NeurIPS 2025 Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization

Reinforcement learning on execution feedback makes generated code progressively faster without losing correctness.

ICML 2026 SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?

Benchmarks whether coding agents can deliver measurable performance gains on real-world repositories.

NeurIPS 2025 EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code

Measures the efficiency of LLM-generated code across six languages against human expert solutions.

NeurIPS 2024 Mercury: A Code Efficiency Benchmark for Code Large Language Models

Scores generated solutions on both correctness and runtime with a distribution-aware efficiency metric.

Preprint · arXiv 2026 Ockhamareto: Pareto-Gated Segment-Level Credit Assignment for Concise Unit-Test Generation with Reinforcement Learning

Rewards fault detection and brevity together so models generate smaller, more effective unit-test suites.

Preprint · arXiv 2026 Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles

Injects known GPU-kernel faults to reveal which bugs benchmark correctness checkers fail to catch.

DL4C @ ICML 2026 · Workshop Nexus: Execution-Grounded Multi-Agent Test Oracle Synthesis

Coordinates specialist agents with sandboxed execution to synthesize and refine test oracles.

TOSEM 2026 Benchmarking LLMs for Unit Test Generation from Real-World Functions

A contamination-aware benchmark of unit-test generation built from complex real-world functions.

ICSE 2026 Measuring the Influence of Incorrect Code on Test Generation

Faulty code in the prompt degrades generated tests, and natural-language descriptions help recover.

Preprint · arXiv 2026 Mastermind: Strategy-Level Learning for Repository-Scale Vulnerability Reproduction

Teaches agents reusable investigation strategies for reproducing and verifying repository-scale vulnerabilities.

Preprint · arXiv 2026 Secure Code Generation via Online Reinforcement Learning with Vulnerability Reward Model

A vulnerability-aware reward trains code models to produce secure code without losing functionality.

ACL 2026 · Oral · Best Paper nomination SeCuRepair: Semantics-Aligned, Curriculum-Driven, and Reasoning-Enhanced Vulnerability Repair Framework

Reason-then-edit vulnerability repair trained with semantic rewards and a difficulty curriculum.

NeurIPS 2026 CodeScaler: Scaling Code LLM Training and Test-Time Inference via Execution-Free Reward Models

An execution-free reward model scales code-model training and ranks candidate solutions at inference time.

ACL 2025 CodeArena: A Collective Evaluation Platform for LLM Code Generation

A collective, continuously refreshed evaluation platform that limits benchmark leakage in code generation.