Google - The Gemma 4 Developer Agent CompetitionGoogle - Gemma 4 开发者智能体竞赛

Note: there are two active competitions for the Gemma 4 Developer Agent Competition: this competition, and a paper track, where you can document your approach for this prediction competition.注意：Gemma 4 开发者智能体竞赛包含两项正在进行的竞赛：本项竞赛，以及一个论文赛道，您可以在该赛道中记录您在本预测竞赛中所使用的方法。

Today’s most capable coding agents rely on internet-connected models that require powerful commercial infrastructure. That puts advanced software engineering out of reach for many developers.当今最强大的编程智能体依赖于需要强大商业基础设施支持的联网模型。这使得许多开发者无法使用先进的软件工程工具。

Small local models have made rapid progress, but they still struggle with complex, multi-turn software engineering. Large codebases are hard to navigate, so fixing bugs and adding features reliably is tough for models that fit on a single accelerator.小型本地模型已取得快速进展，但在处理复杂的多轮软件工程任务时仍存在困难。大型代码库难以浏览，因此对于那些只能在单个加速器上运行的模型来说，可靠地修复漏洞和添加功能具有很大挑战性。

In this competition, you’ll post-train Gemma 4 to build autonomous software engineering agents designed to amplify developers’ talents. You’ll explore new fine-tuning and reinforcement learning techniques to create agents that can independently navigate repositories and draft solutions, freeing up human engineers to focus on architecture, logic, and review.在本次竞赛中，您将对 Gemma 4 进行训练后微调，以构建旨在增强开发者才华的自主软件工程智能体。您将探索新的微调和强化学习技术，以创建能够独立浏览存储库并起草解决方案的智能体，从而让工程师能够专注于架构、逻辑和代码审查。

Your solution could help put powerful coding assistants in the hands of millions more developers. Sharing how you did it makes advanced coding tools accessible to everyone. 您的解决方案可能有助于让数百万更多的开发者使用上强大的编程助手。分享您的实现过程，能让每个人都能接触到先进的编程工具。

You must provide a zip archive (submission.zip) that contains your Agent Config, comprising system prompts, custom tools, skills, and LoRA adapters. An agent.yaml file must be located at the root of the archive.您必须提供一个包含智能体配置（submission.zip）的压缩包，其中应包括系统提示词、自定义工具、技能和 LoRA 适配器。压缩包的根目录下必须包含一个 agent.yaml 文件。

submission.zip
├── agent.yaml                  # REQUIRED: Root agent config
├── configs/
│   └── sampling.yaml           # Optional: Generation parameters loaded via !include
├── prompts/
│   ├── system.md               # Optional: System instructions loaded via !include
│   └── analyzer.md
├── sub_agents/
│   └── code_analyzer.yaml      # Optional: Sub-agent or AgentTool YAML configurations
├── adapters/                   # Optional: Fine-tuned PEFT LoRA adapters or model weights
│   ├── main_lora/
│   │   ├── adapter_config.json
│   │   └── adapter_model.safetensors
│   └── tool_lora/
│       ├── adapter_config.json
│       └── adapter_model.safetensors
└── skills/                     # Optional: ADK Skill directories
    └── repo_navigation/
        ├── SKILL.md
        ├── scripts/            # Python or bash scripts executed in a sandbox
        └── resources/          # Domain knowledge markdown files
content_copy

The config language follows the Google ADK Agent Config specification with additional restrictions to prevent code execution outside of a sandbox. The submissions themselves are compiled into ADK agents to be evaluated.配置语言遵循 Google ADK 智能体配置规范，并附加了限制条件以防止在沙箱之外执行代码。提交的内容将被编译为 ADK 智能体进行评估。

Issue Scoring问题评分

Similar to SWE-Bench, your agent's submitted patches are evaluated by a PASS/FAIL metric. For each issue, the submitted patch is applied to that issue's code repository and that issue's validation tests are run. Your submission's overall score is the percentage of patched repositories that pass the validation tests.与 SWE-Bench 类似，您提交的智能体补丁将通过“通过/失败”（PASS/FAIL）指标进行评估。对于每个问题，提交的补丁将被应用到该问题的代码库中，并运行该问题的验证测试。您提交内容的总体得分是成功通过验证测试的补丁库所占的百分比。

Your agent has a limit of 12 hours to submit patches for all tasks, inclusive of sandbox setup time, but excluding time for patch validation. You may, but are not required, to set per-task time limits in your submission's eval_config.yaml.您的智能体在提交所有任务补丁时有 12 小时的限时，其中包括沙箱设置时间，但不包括补丁验证时间。您可以（但并非必须）在提交内容的 eval_config.yaml 中设置每个任务的时间限制。

• September 23, 2026 - Start Date.2026 年 9 月 23 日 - 开始日期。
• November 12, 2026 (optional) - Deadline to submit Research Paper.2026 年 11 月 12 日（可选）- 提交研究论文的截止日期。
• November 25, 2026 - Entry Deadline. You must accept the competition rules before this date in order to compete.2026 年 11 月 25 日 - 参赛截止日期。您必须在此日期之前接受竞赛规则才能参赛。
• November 25, 2026 - Team Merger Deadline. This is the last day participants may join or merge teams.2026 年 11 月 25 日 - 团队合并截止日期。这是参赛者加入或合并团队的最后一天。
• December 2, 2026 - Final Submission Deadline.2026 年 12 月 2 日 - 最终提交截止日期。

