WARP - Web-Augmented Real-time Program Repairer: A Real-Time Compilation Error Resolution using LLMs and Web-Augmented Synthesis

Anderson · 2025

Read WARP - Web-Augmented Real-time Program Repairer: A Real-Time Compilation Error Resolution using LLMs and Web-Augmented Synthesis by Anderson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-25192.txt.

Introduction

Anderson de Lima Luiz builds a real time repairer for compilation errors in this 2025 paper. Such errors stall programmers daily, and built in quick fixes in editors are shallow. His system, WARP, watches the developer's terminal and spots errors as they appear.

WARP pairs a fine tuned code language model with live web search, pulling fixes, explanations, and snippets from forums and official documents. On a curated benchmark of C, C++, Python, and Go errors, it got 72.5 percent of broken programs compiling again. Its fixes also matched the intended meaning better than model only baselines and editor quick fixes. The paper also weighs the hard part: building accurate fixes from noisy web material.

Catalog record

CREATORAnderson
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-25192

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

WARP - Web-Augmented Real-time Program Repairer: A Real-Time Compilation Error Resolution using LLMs and Web-Augmented Synthesis Anderson Abstract Compilation errors represent a significant bottleneck in software development productivity. This paper introduces WARP (Web-Augmented Real-time Program Repairer), a novel system that leverages Large Language Models (LLMs) and dynamic web-augmented synthesis for real-time resolution of these errors. WARP actively monitors developer terminals, intelligently detects compilation errors, and synergistically combines the understanding of a fine-tuned Code-LLM with relevant solutions, explanations, and code snippets retrieved from up-to-date web sources like developer forums and official documentation. Experimental results on our curated benchmark, CGP (featuring C/C++, Python, and Go errors), demonstrate WARP achieves a superior fix rate (72.5% Compiles correctly) and higher semantic correctness compared to baseline LLM-only approaches and traditional IDE quick-fixes. Key technical challenges in achieving high-accuracy synthesis from noisy web data. de Lima Luiz Germany keywords LLM, compilation errors, real-time, web-augmented synthesis, RAG, automated program repair, developer tools, AI for SE, program synthesis, context-aware AI † † email: [email protected] 1 Introduction The iterative nature of software development is frequently punctuated by compilation errors, whose diagnosis and resolution consume disproportionate developer time and cognitive resources [ 1 , 2 ] . Existing Integrated Development Environments (IDEs), while offering some assistance, often fall short with complex, context-dependent, or novel errors, frequently providing generic suggestions [ 3 ] . This paper introduces WARP (Web-Augmented Real-time Program-repairer), a system advancing automated compilation error resolution by synergizing Large Language Models (LLMs) [ 4 , 5 , 6 ] with vast, dynamic web knowledge. The primary contributions are: (i) a novel, comprehensive architecture for WARP; (ii) specific techniques for AST-based contextual feature extraction, prompt engineering for code-centric LLMs, and effective LLM-RAG synergy for compilation errors; and (iii) empirical validation of WARP’s efficacy on ComErrFix-CGP-v1.1, a new benchmark of C/C++, Python, and Go errors, showing significant improvements over baseline approaches. The remainder of this paper is structured as follows. Section 2 reviews related work. Section 3 details the architecture and methodology of WARP. Section 4 presents our experimental setup, benchmark, and results. Section 5 discusses challenges and future research directions. Finally, Section 6 concludes the paper. Figure 1: Interface of the WARP system, displaying a detected error, the LLM-generated explanation, suggested fix (diff), and supporting web evidence. This interface aims for clarity and actionable insights for the developer. 2 Related Work WARP intersects with real-time error analysis, Automated Program Repair (APR), LLMs applied to code, and Information Retrieval (IR) for software engineering. Contemporary IDEs integrate static analysis tools for real-time feedback [ 7 ] , but suggestions are often limited to simple syntactic errors, lacking deep semantic understanding or awareness of broader community solutions. APR [ 8 ] has a rich history, including search-based, template-based, and constraint-based methods. LLMs have spurred learning-based APR. DeepDelta [ 1 ] used NMT for specific Java compilation errors, predating modern LLMs. More recent tools like RustAssistant [ 9 ] leverage LLMs like GPT-4 [ 10 ] for Rust compilation errors via iterative compiler feedback; however, it is language-specific and lacks dynamic web search. Other systems like