Research map / Large language models / Agents, code and math
Agnostics: Learning to Synthesize Code in Any Programming Language with a Universal Reinforcement Learning Environment
Aleksander Boruch-Gruszecki, Yangtian Zi, Zixuan Wu, Tejas Oberoi, Carolyn Jane Anderson, Joydeep Biswas, Arjun Guha
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 |
|---|---|---|---|
| train with Agnostics RL on Ag-Codeforces-X (Fortran) | Qwen3-4B · Ag-LiveCodeBench-Fortran | pass@1: 0 -> 15 | Improved |
| train with Agnostics RL on Ag-Codeforces-X (Lua) | Qwen3-4B · Ag-LiveCodeBench-Lua | pass@1: 11 -> 23 | Improved |
| scale base model from 4B to 8B | Agnostics-trained Qwen3 · Ag-LiveCodeBench-R | pass@1: 15 -> 19 | Improved |
| train on easier MBPP problems instead of Codeforces | Qwen3-4B · Ag-LiveCodeBench-R | pass@1: 10 -> 9 | Hurt |
| train on easier MBPP problems instead of Codeforces | Qwen3-4B · MultiPL-E Julia | pass@1: 51 -> 62 | Improved |
| train on Codeforces rather than MBPP | Qwen3-4B · Ag-LiveCodeBench-Lua | pass@1: 15 -> 23 | Improved |
| apply Agnostics to non-Qwen model | DeepSeek Coder 6.7B Ins · MultiPL-E R | pass@1: 37 -> 52 | Improved |
| apply Agnostics to non-Qwen model | SmolLM3 3B · Ag-LiveCodeBench-Lua | pass@1: 1 -> 8 | Improved |
| RL training via Agnostics instead of distilling from Sonnet 4 Thinking | Qwen3-4B Fortran · Ag-LiveCodeBench-Fortran | pass@1: 2 -> 15 | Improved |
| add partial reward for runnable but wrong programs (ablation) | GRPO reward · Ag-LiveCodeBench-Lua | pass@1: 23.00 -> 18.57 | Hurt |
| add partial reward for runnable but wrong programs (ablation) | GRPO reward · Ag-Codeforces-Lua test | pass@1: 24.76 -> 13.33 | Hurt |
| increase GRPO group size from 32 to 64 (ablation) | Qwen3-4B-CF-Lua · Ag-LiveCodeBench-Lua | pass@1: 23.00 vs 21.71 | No clear effect |
Closest papers by meaning
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.88)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.88)
- R1-Code-Interpreter: LLMs Reason with Code via Supervised and Multi-stage Reinforcement Learning (ICLR 2026 · similarity 0.87)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.87)
- Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models (ICLR 2025 · similarity 0.87)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 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.