All deadlines are at 11:59 PM UTC on the corresponding day unless otherwise noted. The competition organizers reserve the right to update the contest timeline if they deem it necessary.除非另有说明，所有截止日期均为相应日期的晚上 11:59（UTC 时间）。竞赛组织者保留在必要时更新竞赛时间表的权利。

Total Prizes Available: $100,000 奖金总额：100,000 美元

The Gemma 4 Developer Agent Competition: $65,000Gemma 4 开发者智能体竞赛：65,000 美元

• 1st Place: $37,000第一名：37,000 美元
• 2nd Place: $18,000第二名：18,000 美元
• 3rd Place: $10,000第三名：10,000 美元

Paper Submission Award (optional): $35,000 论文提交奖（可选）：35,000 美元

Participants are encouraged to submit original, unpublished research submissions on topics including, but not limited to:鼓励参赛者提交原创、未发表的研究报告，主题包括但不限于：

• Tuning & Optimization: Parameter-efficient fine-tuning and Reinforcement Learning (RL) approaches tailored to improve SWE agent capabilities.调优与优化：旨在提高软件工程（SWE）智能体能力的参数高效微调和强化学习（RL）方法。
• Code Comprehension: Novel techniques for code-graph generation, parsing, and embedding to enhance deep comprehension of complex repositories.代码理解：用于代码图生成、解析和嵌入的新颖技术，以增强对复杂存储库的深度理解。
• Tasks & Benchmarks: Development of new tasks, evaluation datasets, or resources targeting code structure, graphs, and structured code generation.任务与基准：开发针对代码结构、图和结构化代码生成的新任务、评估数据集或资源。
• Graph Reasoning: Methods and architectures for improving Large Language Model (LLM) reasoning over large-scale graphs.图推理：改进大语言模型（LLM）在大型图上进行推理的方法和架构。

Top submissions will be highlighted at a Google hosted event (such as a NeurIPS expo workshop) and awarded prizes. Submissions leveraging the accompanying released graph and embedding dataset are highly encouraged, but not a requirement for submission. Submissions must be made by November 12, 2026 at 11:59PM UTC.顶级提交作品将在 Google 主办的活动（如 NeurIPS 博览会研讨会）上进行展示并获得奖励。强烈鼓励提交利用随附发布的图和嵌入数据集的作品，但这并非提交的必要条件。提交截止日期为 2026 年 11 月 12 日晚上 11:59（UTC 时间）。

To submit, sign up for the paper track. 如需提交，请注册论文赛道。

Model Selection and LoRA Adapters模型选择和 LoRA 适配器

At present, we only support the gemma-4-31b-it-qat-w4a16-ct Gemma 4 model variant. You must choose this model for every agent and subagent.目前，我们仅支持 gemma-4-31b-it-qat-w4a16-ct 这一 Gemma 4 模型变体。您必须为每个智能体和子智能体选择此模型。

You may, but are not required, to include one or more .safetensors format LoRA adapters with your submission. While you are restricted to a single base model, you may use different adapters for each agent in your submission:您可以（但并非必须）在提交内容中包含一个或多个 .safetensors 格式的 LoRA 适配器。虽然您仅限于使用单一基础模型，但您可以为提交内容中的每个智能体使用不同的适配器：

1. Place PEFT LoRA directories (containing adapter_config.json and adapter_model.safetensors) inside
   adapters/<adapter_name>/.将 PEFT LoRA 目录（包含 adapter_config.json 和 adapter_model.safetensors）放置在 adapters/<adapter_name>/ 文件夹内。
2. Reference adapter: <adapter_name> on any LlmAgent in agent.yaml or sub_agents/*.yaml (for example,
   adapter: main_lora on the root coder agent and adapter: tool_lora on a read-only analyzer AgentTool).在 agent.yaml 或 sub_agents/*.yaml 中的任何 LlmAgent 上引用适配器：<adapter_name>（例如，在根 coder 智能体上使用 adapter: main_lora，在只读 analyzer AgentTool 上使用 adapter: tool_lora）。

