Reported Scores
| Model | Score | Source paper | Year |
|---|---|---|---|
| Qwen2.5 | 96.3% | ReflexiCoder: Teaching Large Language Models to Self-Reflect on Generated Code and Self-Correct It via Reinforcement Learning / arxiv.org | 2026 |
| Qwen2.5 | 82.2% | FeedbackEval: A Benchmark for Evaluating Large Language Models in Feedback-Driven Code Repair Tasks / arxiv.org | 2025 |
| Qwen2.5 | 79.6% | FeedbackEval: A Benchmark for Evaluating Large Language Models in Feedback-Driven Code Repair Tasks / arxiv.org | 2025 |
| Qwen2.5 | 59.6% | HumanEval Pro and MBPP Pro: Evaluating Large Language Models on Self-invoking Code Generation / arxiv.org | 2024 |
| Qwen2.5 | 59.1% | Qwen2.5 Technical Report / arxiv.org | 2024 |
| Qwen2.5 | 41.0% | Assessing Small Language Models for Code Generation: An Empirical Study with Benchmarks / arxiv.org | 2025 |
| Qwen2.5 | 32.6% | LLaDA-MoE: A Sparse MoE Diffusion Language Model / arxiv.org | 2025 |
Interpretation
This page groups score claims extracted from papers for the same model and benchmark label. A nonzero spread means the public literature reports different values for this cluster.
Differences are not automatically errors. They may come from prompt choices, dataset versions, evaluation protocol, scoring rule, preprocessing, fine-tuning, or reporting convention. Source papers remain authoritative for their own claims. See the quality guide for how to read evidence links, manifests, and automated assessment fields.
Source coverage is a conservative count of distinct public paper URLs or titles in the cluster. It measures coverage breadth, not correctness.
Source profile reports public URL domains and publication years when they are available in extracted records. It is included for auditability only.