LLMs tackle null-pointer bugs at scale with NullRepair system

Integrating static analysis tools that catch null-pointer exceptions into mature Java codebases is notoriously difficult. Even after automated annotation inference removes many errors, a tough residue of issues—some genuine bugs, others false alarms—requires manual code fixes that are slow and mistake-prone.

A new tool called NullRepair brings large language models into this workflow, automating repairs while preserving code correctness. Rather than simply asking an LLM to fix errors, the system uses a decision framework grounded in analysis of 200 real-world cases to guide the model's output.

How NullRepair Works#

The system combines LLM reasoning with static analysis to understand which code regions are safe or unsafe. It feeds the model examples of error-free usage patterns to provide concrete context, then iteratively refines patches using broader project knowledge and decision logic. This structured approach avoids the pitfalls of naive prompting, which often produces semantically broken or contextually misaligned edits.

Real-World Results#

When tested on 12 active Java projects containing 1,119 residual nullability errors—the type that slip through standard annotation inference—NullRepair successfully resolved 63% of them. Crucially, the fixes maintain program integrity: all unit tests passed across 10 of the 12 projects after applying every suggested edit, and 98%+ tests passed in the remaining two. Competing baselines like single-shot prompting and a mini-SWE-agent variant both performed worse and left more tests broken.

Why It Matters#

Nullability checkers are becoming standard in modern Java development, but adoption in large legacy codebases has been bottlenecked by manual remediation. An automated, reliable repair mechanism could lower that adoption barrier significantly, letting teams catch whole classes of runtime errors earlier in development.

The research is available on arXiv.