Tool, Prompt, and Skill Rules工具、提示词和技能规则

• !include Directives: Resolves file paths relative to the directory of the file containing the tag. For instance, in agent.yaml, !include prompts/system.md loads prompts/system.md.!include 指令：解析相对于包含该标签文件的目录的文件路径。例如，在 agent.yaml 中，!include prompts/system.md 会加载 prompts/system.md。
• Sandboxing: Path traversal outside the submission root (e.g., via ../ or symlinks) is not allowed.沙箱：不允许在提交根目录之外进行路径遍历（例如通过 ../ 或符号链接）。
• Allowed Tools: Your agent can only request tools provided by the competition harness or custom subagents defined via agent_tool.允许使用的工具：您的智能体只能请求由竞赛测试框架提供的工具，或通过 agent_tool 定义的自定义子智能体。
• Skill Structure & Sandboxing: Each skill must be a directory containing a SKILL.md manifest with YAML frontmatter (name: <skill-name>). Scripts executed via run_skill_script run securely inside the competition's persistent Docker container, sharing filesystem access with run_command and debiting execution time against your central budget. Agents can inspect domain knowledge files using load_skill_resource.技能结构与沙箱：每项技能必须是一个包含带有 YAML 元数据的 SKILL.md 清单（name: <skill-name>）的目录。通过 run_skill_script 执行的脚本在竞赛的持久化 Docker 容器内安全运行，与 run_command 共享文件系统访问权限，并从您的中心预算中扣除执行时间。智能体可以使用 load_skill_resource 检查领域知识文件。

The predefined tools available to your agent are:您的智能体可使用的预定义工具包括：

1. run_command(command: str) -> str - Executes a shell command in /bin/bash -c inside /workspace.run_command(command: str) -> str - 在 /workspace 内执行 /bin/bash -c shell 命令。
2. submit_patch() -> str - Stages untracked file intents (git add -N .) and captures git diff HEAD from /workspace.submit_patch() -> str - 暂存未追踪的文件意图（git add -N .）并捕获 /workspace 中的 git diff HEAD。
3. get_status() -> str - Returns live budget consumption and patch status.get_status() -> str - 返回实时预算消耗情况和补丁状态。
4. read_file(filepath: str, start_line: int | None = None, end_line: int | None = None) -> str - Reads a file from /workspace with 1-indexed inclusive line slicing.read_file(filepath: str, start_line: int | None = None, end_line: int | None = None) -> str - 从 /workspace 读取文件，支持以 1 为起始索引的包含性行切片。
5. edit_file(filepath: str, old_string: str, new_string: str, allow_multiple: bool = False) -> str - Replaces old_string with new_string in an existing non-empty file inside /workspace.edit_file(filepath: str, old_string: str, new_string: str, allow_multiple: bool = False) -> str - 将 /workspace 内现有非空文件中的 old_string 替换为 new_string。
6. write_file(filepath: str, content: str) -> str - Creates or overwrites a file at /workspace/<filepath>, automatically creating parent directories (mkdir -p).write_file(filepath: str, content: str) -> str - 在 /workspace/<filepath> 创建或覆盖文件，并自动创建父目录（mkdir -p）。
7. get_code_neighbors(node: str, edge_type: str | None = None, max_neighbors: int = 50) -> str - Finds incoming and outgoing neighbors of a symbol (node) in the repository call/dependency graph.get_code_neighbors(node: str, edge_type: str | None = None, max_neighbors: int = 50) -> str - 在存储库调用/依赖图中查找符号（节点）的入度和出度邻居。
8. search_similar_code(query: str, k: int = 10) -> str - Finds top-k graph nodes with highest cosine similarity to query in the pre-computed embeddings.search_similar_code(query: str, k: int = 10) -> str - 在预计算的嵌入中查找与查询余弦相似度最高的 top-k 个图节点。
9. get_code_subgraph(nodes: list[str]) -> str - Extracts the induced subgraph (all nodes and interconnecting edges) for a list of symbols.get_code_subgraph(nodes: list[str]) -> str - 为一组符号提取诱导子图（所有节点及互连边）。

See the HARNESS_README.md file in the dataset for detailed information about the submission format and execution environment.有关提交格式和执行环境的详细信息，请参阅数据集中的 HARNESS_README.md 文件。

This competition has access to Kaggle's pool of powerful new L4x4 machines! These machines offer 96GB of GPU memory enabling submissions with much larger models. See this page for more on these machines.本次竞赛可以使用 Kaggle 强大的新型 L4x4 机器资源池！这些机器提供 96GB 的 GPU 显存，使提交内容能够使用规模更大的模型。有关这些机器的更多信息，请参阅此页面。

What you need to know:您需要了解的内容：

• Quota usage - Due to their limited availability, notebooks with L4x4s consume GPU quota at twice the rate of the older T4x2 and P100 machines. We may increase this rate as necessary to ensure these machines are available for submission scoring.配额使用 - 由于可用性有限，使用 L4x4 的笔记本电脑消耗 GPU 配额的速度是旧款 T4x2 和 P100 机器的两倍。我们可能会根据需要提高此速率，以确保这些机器可用于提交评分。
• Restricted Use - L4s are only available for notebooks attached to this competition. We will build tooling to enforce this if necessary. In the meantime, attempts to circumvent this restriction may be enforced by Kaggle moderation with consequences including team bans from the competition or account bans.限制使用 - L4 仅适用于连接到本次竞赛的笔记本电脑。如有必要，我们将构建工具来强制执行此规定。在此期间，Kaggle 管理部门可能会对规避此限制的行为采取强制措施，后果包括团队被禁止参加竞赛或封禁账户。
• No Internet - All L4 sessions must have internet disabled.无网络 - 所有 L4 会话必须禁用互联网。