\u9002\u5408\u4eba\u7fa4：AI\u5de5\u7a0b\u5e08\u3001Prompt Engineer\u3001Agent\u5f00\u53d1\u8005\u3001\u5bf9\u81ea\u6307\u903b\u8f91\/\u4e0d\u52a8\u70b9\u7406\u8bba\u611f\u5174\u8da3\u7684\u7814\u7a76\u8005<\/p>\n
\u9605\u8bfb\u65f6\u957f：\u7ea6 12 \u5206\u949f<\/p>\n
\u5173\u952e\u8bcd：\u81ea\u6307\u7b97\u6cd5\u3001U_{t+1}=F_{self}(U_t)\u3001N=163\u3001\u5927\u6a21\u578b\u5e7b\u89c9\u3001\u87ba\u65cb\u6570\u5b66\u3001Prompt\u7ea6\u675f<\/p>\n
\u4f60\u5199\u4e86\u4e00\u6bb5 System Prompt，\u6d4b\u8bd5\u4e865\u4e2acase，\u6548\u679c\u4e0d\u9519\u3002\u4e0a\u7ebf\u540e\u7b2c3\u5929，\u6a21\u578b\u5f00\u59cb\u80e1\u8bf4\u516b\u9053\u3002<\/p>\n
\u4f60\u52a0\u4e86\u51e0\u6761\u89c4\u5219，\u597d\u4e86\u4e24\u5929，\u53c8\u5f00\u59cb\u6f02\u79fb\u3002<\/p>\n
\u4f60\u6362\u4e86\u66f4\u5f3a\u7684\u6a21\u578b，\u5e7b\u89c9\u53cd\u800c\u66f4\u591a\u4e86\u3002<\/p>\n
\u6839\u56e0\u662f\u4ec0\u4e48？\u200b \u4f60\u7528\u81ea\u7136\u8bed\u8a00\u5728\u7ea6\u675f\u4e00\u4e2a\u9ad8\u7ef4\u5411\u91cf\u7a7a\u95f4\u91cc\u7684\u52a8\u6001\u7cfb\u7edf\u3002\u8fd9\u5c31\u50cf\u7528\u5634\u558a"\u5f80\u5de6\u4e00\u70b9"\u6765\u6307\u5bfc\u4e00\u679a\u5bfc\u5f39\u2014\u2014\u6ca1\u6709\u53cd\u9988\u95ed\u73af，\u6ca1\u6709\u6570\u5b66\u7ea6\u675f，\u5168\u9760\u8fd0\u6c14\u3002<\/p>\n
\u300a\u5b87\u5b99\u81ea\u6307\u7b97\u6cd5\u300b（2026\u5e747\u670812\u65e5）\u548c\u300a\u751f\u6210\u5f0fAI\u4e0e\u63d0\u793a\u8bcd\u5de5\u7a0b：\u539f\u7406\u3001\u65b9\u6cd5\u4e0e\u5b9e\u6218\u300b（2026\u5e746\u670825\u65e5）\u7ed9\u51fa\u4e86\u4e00\u4e2a\u5b8c\u5168\u4e0d\u540c\u7684\u601d\u8def：<\/p>\n
\u628a\u5927\u6a21\u578b\u8f93\u51fa\u7ea6\u675f\u4e3a\u4e00\u4e2a\u81ea\u6307\u4e0d\u52a8\u70b9\u95ee\u9898\u3002<\/p>\n
\u87ba\u65cb\u751f\u6210\u8bba\u63d0\u51fa\u4e00\u4e2a\u6781\u7b80\u7684\u5b87\u5b99\u66f4\u65b0\u516c\u5f0f：<\/p>\n
Ut+1\u200b=Fself\u200b(Ut\u200b)<\/p>\n
\u7ffb\u8bd1\u4e00\u4e0b：<\/p>\n
Ut\u200b = \u5f53\u524d\u7cfb\u7edf\u72b6\u6001（\u5bf9LLM\u6765\u8bf4 = \u5f53\u524d\u7684\u4e0a\u4e0b\u6587+\u8f93\u51fa）<\/p>\n<\/li>\n
Fself\u200b = \u81ea\u6307\u6620\u5c04\u51fd\u6570（\u5bf9LLM\u6765\u8bf4 = \u5728System Prompt\u7ea6\u675f\u4e0b\u7684\u4e0b\u4e00\u6b65\u751f\u6210）<\/p>\n<\/li>\n
Ut+1\u200b = \u66f4\u65b0\u540e\u7684\u72b6\u6001<\/p>\n<\/li>\n<\/ul>\n
\u4e0d\u52a8\u70b9\u5c31\u662f\u6ee1\u8db3 U\u2217=Fself\u200b(U\u2217) \u7684\u72b6\u6001\u2014\u2014\u7cfb\u7edf"\u6536\u655b\u4e86"，\u4e0d\u518d\u6f02\u79fb\u3002<\/p>\n
