Code LLM Papers Guide

Code LLM Papers Guide

Overview

This curated collection covers LLMs for code — from foundational models (Codex, CodeGen, StarCoder) through code generation, completion, repair, translation, and understanding. Accompanies a TMLR survey paper providing systematic categorization. Tracks 500+ papers across pre-training, fine-tuning, evaluation, and application of code-focused language models.

Taxonomy

Code LLMs
├── Pre-training
│   ├── Encoder-only (CodeBERT, GraphCodeBERT)
│   ├── Decoder-only (Codex, CodeGen, StarCoder, DeepSeek-Coder)
│   └── Encoder-Decoder (CodeT5, PLBART)
├── Fine-tuning & Alignment
│   ├── Instruction tuning (WizardCoder, Magicoder)
│   ├── RLHF for code (CodeRL)
│   └── Self-play (AlphaCode)
├── Applications
│   ├── Code generation (NL → Code)
│   ├── Code completion (infilling)
│   ├── Code repair (bug fixing)
│   ├── Code translation (language conversion)
│   ├── Code summarization (Code → NL)
│   ├── Test generation
│   └── Code review
└── Evaluation
    ├── Benchmarks (HumanEval, MBPP, SWE-bench)
    ├── Metrics (pass@k, CodeBLEU)
    └── Security analysis

Key Models Timeline

Benchmark Tracking

# Track model performance on HumanEval
humaneval_scores = {
    "GPT-4": {"pass_at_1": 67.0, "pass_at_10": 86.0},
    "Claude 3.5 Sonnet": {"pass_at_1": 64.0},
    "DeepSeek-Coder-33B": {"pass_at_1": 56.1},
    "CodeLlama-34B": {"pass_at_1": 48.8},
    "StarCoder2-15B": {"pass_at_1": 46.3},
    "GPT-3.5-Turbo": {"pass_at_1": 48.1},
}

print(f"{'Model':<25} {'pass@1':>8} {'pass@10':>8}")
print("-" * 43)
for model, scores in sorted(
    humaneval_scores.items(),
    key=lambda x: x[1].get("pass_at_1", 0),
    reverse=True,
):
    p1 = scores.get("pass_at_1", "—")
    p10 = scores.get("pass_at_10", "—")
    print(f"{model:<25} {str(p1):>8} {str(p10):>8}")

Research Directions

### Active Areas (2024-2025)
1. **Repository-level generation** — Understanding full codebases
2. **Agentic coding** — LLMs using tools (debugger, terminal)
3. **Formal verification** — Proving correctness of generated code
4. **Multi-language** — Cross-language transfer and translation
5. **Security** — Detecting and avoiding vulnerable code
6. **Long context** — Processing large codebases (100k+ tokens)
7. **Code editing** — Natural language instructions for code changes

Paper Search

import arxiv

def find_code_llm_papers(topic="code generation", max_results=20):
    """Find recent Code LLM papers on arXiv."""
    query = f"abs:{topic} AND (abs:large language model OR abs:LLM)"

    search = arxiv.Search(
        query=query,
        max_results=max_results,
        sort_by=arxiv.SortCriterion.SubmittedDate,
    )

    for result in search.results():
        print(f"[{result.published.strftime('%Y-%m-%d')}] "
              f"{result.title}")

find_code_llm_papers("code generation")
find_code_llm_papers("automated program repair")

Use Cases

1. Literature survey: Map the Code LLM research landscape
2. Model selection: Compare code models for specific tasks
3. Benchmark analysis: Track state-of-the-art on standard benchmarks
4. Research planning: Identify open problems and trends
5. Course material: Teach software engineering + AI intersection

References

• Awesome-Code-LLM
• TMLR Survey Paper
• HumanEval
• SWE-bench