ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

ProjAgent is a novel repository-level code generation system that introduces procedural similarity as a key retrieval signal, addressing challenges posed by complex cross-file dependencies and project-specific conventions. It achieves 41.14% Pass@1 on the REPOCOD benchmark, outperforming existing retrieval-based baselines, including SpecAgent, by 6.62%. The system decomposes target functions into intermediate reasoning steps and employs an agentic workflow to retrieve repository functions exhibiting similar procedural behavior, even when lexical or semantic similarity is low. This procedural context is integrated with conventional semantic retrieval and a conservative static-analysis feedback loop for iterative code repair, demonstrating that procedural similarity is an effective and previously unexplored dimension for improving repository-level code generation.

Key takeaway

For ML Engineers developing repository-level code generation systems, integrating procedural similarity is crucial. ProjAgent's 6.62% Pass@1 improvement over SpecAgent on REPOCOD highlights that focusing on underlying computational patterns, not just surface-level text, yields superior results. Consider adopting agentic workflows and static analysis feedback to enhance your context retrieval and code correctness.

Key insights

Procedural similarity, identified via LLM hidden-state projections, significantly enhances repository-level code generation.

Principles

• Procedural similarity complements lexical and semantic retrieval.
• LLM hidden-state projections capture underlying code logic.
• Agentic workflows improve context discovery and refinement.

Method

ProjAgent decomposes target functions into steps, uses an agentic workflow with hidden-state projection similarity for procedural context, combines it with BM25/dense semantic retrieval, and refines code via static-analysis feedback.

In practice

• Use LLM hidden-state projections for behavioral code similarity.
• Combine procedural and semantic context for robust generation.
• Implement conservative static-analysis feedback loops.

Topics

• Repository-Level Code Generation
• Procedural Similarity
• LLM Hidden States
• Agentic Code Generation
• Code Context Retrieval
• Static Analysis Feedback

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

Related on AIssential

• ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation — Procedural similarity, not just lexical or semantic, enhances repository-level code generation by matching functional logic.
• Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation — Long-horizon agentic frameworks with shared memory improve code documentation consistency and efficiency.
• Contract-Coding: Towards Repo-Level Generation via Structured Symbolic Paradigm — Contract-Coding uses a formal "Language Contract" to enable parallel, intent-driven repository generation, mitigating context and hallucination issues.
• DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch — Automated, structured data generation with difficulty-aware filtering scales long-horizon software engineering training for LLM agents.
• On Accelerating Grounded Code Development for Research — Lexical search and structured workflows can effectively ground coding agents in dynamic, domain-specific research.
• Agent trajectories as programs: fingerprinting and programming coding-agent behavior — Agent problem-solving trajectories can be fingerprinted and analyzed procedurally to reveal distinct behavioral patterns beyond mere success rates.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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