Sabitlenmiş Tweet
It’s a rare joy to work on something with real-world impact. Meet Chronos LTV, a system we built at Uber to estimate the causal, long-term impact of short-term delays. Links below.
This problem is notoriously difficult because of the cascading effects inherent in long-horizon learning: a single event does not just change a user’s current experience; it can also update hidden state (e.g., beliefs, preferences, habits) in ways that in turn would have lasting effects on future behaviors.
Chronos LTV combines ideas from Markov decision processes, off-policy evaluation in reinforcement learning, and causal inference. It then enables accurate long-horizon evaluation using observational data, without relying on expensive, long-running randomized controlled trials.
Beyond its immediate application here, I am optimistic that Chronos could also be useful in other long-horizon evaluation tasks in agentic AI and world models.
English
