LLM-based code completion systems typically offer only one suggestion, often failing to cover developers’ intent, especially with ambiguous context. Current multiple-hypotheses code completion, i.e., generating multiple full-code candidates to increase the possibility to hit the user intent, can mitigate this but with a price of high latency and developers’ cognitive burden. To fill this gap, we propose FastMCC, which presents multiple hypotheses as concise natural language phrases instead of full code. It uses a two-stage inference: first generating a initial code suggestion followed by alternative intent phrases; then generating intent-aligned code if a phrase is selected. We design a lightweight fine-tuning framework to equip base code LLMs with this capability without compromising their original code completion performance. Experiments on three code LLMs show FastMCC maintains the pass rate of initial suggestions, while saving 76.6% generated tokens for multiple-hypotheses options.

Currently, I’m a PhD student at Singapore Management University, supervised by Prof. David Lo.

My research interest is SE4AI and AI4SE.

The goal of my research is to build effective, efficient, and trustworthy AI tools for human coders and software tools for AI coders.