iCoRe: An Iterative Correlation-Aware Retriever for Bug Reproduction Test Generation

2026-03-24 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

iCoRe, an iterative correlation-aware context retrieval approach, significantly enhances the automatic generation of bug reproduction tests (BRT) from issue descriptions. It addresses key limitations in existing Large Language Model (LLM)-based retrieval methods by employing a differentiated strategy for source code and test cases, incorporating function call relationships beyond semantic similarity, and utilizing a feedback loop from the generation phase for iterative refinement. Evaluated with an LLM-based BRT generator on SWT-bench Lite and TDD-bench Verified benchmarks, iCoRe achieved Fail-to-Pass rates of 42.0% and 52.8% respectively. This represents a 19.7%–31.7% relative improvement over prior retrieval methods, outperforming state-of-the-art e-Otter++ with GPT-4o (40.2% and 51.4%) while being 6.4 times more cost-effective. It also provides more concise context, averaging approximately 2,500 tokens.

Key takeaway

For Machine Learning Engineers developing automated bug reproduction systems, you should consider adopting iCoRe's correlation-aware retrieval. Its differentiated strategy for code and tests, combined with iterative refinement and function call analysis, significantly improves BRT generation accuracy and cost-effectiveness. Integrating this generator-agnostic approach can lead to higher Fail-to-Pass rates and reduce expensive LLM calls, making your automated testing pipelines more robust and efficient.

Key insights

iCoRe enhances bug reproduction test generation via correlation-aware, iterative context retrieval, differentiating code/tests and integrating function call analysis.

Principles

• Production code and test case retrieval demand distinct strategies.
• Function call relationships reveal behavioral relevance beyond text.
• Iterative feedback from generation refines context retrieval.

Method

iCoRe employs a two-stage process: heuristic-based production code retrieval and iterative test code retrieval. This involves sketch BRT generation, textual/function-call similarity calculation, and LLM-based reranking.

In practice

• Differentiate retrieval for production code and test cases.
• Use function call analysis for behavioral code relevance.
• Implement generation-to-retrieval feedback loops.

Topics

• Bug Reproduction
• LLM-based Code Retrieval
• Software Testing Automation
• Function Call Graphs
• Iterative Context Refinement
• Test Generation

Code references

• ZJU-CTAG/iCoRe

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• Automated Root-Cause Subclassification and No-Code Fix Generation for Invalid Bug Reports — Automating invalid bug report subclassification and no-code fix generation significantly reduces software maintenance overhead.
• When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in Scientific Multi-hop Question Answering — Iterative RAG's staged retrieval and reasoning process outperforms static ideal evidence in complex multi-hop QA.
• Meta Reports 4x Higher Bug Detection with Just-in-Time Testing — JiT testing dynamically generates targeted tests during code review, enhancing bug detection in AI-driven development.
• Revisiting Vul-RAG: Reproducibility and Replicability of RAG-based Vulnerability Detection with Open-Weight Models — RAG-based vulnerability detection with open-weight LLMs faces a performance plateau around 0.30 pairwise accuracy.
• Towards Functional Correctness of Large Code Models with Selective Generation — Code hallucination in LLMs can be controlled by a selective generator using fuzzing-generated unit tests and "α-code entailment".
• Livecoding: ICE and the Factored Cognition Primer by Ought — Iterated decomposition and execution tracing significantly enhance language model reliability and debugging.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.