New LLM method localizes bugs with single token per file

Researchers have developed a new method to make bug localization in code dramatically faster and more precise. The approach, called MLC (Multi-task LLM for Bug Classification), uses a lightweight LLM with auxiliary decoding heads to identify bugs at the line level using just a single generated token per file. This is a major leap over existing agentic methods that require thousands of tokens and minutes of reasoning per file.

The method introduces three key innovations: a token alignment algorithm to handle tokenization mismatches, a multi-task training recipe optimized for multi-line predictions, and a model architecture that classifies tokens as buggy or clean directly during decoding. On Defects4J and PypiBugs benchmarks, MLC achieves accuracy comparable to heavyweight agentic systems but with orders of magnitude lower latency. The method also generalizes to out-of-domain Python datasets. The author plans to open source the code, models, and datasets upon paper acceptance.