Research map / Large language models / Agents, code and math
Bringing Code ALIVE: Optimizing Interactive Frontend Mini-Games via Automated Play and Reinforcement Learning at Scale
Jiajun Zhang, Yuheng Jing, Zeyu Cui, Hao Zheng, Wentao Chen, Kaixin Li, Jiaxi Yang, Tianbao Xie, Zeyao Ma, Tianyi Bai, KaShun SHUM, Lei Zhang and 7 more
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 |
|---|---|---|---|
| curate SFT data with ALIVE-Score (best of five Gemini-3-Pro candidates) and fine-tune | Qwen3-Next · ALIVE-Eval | ALIVE-Score: 25.4 -> 29.1 (+3.7) | Improved |
| curate SFT data with ALIVE-Score and fine-tune | Qwen3-Next · ArtifactsBench-Game | score: 49.2 -> 58.4 (+9.2) | Improved |
| add GRPO RL with ALIVE reward after SFT | ALIVE-Coder-SFT · ALIVE-Eval | ALIVE-Score: 29.1 -> 40.5 | Improved |
| add GRPO RL with ALIVE reward after SFT | ALIVE-Coder-SFT · ArtifactsBench-Game | score: 58.4 -> 61.2 | Improved |
| use ALIVE as RL reward instead of Hybrid Judge (code+visual static judge) (ablation) | RL on ALIVE-Coder-SFT · ALIVE-Eval test score | test score: Hybrid test score collapses; response length 7k -> over 14k tokens | Improved |
| use ALIVE as RL reward instead of Visual Judge (screenshots) (ablation) | RL on ALIVE-Coder-SFT · ALIVE-Eval test score | test score | Improved |
| ALIVE one-shot planning instead of GUI agent serial loop | evaluation framework · evaluation cost | token consumption / latency: approximately 12× lower token consumption and 22× lower latency | Improved |
| ALIVE instead of GUI agent for RL reward computation | RL reward pipeline · 4,500 prompts x 16 samples | cost ($): 4,800 -> 396 | Improved |
| allow up to three replanning attempts versus one (ablation) | ALIVE · evaluation accuracy | accuracy parity with GUI agent steps: 9-10 steps (Replan=1) -> 11-12 steps (Replan=3) | Improved |
| add replanning mechanism in SFT data curation (ablation) | ALIVE · SFT dataset selection | fraction of samples rectified: 6% of the total dataset rectified | Improved |
| use ALIVE planning instead of heuristic random monkey policies (ablation) | ALIVE · fully functional games | mechanism exposure | Improved |
| use ALIVE instead of code/visual/hybrid judges for human alignment (ablation) | evaluation judge · 500 mini-game queries preliminary study | Rank Agreement / Score Precision: baselines Rank Agreement <= 0.5; ALIVE > 0.70, Score Precision > 0.85 | Improved |
Closest papers by meaning
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.88)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.87)
- Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning (ICML 2024 · similarity 0.87)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.86)
- Evaluating LLMs in Open-Source Games (NeurIPS 2025 · similarity 0.86)
- CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI (NeurIPS 2026 · similarity 0.85)
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.
