Distribution-Aware AI Model Predicts Concurrent Go Program Behavior.

Predicting the next step in a concurrent program is inherently difficult due to the nondeterministic nature of schedulers, meaning the same program prefix can lead to multiple valid subsequent events. Training a model with a single label for such a process often results in the model merely guessing one of many possible outcomes. Researchers have turned this challenge into a training signal by running concurrent Go programs multiple times and aggregating the observed next events into an empirical distribution. A 7B language model is then fine-tuned using a Kullback-Leibler (KL) objective to match this distribution, effectively learning the probabilities of different valid next steps. Evaluated on 798 held-out predictions from real-world Go bugs (e.g., CockroachDB, Kubernetes), this distribution-aware fine-tuning achieved 36.2% accuracy, outperforming Gemini 3.5 Flash zero-shot (34.8%) and the unfine-tuned model (28.6%). While accuracy was similar to cross-entropy training, this method significantly reduced Expected Calibration Error from 0.205 to 0.169. The work also formally identifies a goroutine-leak signature. All datasets, trained adapters, and tooling are open-sourced.