When the Next Step Is Not One Step: Distribution-Aware Execution Modeling for Concurrent Go Programs

A research team has developed a method that treats the inherent nondeterminism of concurrent Go programs as a training signal, improving a language model’s ability to predict program behavior. The approach, detailed in a paper submitted to arXiv, fine-tunes a 7-billion-parameter model to match empirical event distributions rather than single outcomes [1]. The technique addresses a core challenge in modeling concurrent systems: two executions of the same program from an identical trace prefix can yield different, yet valid, next events due to scheduler nondeterminism. A model trained on a single label is effectively learning to guess one outcome of a random process [2]. The researchers, led by Kaviru Hapuarachchi, instead run each program many times and aggregate the observed next events into an empirical distribution. A 7B model is then fine-tuned to match this distribution using a Kullback-Leibler divergence objective [1][2]. The model was evaluated on 798 held-out predictions drawn from real production Go bugs found in systems including CockroachDB, Kubernetes, gRPC, and etcd [1][2]. CockroachDB is a distributed SQL database designed to survive infrastructure failures, while etcd serves as a distributed key-value store often used for cluster coordination [8]. After fine-tuning on fewer than a thousand traces, the model reached 36.2% accuracy on these predictions. This outperformed Google’s Gemini 3.5 Flash used in a zero-shot setting, which achieved 34.8% accuracy, and the same base model without fine-tuning, which scored 28.6% [1][2]. Distribution training matched the accuracy of a cross-entropy baseline (35.8% versus 36.2%) while reducing the Expected Calibration Error from 0.205 to 0.169, indicating more reliable confidence estimates [1][2]. The underlying model architecture belongs to the transformer family, which relies on multi-head attention mechanisms to contextualize tokens and has become foundational for large language models since its introduction in 2017 [3]. Beyond the empirical results, the paper derives a formal goroutine-leak signature for a class of select-blocked goroutines. In these cases, the probability of a goroutine unblocking is zero by scheduler semantics, not by learned behavior [1][2]. The researchers have released the dataset, trained adapters, and associated tooling alongside the paper [1].