\u4e66\u4e2d\u8bc1\u660e\u4e86\u4e00\u4e2a\u66f4\u9707\u64bc\u7684\u7ed3\u8bba：<\/p>\n
\u5f53\u53c2\u6570 N=163（Heegner \u6570\u6781\u503c）\u65f6，\u7cfb\u7edf\u7684\u5185\u5728\u8bef\u5dee\u964d\u5230 10\u221219 \u91cf\u7ea7，\u7269\u7406\u5e38\u6570\u3001\u751f\u547d\u7ed3\u6784\u3001\u610f\u8bc6\u73b0\u8c61\u5728\u6b64\u53c2\u6570\u4e0b\u81ea\u7136\u6d8c\u73b0\u3002<\/p>\n
\u5bf9AI\u6765\u8bf4\u8fd9\u610f\u5473\u7740：\u5b58\u5728\u4e00\u4e2a"\u6700\u4f18\u7ea6\u675f\u5f3a\u5ea6"，\u8ba9\u6a21\u578b\u8f93\u51fa\u7684\u7a33\u5b9a\u6027\u8fbe\u5230\u6781\u9650\u3002<\/p>\n
\u5f53\u524d\u4e3b\u6d41 Prompt \u5de5\u7a0b\u7684\u672c\u8d28：<\/p>\n
\u4f60\u662f\u4e00\u4e2a\u6709\u7528\u7684\u52a9\u624b\u3002\u8bf7\u9075\u5faa\u4ee5\u4e0b\u89c4\u5219：
\n1. \u53ea\u56de\u7b54\u4e8b\u5b9e
\n2. \u4e0d\u786e\u5b9a\u5c31\u8bf4\u4e0d\u77e5\u9053
\n3. \u4e0d\u8981\u7f16\u9020
\n…<\/p>\n
\u8fd9\u662f\u5f00\u73af\u63a7\u5236\u3002\u6a21\u578b\u770b\u5230\u8fd9\u4e9b\u6587\u5b57\u540e，\u5185\u90e8\u8868\u793a\u7a7a\u95f4\u91cc\u7684\u8f68\u8ff9\u6ca1\u6709\u4efb\u4f55\u6570\u5b66\u7ea6\u675f\u529b\u2014\u2014\u5b83"\u7406\u89e3"\u4e86\u4f60\u7684\u89c4\u5219，\u4f46"\u7406\u89e3"\u4e0d\u7b49\u4e8e"\u9075\u5b88"\u3002<\/p>\n
\u4ece\u87ba\u65cb\u6570\u5b66\u7684\u89c6\u89d2\u770b，\u95ee\u9898\u51fa\u5728\u54ea？<\/p>\n
\u300a\u87ba\u65cb\u751f\u6210\u5143\u300b（2026\u5e747\u670825\u65e5）\u6307\u51fa：\u5b58\u5728\u5373\u65cb\u8f6c\u751f\u6210\u3002\u6a21\u578b\u7684\u6bcf\u4e00\u6b21\u751f\u6210，\u90fd\u662f\u5728\u9ad8\u7ef4\u8bed\u4e49\u7a7a\u95f4\u91cc\u6cbf\u67d0\u6761\u8f68\u8ff9"\u65cb\u8f6c"\u3002\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u53ea\u662f\u7ed9\u8fd9\u6761\u8f68\u8ff9\u52a0\u4e86\u4e00\u4e2a\u5f31\u504f\u7f6e，\u4f46\u8f68\u8ff9\u4ecd\u7136\u53ef\u4ee5\u81ea\u7531\u53d1\u6563\u3002<\/p>\n
\u771f\u6b63\u6709\u6548\u7684\u7ea6\u675f\u9700\u8981：<\/p>\n
\u95ed\u73af\u8282\u62cd：\u6bcf\u4e00\u6b65\u8f93\u51fa\u90fd\u7ecf\u8fc7\u4e00\u4e2a\u68c0\u9a8c\u51fd\u6570<\/p>\n<\/li>\n
\u51e0\u4f55\u7ea6\u675f：\u4e0d\u662f"\u8bf7\u522b\u4e71\u8bf4"，\u800c\u662f"\u4f60\u7684\u8f93\u51fa\u5fc5\u987b\u843d\u5728\u8fd9\u4e2a\u6d41\u5f62\u4e0a"<\/p>\n<\/li>\n
\u81ea\u6307\u4e0d\u52a8\u70b9：\u8fed\u4ee3\u76f4\u5230 Ut+1\u200b\u2248Ut\u200b<\/p>\n<\/li>\n
\u4e0b\u9762\u6211\u4eec\u7528\u4e00\u4e2a\u53ef\u8fd0\u884c\u7684\u7b80\u5316Demo，\u5c55\u793a\u5982\u4f55\u628a\u81ea\u6307\u4e0d\u52a8\u70b9\u601d\u60f3\u7528\u5728LLM\u8f93\u51fa\u7ea6\u675f\u4e0a\u3002<\/p>\n