SRepair [ 11 ] and RepairAgent [ 12 ] use LLMs for runtime bugs from benchmarks like Defects4J [ 13 ] , typically operating offline and targeting different bug categories than WARP’s focus on real-time compilation error resolution augmented by web knowledge. Foundational LLMs like OpenAI’s Codex [ 5 ] and GPT series [ 10 ] , and Meta’s Llama family [ 6 , 14 ] , excel at code tasks but can ”hallucinate” [ 15 ] and may lack knowledge of recent API changes or specific error contexts post-training. Retrieval Augmented Generation (RAG) [ 16 ] mitigates these issues by grounding LLM responses in external information. Applying RAG effectively to diverse, noisy developer-focused web data for real-time code repair, as WARP does, poses challenges in query formulation, source selection, information extraction, low-latency processing, and synthesis of potentially conflicting information [ 17 , 18 ] . WARP’s architecture is designed to address these. 3 The WARP System: Architecture and Methodology WARP’s operational pipeline, depicted conceptually in Figure 2 , initiates with real-time capture of compiler outputs. Upon error detection, it formulates a rich contextual prompt incorporating the error message, Abstract Syntax Tree (AST)-derived code snippets, and project metadata. This context is fed to a fine-tuned CodeLlama-70B variant [ 14 ] for initial analysis and a candidate fix. Concurrently, WARP’s specialized Retrieval Augmented Generation (RAG) module [ 16 , 19 ] queries web sources like Stack Overflow, GitHub Issues [ 20 ] , and official documentation to retrieve relevant code examples, explanations, and solutions. In the final stage, the LLM synthesizes its pre-trained knowledge with this retrieved web intelligence to generate ranked potential fixes (as diffs), detailed explanations, and references to supporting web evidence, presented via the interface shown in Figure 1 . The system synergizes deep LLM analytics with dynamic web knowledge for real-time, low-latency operation. Figure 2: High-level architecture of the WARP system, illustrating information flow from real-time error capture, through contextual feature extraction, initial LLM analysis, web augmentation via the WAR-Module, and final LLM-driven synthesis, before presenting the solution to the developer. 3.1 Real-time Error Ingestion and Contextual Feature Extraction WARP monitors ‘stdout‘ and ‘stderr‘ from compilation commands (e.g., ‘gcc‘, ‘python script.py‘) or integrates with Language Server Protocol (LSP) implementations. On detecting error signatures, the Ingestion module parses compiler/interpreter output to extract a structured error context, E_{ctx} . This context includes: • Canonicalized Error Identifier ( E_{id} ) : Standardized error type (e.g., C_SEMICOLON_EXPECTED ). • Tokenized Error Message ( M_{tok} ) : Primary error message, tokenized. • File Path ( F_{path} ) : Source file path. • Line Number ( L_{num} ) : Error line number. • AST-based Code Context ( C_{AST} ) : AST nodes around L_{num}\pm k lines, enclosing function, or block, generated using a parser (e.g., ‘tree-sitter‘). Provides richer structural information than raw text. • Programming Language ( L_{lang} ) : Inferred from file extension or command. • Project-Specific Metadata ( P_{meta} ) : Dependencies, library versions, compiler options from build files (e.g., ‘pom.xml‘, ‘requirements.txt‘) or flags. This E_{ctx}=f_{extract}(\text{raw\_compiler\_output},\text{source\_file\_content},\text{project\_config}) is input for subsequent stages. 3.2 LLM-Internal Hypothesis Generation The structured E_{ctx} is formatted into an initial prompt, P_{hypo} , processed by a core LLM, \text{LLM}_{core} (e.g., a fine-tuned CodeLlama-70B variant or similar). This \text{LLM}_{core} is fine-tuned on a large dataset of erroneous code snippets, error messages, metadata, corrected code diffs, explanations, and error categories to maximize correction accuracy and enhance explanation generation in diff format. P_{hypo} guides \text{LLM}_{core} for analysis and candidate fix generation. A template: Task: Analyze and resolve a compilation error. Language: ${L_lang} Error Type: ${E_id} Error Message: "${M_tok}" File: ${F_path} at Line: ${L_num} AST Context Snippet: ${C_AST} Relevant Project Metadata: ${P_meta} Instruction: Provide a concise textual explanation of the root cause. Then, provide a suggested code modification in ’diff -u’ format. Estimate confidence. \text{LLM}_{core} generates S_{hypo}=(\text{Fix}_{hypo},\text{Expl}_{hypo},\text{Conf}_{hypo})=\text{LLM}_{core}(P_{hypo}) , containing the proposed diff, explanation, and an initial confidence score (e.g., from token log-probabilities or a confidence head). 3.3 Web-Augmented Retrieval (WAR) Module To complement \text{LLM}_{core} ’s knowledge, the WAR module, often invoked in parallel or post- S_{hypo} generation, performs: 1) Dynamic and Contextual Query Formulation : Generates diverse search queries \{Q_{1},...,Q_{N}\}

Read the complete text