Evolve your
Codex
on
Terminal-Bench 2
在
Terminal-Bench 2
上进化
Codex
RSIHub is a framework focused on providing the environment for the RSI process, integrating multiple RSI methods and benchmarks. RSIHub 是一个专注于为 RSI 过程提供环境的框架，集成了多种 RSI 方法与 benchmark。
$ git clone https://github.com/simple-agent-lab/RSIHub.git
$ cd RSIHub && ./scripts/setup_terminal_bench.sh ahe
$ ./scripts/run_recipe_demo.sh ahe
gen 0 seed train 60.0% baseline tagged
gen 2 candidate train 68.0% gate ✓ admitted
gen 5 best train 74.0% archive.jsonl stampedSeed and best train scores from the published Codex × AHE run; intermediate generations abridged. seed 和最佳 train 分数来自真实的 Codex × AHE 运行，中间几代有省略。
Results Benchmark 结果
Audited gains on real benchmarks. 在真实 benchmark 上的提升
Scores are percentages shown as seed → evolved agent. The train score is measured on the recipe’s training split; the full benchmark score is measured across the complete benchmark. All runs use a GPT-5.4-high target model and a GPT-5.4-xhigh Codex mutate operator. 分数格式为 seed → 进化后。Train 分数在 recipe 的训练集上得出，完整分数在整个 benchmark 上得出。所有实验的目标模型是 GPT-5.4-high，mutate 使用 GPT-5.4-xhigh 的 Codex。
Terminal-Bench 2
Tau³ Banking
Full results table 完整结果表格
| Benchmark Benchmark | Target agent 目标 agent | Method 方法 | Train score Train 分数 | Full benchmark score 完整 benchmark 分数 |
|---|---|---|---|---|
| Terminal-Bench 2 50 train / 19 gate / 20 sealed |
MiniSWE | AHE | 70.0% → 74.0% (+4.0%) |
55.1% → 56.2% (+1.1%) |
| Hyperagents | 58.0% → 68.0% (+10.0%) |
55.1% → 68.5% (+13.4%) |
||
| A-Evolve | 66.0% → 68.0% (+2.0%) |
55.1% → 65.2% (+10.1%) |
||
| GEPA | 58.0% → 68.0% (+10.0%) |
55.1% → 59.6% (+4.5%) |
||
| Codex | AHE | 60.0% → 74.0% (+14.0%) |
60.7% → 66.3% (+5.6%) |
|
| Hyperagents | 58.0% → 72.0% (+14.0%) |
60.7% → 70.8% (+10.1%) |
||
| A-Evolve | 62.0% → 62.0% (0.0%) |
60.7% → 64.0% (+3.3%) |
||
| GEPA | 64.0% → 64.0% (0.0%) |
60.7% → 60.7% (0.0%) |
||
| Tau³ Banking 50 train / 20 gate / 27 sealed |
MiniSWE | AHE | 34.0% → 36.0% (+2.0%) |
12.4% → 22.7% (+10.3%) |
| Hyperagents | 30.0% → 38.0% (+8.0%) |
12.4% → 28.9% (+16.5%) |
||
| A-Evolve | 30.0% → 34.0% (+4.0%) |
12.4% → 24.7% (+12.3%) |
||
| GEPA | 30.0% → 32.0% (+2.0%) |
12.4% → 22.7% (+10.3%) |
||
| Codex | AHE | 32.0% → 36.0% (+4.0%) |
11.3% → 26.8% (+15.5%) |
|
| Hyperagents | 34.0% → 36.0% (+2.0%) |
11.3% → 39.2% (+27.9%) |
||
| A-Evolve | 30.0% → 38.0% (+8.0%) |
11.3% → 13.4% (+2.1%) |
||
| GEPA | 10.0% → 16.0% (+6.0%) |
11.3% → 15.5% (+4.2%) |
How it works 工作原理
A typical iteration loop. 一个典型的迭代循环
A recipe decides how parents are selected, how traces are analyzed, what may be modified, and how evaluation runs. The framework carries those decisions: managing candidates, running evals, and archiving changes. Recipe 决定如何选择父代、如何分析轨迹、修改范围、以及如何eval。框架负责承载这些决策：管理 candidate、运行 eval、归档改动。
Five built-in recipes compose the loop for different goals: 内置的五个 recipe 各有侧重：
hill_climb
Improve one candidate from its current best parent. 在当前最优 candidate 上持续改进。
Classic loop 经典循环See the recipe guide for each strategy’s workflow and configuration. 每个策略的用法和配置见 recipe 指南。
Showcase
A Skill that learned poster design. 一个学会海报设计的 Skill。
RSIHub can improve a Skill as a complete package: instructions, references, and validation scripts evolve together while a frozen evaluator keeps the comparison honest. In this local Paper2Poster run, the same Codex model and paper prompt produced both LoRA posters below. RSIHub 可以把 Skill 当成一个整体来进化：指令、参考资料、验证脚本一起改，eval 保持冻结来保证对比公平。下面两张 LoRA 海报来自同一个 Codex 模型和同一条 prompt，只是 Skill 的代数不同。
Across the four-paper showcase, the deterministic completion pass rate moved from 1/4 at Gen 0 to 4/4 at Gen 2. This is a representative evolution run rather than a broad benchmark; see the result snapshot, frozen rubric, and minimal seed Skill. 四篇论文的测试中，通过率从 Gen 0 的 1/4 涨到 Gen 2 的 4/4。这只是一次示例运行，不是严格的 benchmark；结果快照、评分标准和种子 Skill 都在仓库里。
Build agents that improve —
and keep the evidence.
让智能体持续改进，
并保留证据。
Clone the repository and run your first generation in an afternoon. 克隆仓库，一个下午跑通第一代。
git clone https://github.com/simple-agent-lab/RSIHub.git