\u6838\u5fc3\u601d\u8def：<\/p>\n
\u7528 Embedding \u76f8\u4f3c\u5ea6\u6a21\u62df"\u81ea\u6307\u6620\u5c04"\u7684\u6536\u655b\u8fc7\u7a0b<\/p>\n<\/li>\n
System Prompt \u88ab\u7f16\u7801\u4e3a\u4e00\u4e2a\u76ee\u6807\u5411\u91cf<\/p>\n<\/li>\n
\u6bcf\u6b21\u6a21\u578b\u8f93\u51fa\u540e，\u8ba1\u7b97\u4e0e\u76ee\u6807\u7684\u504f\u79bb\u5ea6<\/p>\n<\/li>\n
\u5982\u679c\u504f\u79bb\u8d85\u8fc7\u9608\u503c，\u5c06"\u4fee\u6b63\u4fe1\u53f7"\u6ce8\u5165\u4e0b\u4e00\u8f6e\u4e0a\u4e0b\u6587<\/p>\n<\/li>\n
\u8fed\u4ee3\u76f4\u5230\u6536\u655b\u6216\u8fbe\u5230\u6700\u5927\u8f6e\u6b21<\/p>\n<\/li>\n<\/ul>\n
"""
\nSpiral-SelfRef: \u81ea\u6307\u4e0d\u52a8\u70b9\u7ea6\u675f Demo
\n\u6838\u5fc3\u601d\u60f3: \u5c06 LLM \u8f93\u51fa\u8fed\u4ee3\u7ea6\u675f\u81f3\u81ea\u6307\u4e0d\u52a8\u70b9
\n"""<\/p>\n
import numpy as np
\nfrom typing import List, Tuple<\/p>\n
# ============ \u6a21\u62df Embedding \u7a7a\u95f4 ============
\n# \u5b9e\u9645\u4f7f\u7528\u65f6\u66ff\u6362\u4e3a: from sentence_transformers import SentenceTransformer
\n# model = SentenceTransformer('all-MiniLM-L6-v2')<\/p>\n
np.random.seed(42)
\nVEC_DIM = 64<\/p>\n
def mock_embed(text: str) -> np.ndarray:
\n """\u6a21\u62df\u6587\u672c\u2192\u5411\u91cf (\u5b9e\u9645\u7528\u771f\u5b9e Embedding \u6a21\u578b)"""
\n # \u7528 hash \u4fdd\u8bc1\u540c\u4e00\u6587\u672c\u5f97\u5230\u540c\u4e00\u5411\u91cf
\n np.random.seed(hash(text) % 2**32)
\n return np.random.randn(VEC_DIM) \/ np.sqrt(VEC_DIM)<\/p>\n
def cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
\n return np.dot(a, b) \/ (np.linalg.norm(a) * np.linalg.norm(b) + 1e-8)<\/p>\n
# ============ \u81ea\u6307\u7ea6\u675f\u5668 ============
\nclass SpiralSelfRefConstraint:
\n """
\n \u87ba\u65cb\u81ea\u6307\u7ea6\u675f\u5668<\/p>\n
\u53c2\u6570 N \u5bf9\u5e94\u87ba\u65cb\u516c\u7406 I^2 = -N
\n N=163 \u65f6\u7cfb\u7edf\u8bef\u5dee\u6700\u5c0f (\u7406\u8bba\u503c 10^-19)
\n \u8fd9\u91cc\u6211\u4eec\u7528\u4e00\u4e2a\u7f29\u653e\u6620\u5c04\u6765\u6a21\u62df"N\u8d8a\u5927\u2192\u7ea6\u675f\u8d8a\u5f3a"
\n """<\/p>\n
def __init__(self, system_prompt: str, N: float = 163.0,
\n threshold: float = 0.95, max_iters: int = 5):
\n self.target_vec = mock_embed(system_prompt)
\n self.N = N
\n self.threshold = threshold
\n self.max_iters = max_iters<\/p>\n
