PLLM+: hybrid replay-and-repair pipeline for Python dependency resolution (arXiv:2609.26952v1)
This paper presents PLLM+, a hybrid dependency-repair pipeline evaluated on the HG2.9K benchmark of 2,891 dependency-failing Python snippets. PLLM+ applies inexpensive deterministic steps first (AST-based interpreter inference, replaying historically successful dependency configurations from a solutions database, and live PyPI validation) and falls back to a structured LLM-based repair loop (typed error classification with Proposer/Critic agents) only when needed; on HG2.9K it solves 1,500/2,891 snippets (vs. 1,169 for the PLLM baseline), reduces average runtime from 368.7s to 71.8s per snippet, and 1,495 of the 1,500 successful fixes come from replayed solutions while the LLM fallback adds five fixes.