A new algorithm can identify why an AI system reaches bad outcomes, even when it doesn't know the underlying probability model.
Researchers studying Markov decision processes (MDPs) - mathematical frameworks for modeling sequential decisions under uncertainty - have long had tools to quantify the chance of an undesired outcome, but not to explain it. This paper proposes a learning approach that finds "probability-raising causes": states whose visitation makes a bad result more likely. The key innovation is a restart-based modification that sidesteps a core technical problem - existing methods require knowing the MDP's reachability probabilities in advance, which isn't possible when the transition model is unknown. The algorithm instead reduces the problem to two conditional queries and uses two-sided value iteration to progressively label states as causal, non-causal, or undecided.
That distinction matters because explainability in autonomous systems is increasingly a regulatory and safety concern, not just a research nicety. Most model-checking tools tell engineers that a failure probability is 12% - this approach would tell them which states are actually driving that number, giving them somewhere to intervene.
Experiments on two benchmarks show the method is both reliable and fast. That's a low bar to clear in a controlled setting; the harder test will be whether it scales to the large, noisy MDPs that show up in real robotics and infrastructure applications.