HASTE: Hybrid AST-guided Selection with Token-bounded Extraction

ICLR 2026 Conference Submission18930 Authors

Keywords: Code Comprehension, Large Language Models, Context Retrieval, Abstract Syntax Tree, Automated Program Repair, Context Compression

Abstract:

The rise of Large Language Models (LLMs) has opened new frontiers in software engineering, promising to automate complex tasks from bug fixing to large-scale refactoring. However, this promise is critically hampered by a fundamental constraint: the limited context window of these models. This limitation forces a difficult trade-off in context retrieval. On one hand, structure-aware approaches preserve syntactic integrity but often fail to pinpoint the most semantically relevant code for a given task. On the other, relevance-focused techniques excel at finding pertinent snippets but risk severing critical structural dependencies, leading to incoherent code that causes LLMs to hallucinate.

To resolve this trade-off, we introduce HASTE (Hybrid AST-guided Selection with Token-bounded Extraction), a novel framework that synergistically integrates robust information retrieval with deep structural analysis. HASTE leverages the Abstract Syntax Tree (AST), combining lexical and semantic search to ensure the extracted code is not only topically relevant but also structurally coherent and executable. Our evaluation, conducted using a robust LLM-as-a-judge framework, demonstrates that HASTE achieves upto 85% code compression while significantly improving the success rate of automated code edits and maintaining high structural fidelity, thereby reducing model-generated hallucinations. HASTE represents a key step towards enabling reliable and scalable AI-assisted software development.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 18930