make to get there. Evaluate on your model today!
