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

2026-06-17 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, long

Summary

This research introduces a novel approach for modeling concurrent Go programs by treating scheduler nondeterminism as a training signal. Instead of predicting a single next event, a 7B Qwen2.5-Coder model is fine-tuned using a KL objective to match an empirical distribution of next events observed across repeated program runs. This method achieved 36.2% accuracy on 798 held-out predictions from real production Go bugs (CockroachDB, Kubernetes, gRPC, etcd), outperforming Gemini 3.5 Flash zero-shot (34.8%) and the same model without fine-tuning (28.6%). While accuracy was comparable to cross-entropy training (35.8% vs. 36.2%), the KL objective reduced Expected Calibration Error from 0.205 to 0.169. The study also formally derived a goroutine-leak signature for select-blocked goroutines where P("GoUnblock")=0 due to scheduler semantics. The dataset, trained adapters, and tooling are publicly released.

Key takeaway

For AI Scientists and Machine Learning Engineers working on concurrent program analysis, recognize that traditional deterministic execution models fall short. You should consider using distribution-aware training with objectives like KL divergence to capture nondeterministic behavior, improving model calibration and accuracy on real-world concurrency bugs. Be aware that current single-step training limits multi-step prediction coherence. Explore enriching trace representations with synchronization state for future work.

Key insights

Nondeterminism in concurrent execution can be a training signal for next-event distribution prediction, improving model calibration.

Principles

• Concurrent execution is inherently nondeterministic.
• Nondeterminism provides a principled distributional training target.
• Single-step training limits multi-step prediction coherence.

Method

Fine-tune a 7B Qwen2.5-Coder model with a KL objective to match empirical next-event distributions from repeated concurrent Go program runs, using 4-bit QLoRA.

In practice

• Use empirical distributions for concurrent program modeling.
• Focus on single-step prediction for current models.
• Consider encoding channel/mutex state for better reasoning.

Topics

• Concurrent Go Programs
• Execution Modeling
• Nondeterminism
• Distribution Estimation
• Model Calibration
• Concurrency Bugs

Code references

• kaviru2/weave

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs — LLMs struggle significantly with generating samples calibrated to true underlying distributions, hindering their use in simulations.
• Instance-Aware Knowledge Distillation for Semi-Supervised Learning of an On-Board Multi-Task Dense Prediction Model for Collision Avoidance System — Instance-aware knowledge distillation improves lightweight models for edge-deployed multi-task dense prediction by mitigating teacher bias.
• Distribution Corrected Offline Data Distillation for Large Language Models — Offline distillation can be improved by correcting teacher-student distribution drift without costly online sampling.
• A Quantitative Experimental Repeated Measures Study of Training Dynamics in a Small Llama Style Language Model Under a Compute-Aware Token Budget — Compute-constrained LLM training trajectories reveal non-monotonic degradation and instability, challenging endpoint-only evaluation.
• Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning — Larger LLMs demonstrate superior efficiency and performance in RL post-training for mathematical reasoning.
• Knowledge-Based Zero-Replay Debugging of Multi-Agent LLM Traces — Predicting high-effect events in multi-agent LLM traces without costly counterfactual replay significantly improves debugging efficiency and localization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.