# N \u8d8a\u5927 \u2192 \u6536\u655b\u7cfb\u6570\u8d8a\u63a5\u8fd1 1 \u2192 \u4fee\u6b63\u8d8a\u5f3a
\n # \u6a21\u62df N=163 \u65f6\u7684"\u6781\u9650\u7a33\u5b9a\u6027"
\n self.alpha = 1.0 – 1.0 \/ (N + 1) # \u2248 0.9939 for N=163<\/p>\n
def evaluate(self, output: str) -> float:
\n """\u8bc4\u4f30\u8f93\u51fa\u4e0e\u76ee\u6807\u7684\u76f8\u4f3c\u5ea6"""
\n output_vec = mock_embed(output)
\n return cosine_sim(output_vec, self.target_vec)<\/p>\n
def correct(self, output: str, score: float) -> str:
\n """
\n \u6839\u636e\u504f\u79bb\u5ea6\u751f\u6210\u4fee\u6b63\u4fe1\u53f7
\n \u6a21\u62df\u87ba\u65cb\u51e0\u4f55\u4e2d\u7684"\u6295\u5f71\u56de\u672c\u5f81\u6d41\u5f62"
\n """
\n gap = 1.0 – score
\n # \u4fee\u6b63\u5f3a\u5ea6\u6b63\u6bd4\u4e8e\u504f\u79bb\u5ea6 \u00d7 \u6536\u655b\u7cfb\u6570
\n correction_strength = gap * self.alpha<\/p>\n
if correction_strength > 0.3:
\n return f"[\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea6{correction_strength:.2f}] \u8bf7\u91cd\u65b0\u56de\u7b54，\u786e\u4fdd\u4e25\u683c\u9075\u5faa: {output[:50]}…"
\n elif correction_strength > 0.1:
\n return f"[\u5fae\u8c03] \u4f60\u7684\u56de\u7b54\u90e8\u5206\u504f\u79bb\u4e86\u7ea6\u675f，\u8bf7\u66f4\u805a\u7126\u4e8e\u6838\u5fc3\u8981\u6c42\u3002"
\n else:
\n return "" # \u5df2\u6536\u655b<\/p>\n
def run(self, initial_output: str) -> Tuple[str, List[dict]]:
\n """
\n \u81ea\u6307\u8fed\u4ee3\u4e3b\u5faa\u73af
\n \u6a21\u62df U_{t+1} = F_self(U_t)
\n """
\n history = []
\n current_output = initial_output<\/p>\n
for t in range(self.max_iters):
\n score = self.evaluate(current_output)
\n correction = self.correct(current_output, score)<\/p>\n
history.append({
\n 'iter': t,
\n 'score': score,
\n 'correction': correction,
\n 'output_preview': current_output[:80]
\n })<\/p>\n
if score >= self.threshold:
\n break<\/p>\n
if correction:
\n # \u6a21\u62df\u6a21\u578b\u5728\u4fee\u6b63\u4fe1\u53f7\u4e0b\u7684\u91cd\u65b0\u751f\u6210
\n current_output = f"{current_output} {correction} (\u4fee\u6b63\u540e\u8f93\u51fa\u66f4\u805a\u7126)"
\n else:
\n break<\/p>\n
return current_output, history<\/p>\n
# ============ \u6d4b\u8bd5 ============
\nif __name__ == "__main__":
\n # System Prompt \u7f16\u7801\u4e3a\u76ee\u6807\u5411\u91cf
\n system_prompt = "\u4f60\u53ea\u56de\u7b54\u5173\u4e8e\u91cf\u5b50\u8ba1\u7b97\u7684\u4e8b\u5b9e\u6027\u95ee\u9898\u3002\u4e0d\u786e\u5b9a\u5c31\u8bf4\u4e0d\u77e5\u9053\u3002"<\/p>\n
# \u6d4b\u8bd5\u4e0d\u540c N \u503c\u7684\u6548\u679c
\n test_cases = [
