Research map / Large language models / Instruction and fine-tuning
Random Selection Reveals Implicit Knowledge Consensus in Code Generation
Ren-Biao Liu, Xin-Ye Li, Hui Sun, Yali Du, Jiang-Tian Xue, Ming Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use random per-problem solution sampling instead of diversity or difficulty selectors | SFT on LeetCode multi-solution pool · LeetCode test Pass@1 (Qwen2.5-7B) | Pass@1: 3.90 -> 6.32 (vs K-Center) | Improved |
| Use K-Center, AST Coverage or IFD ranking instead of random selection | SFT code data selection · LeetCode, Qwen2.5-7B | Pass@1: 6.32 -> 3.90 / 3.73 / 3.42 | Hurt |
| Use K-Means or Kernel Herding instead of random | SFT code data selection · LeetCode, Qwen2.5-Coder-7B and DeepSeek-Coder-6.7B | Pass@1: 5.57 -> 6.14 (K-Means, Coder); 3.68 -> 4.56 (Herding, DeepSeek) | Mixed |
| Use random 11-solution subset instead of the full dataset | SFT · LeetCode test Pass@1 (Llama-3.1-8B) | Pass@1: 3.68 -> 4.04 | Improved |
| Use complex selectors (Facility Location, K-Means, Kernel Herding) at 32B scale | Qwen2.5-32B · LeetCode test Pass@10 | Pass@10: 31.58 -> 35.53 (Facility Location) | Improved |
| Use K-Means / Facility Location instead of Random on constrained Medium/Hard problems | Qwen2.5-7B · 59 tagged Medium/Hard problems | Pass@10: 6.78 -> 8.47 | Improved |
| Use Kernel Herding instead of Random on Hard problems | Qwen2.5-7B · LeetCode Hard (79 tasks) | Pass@10: 5.06 -> 6.33 | Improved |
| Use cosine embedding distance instead of Levenshtein/Jaccard/syntax distance in K-Center (ablation) | K-Center · Train Pass@1, Qwen2.5-7B | Train Pass@1: 49.41 -> 58.69 | Improved |
| Increase per-problem selection budget k (ablation) | Random selection · Qwen2.5-7B | Train Pass@1 / Test Pass@10 | Improved |
| Vary learning rate from 1e-6 to 5e-5 (ablation) | Random selection · Qwen2.5-7B | Test Pass@1: 3.42 (1e-6) -> 6.32 (1e-5) -> 5.35 (5e-5) | Mixed |
Closest papers by meaning
- CodeChemist: Test-Time Scaling for Low-Resource Code Generation via Functional Knowledge Transfer (ICML 2026 · similarity 0.88)
- RPM-MCTS: Knowledge-Retrieval as Process Reward Model with Monte Carlo Tree Search for Code Generation (AAAI 2026 · similarity 0.88)
- Simple Baselines are Competitive with Code Evolution (NeurIPS 2026 · similarity 0.88)
- Divide-and-Conquer Meets Consensus: Unleashing the Power of Functions in Code Generation (NeurIPS 2024 · similarity 0.87)
- Chain of Execution Supervision Promotes General Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.87)
- Planning in Natural Language Improves LLM Search for Code Generation (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
