Don't Complete It! Preventing Unhelpful Code Completion for Productive and Sustainable Neural Code Completion Systems

TL;DR

This work addresses the productivity and energy cost waste caused by unhelpful code completions in neural code completion systems. It introduces FrugalCoder, an early-rejection mechanism that uses lightweight, learning-based estimators to predict completion quality from a code prompt and blocks prompts when the estimated quality is below a threshold $t$, thereby avoiding expensive LCM inferences. The authors identify four patterns of low-return prompts and validate FrugalCoder with five estimators across two LCMs (StarCoder and CodeGen2) and two languages (Java and Python), using BLEU-based scores and a human-annotated benchmark (HumanAccept). Key findings show that high-precision estimators (notably an encoder-only Transformer) can reject a meaningful portion of prompts with minimal overhead, improving acceptance rates and BLEU metrics for retained completions while significantly reducing computation, demonstrating practical potential for sustainable AI-assisted coding. The work provides publicly released data and artifacts and outlines future directions for more accurate estimators and broader industrial validation to maximize impact.

Abstract

Currently, large pre-trained language models are widely applied in neural code completion systems. Though large code models significantly outperform their smaller counterparts, around 70\% of displayed code completions from Github Copilot are not accepted by developers. Being reviewed but not accepted, their help to developer productivity is considerably limited and may conversely aggravate the workload of developers, as the code completions are automatically and actively generated in state-of-the-art code completion systems as developers type out once the service is enabled. Even worse, considering the high cost of the large code models, it is a huge waste of computing resources and energy, which severely goes against the sustainable development principle of AI technologies. However, such waste has never been realized, not to mention effectively addressed, in the research community for neural code completion. Hence, preventing such unhelpful code completions from happening in a cost-friendly way is of urgent need. To fill this significant gap, we first investigate the prompts of unhelpful code completions, called "low-return prompts". We empirically identify four observable patterns in low-return prompts, each lacking necessary information, making it difficult to address through enhancements to the model's accuracy alone. This demonstrates the feasibility of identifying such low-return prompts based on the prompts themselves. Motivated by this finding, we propose an early-rejection mechanism to turn down low-return prompts by foretelling the code completion qualities. The prompts that are estimated to receive unhelpful code completions will not be sent to the model. Furthermore, we investigated five types of estimators to demonstrate the feasibility of the mechanism. The experimental results show that the estimator can reject 20% of code completion requests with a 97.4% Precision.

Figures (3)

• Figure 1: The workflow of an LCM-based code completion system with FrugalCoder: Given a code prompt, the estimator first produces an estimation score $s$ that reflects the anticipated quality of completion. If $s$ exceeds a predefined threshold $t$, the LCM is activated to fulfill the completion request. Conversely, if $s$ falls below $t$, the request for completion is denied.
• Figure 2: The simple examples of the prompts following each summarized pattern. The example following the pattern is in red, and the slightly shaped variant is in green. The completions are marked with the gray background.
• Figure 3: The training process of a learning-based estimator.