提示词工程
模块化设计
本项目的提示词系统遵循标准化模块设计,每个游戏的提示词由 4 个独立函数组成,存放在 prompt/prompt_for_xxx.py 中。
prompt/
├── prompt_for_simple.py # 4 个标准函数
├── prompt_for_spread.py
├── prompt_for_adv.py
├── prompt_for_push.py
├── prompt_for_tag.py
├── prompt_for_crypto.py
├── prompt_for_reference.py
├── prompt_for_speaker_listener.py
└── prompt_for_world_comm.py四大标准模块
模块 A: 任务与奖励 — get_task_and_reward()
作用: 告诉 LLM "你在玩什么游戏,怎么得分"
典型内容:
- 角色定义(如 Predator / Prey)
- 游戏目标
- 奖励公式(量化指标)
- 胜利/失败条件
示例 (Simple):
python
def get_task_and_reward() -> str:
return (
"TASK / GOAL:\n"
"- Single agent must move to the single landmark.\n\n"
"REWARD (from source):\n"
"- reward = -||agent_pos - landmark_pos||^2 = -(dx^2 + dy^2).\n"
"- Maximize reward by driving dx, dy toward 0.\n"
)角色参数化示例 (Tag):
python
def get_task_and_reward(is_predator: bool) -> str:
if is_predator:
return (
"TASK: HUNTING (You are a PREDATOR / WOLF)\n"
"- Goal: Catch the Prey (Green).\n"
"REWARD: +10.0 for hitting the prey; -0.1 per step distance."
)
return (
"TASK: SURVIVAL (You are the PREY / SHEEP)\n"
"- Goal: Run away from Predators (Red).\n"
"REWARD: -10.0 if caught; +0.1 per safe step."
)模块 B: 物理规则 — get_physics_rules()
作用: 描述物理引擎参数,帮助 LLM 理解动力学
核心参数:
| 参数 | 值 | 含义 |
|---|---|---|
dt | 0.1 | 时间步长 |
damping | 0.25 | 阻尼系数 (每步速度衰减 25%) |
mass | 1.0 | 默认质量 |
sensitivity | 5.0 | 动作到力的映射 |
更新顺序:
(1) 位置更新: p ← p + v × dt (使用旧速度)
(2) 阻尼应用: v ← v × (1 - damping)
(3) 推力应用: v ← v + (F / mass) × dt模块 C: 动作格式 — get_action_and_response_format()
作用: 定义 LLM 的输出格式,包含 few-shot 示例
标准输出:
json
{"action": [a0, a1, a2, a3, a4], "notes": "Short Strategy"}Few-shot 示例:
| 场景 | 推荐动作 |
|---|---|
目标在左上 (dx<0, dy>0) | [0, 0.8, 0, 0, 0.8] |
目标在右下 (dx>0, dy<0) | [0, 0, 0.9, 0.9, 0] |
接近目标 (dist<0.2) | [0, 0, 0, 0, 0] (刹车) |
| 接近边界 (` | x |
模块 D: 导航策略 — get_navigation_hints()
作用: 提供角色特定的策略指导
通用内容:
- 相对坐标语义解释
- 动作互斥规则
- 边界处理 few-shot
- 制动 few-shot
角色特定内容:
- Predator: 追击策略、预判拦截
- Prey: 逃跑策略、边界警告
- Good Agent: 欺骗策略
- Adversary: 推理策略
完整提示词组装流程
以 spread_API.py 为例,每一步的提示词组装:
python
def user_prompt(agent_id, step, obs_struct):
# 1. 基础信息
prompt = f"Step {step}, You are {agent_id}.\n"
prompt += f"Observation:\n{json.dumps(obs_struct, indent=2)}\n\n"
# 2. 拼接 4 大模块
prompt += get_task_and_reward(num_agents=3, local_ratio=0.5) + "\n"
prompt += get_physics_rules() + "\n"
prompt += get_action_and_response_format() + "\n"
prompt += get_navigation_hints() + "\n"
return promptSystem Prompt
System Prompt 通常很简短,定义 LLM 的角色:
python
system_prompt = (
"You are an expert game-playing AI agent in a 2D physics environment. "
"Analyze your observation carefully and output a precise action "
"in the specified JSON format."
)提示词设计最佳实践
1. 明确的坐标语义
OBS RELATIVE COORD: obs gives (other - you).
Example: other at (0,1), you at (1,0) => relative = [-1, 1]始终在提示词中解释"相对坐标"的含义,避免 LLM 混淆方向。
2. 互斥动作规则
NEVER output opposing actions like [0, 0.6, 0.2, ...]
(Left & Right both active). This wastes force.
One side MUST be 0.0.3. 边界警告
BOUNDARY WARNING: Stay within [-1.0, 1.0].
If |x| or |y| > 0.9, thrust back toward center.4. 量化奖励公式
提供精确的数学公式而非模糊描述:
# ✅ 好的
reward = -||agent - landmark||^2 = -(dx^2 + dy^2)
# ❌ 不好的
reward: you get more points when closer to the landmark自定义提示词
如需为新游戏创建提示词,只需在 prompt/ 目录创建新文件并实现这 4 个函数:
python
# prompt/prompt_for_my_game.py
__all__ = [
"get_task_and_reward",
"get_action_and_response_format",
"get_physics_rules",
"get_navigation_hints",
]
def get_task_and_reward(**kwargs) -> str:
return "..."
def get_physics_rules(**kwargs) -> str:
return "..."
def get_action_and_response_format(**kwargs) -> str:
return "..."
def get_navigation_hints(**kwargs) -> str:
return "..."