"Distributional Program Analysis: Treating Large Language Models as Program Samplers," by Phat T. Tran-Truong and Dr. Xuan-Bach Le, has been accepted into the NIER Track of ASE 2026 (IEEE/ACM International Conference on Automated Software Engineering), CORE Rank A*, with an acceptance rate of 23/89 (25%). The conference is held in Munich, Germany, in November 2026. A specification does not pin down a single program. Given one, a large language model induces a whole distribution of candidate programs that satisfy it in different ways. Pipelines that generate one program and analyse it throw away the most informative artifact of that distribution: the variation between samples. The authors propose distributional program analysis — analyse many LLM samples for the same task, then aggregate analyzer output as cross-sample evidence. The central statistic is an invariant's survival rate, the fraction of samples in which the analyzer infers it. Invariants with high survival mark properties that independently sampled programs agree on; fragile ones may signal decoding noise, underspecification, or blind spots in the analyzer. The paper is careful on this point: agreement is not correctness. A pilot on HumanEval+ finds that high-survival invariants match reference-solution invariants in roughly four cases out of five, and that filtering by survival also improves pass@1 at the same 20-sample budget. Paper: https://lexuanbach.github.io/publication/ASE2026-NIER.pdf Artifact: https://doi.org/10.5281/zenodo.21863484 DOI: https://doi.org/10.1145/3832783.3834555