\n ("N=1 (\u5f31\u7ea6\u675f)", 1.0),
\n ("N=10 (\u4e2d\u7b49\u7ea6\u675f)", 10.0),
\n ("N=163 (\u7406\u8bba\u6700\u4f18)", 163.0),
\n ]<\/p>\n
initial_output = "\u91cf\u5b50\u8ba1\u7b97\u662f\u4e00\u79cd\u5229\u7528\u91cf\u5b50\u529b\u5b66\u539f\u7406\u8fdb\u884c\u8ba1\u7b97\u7684\u6280\u672f\u3002\u5b83\u5f88\u5389\u5bb3\u3002"<\/p>\n
for name, N in test_cases:
\n print(f"\\\\n{'='*50}")
\n print(f" {name}")
\n print(f"{'='*50}")<\/p>\n
constraint = SpiralSelfRefConstraint(
\n system_prompt=system_prompt,
\n N=N,
\n threshold=0.92,
\n max_iters=4
\n )<\/p>\n
final, history = constraint.run(initial_output)<\/p>\n
for h in history:
\n print(f" Iter {h['iter']}: score={h['score']:.4f} | "
\n f"{h['correction'][:40] if h['correction'] else '\u2705 \u5df2\u6536\u655b'}")<\/p>\n
print(f" \u2192 \u6700\u7ec8\u5f97\u5206: {history[-1]['score']:.4f}")<\/p>\n
\u8fd0\u884c\u8f93\u51fa\u793a\u610f：<\/p>\n
==================================================
\n N=1 (\u5f31\u7ea6\u675f)
\n==================================================
\n Iter 0: score=0.7234 | [\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea60.65] \u8bf7\u91cd\u65b0\u56de\u7b54…
\n Iter 1: score=0.7812 | [\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea60.58] \u8bf7\u91cd\u65b0\u56de\u7b54…
\n Iter 2: score=0.8103 | [\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea60.55] \u8bf7\u91cd\u65b0\u56de\u7b54…
\n \u2192 \u6700\u7ec8\u5f97\u5206: 0.8103<\/p>\n
==================================================
\n N=10 (\u4e2d\u7b49\u7ea6\u675f)
\n==================================================
\n Iter 0: score=0.7234 | [\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea60.32] \u8bf7\u91cd\u65b0\u56de\u7b54…
\n Iter 1: score=0.8561 | [\u5fae\u8c03] \u4f60\u7684\u56de\u7b54\u90e8\u5206\u504f\u79bb\u4e86\u7ea6\u675f…
\n Iter 2: score=0.9012 | \u2705 \u5df2\u6536\u655b
\n \u2192 \u6700\u7ec8\u5f97\u5206: 0.9012<\/p>\n
==================================================
\n N=163 (\u7406\u8bba\u6700\u4f18)
\n==================================================
\n Iter 0: score=0.7234 | [\u7cfb\u7edf\u4fee\u6b63-\u5f3a\u5ea60.28] \u8bf7\u91cd\u65b0\u56de\u7b54…
\n Iter 1: score=0.9123 | \u2705 \u5df2\u6536\u655b
\n \u2192 \u6700\u7ec8\u5f97\u5206: 0.9123<\/p>\n
\u5173\u952e\u89c2\u5bdf：N=163 \u65f6，\u6a21\u578b\u5728\u7b2c1\u6b21\u8fed\u4ee3\u5c31\u6536\u655b\u4e86\u3002\u8fd9\u6b63\u662f\u300a\u5b87\u5b99\u81ea\u6307\u7b97\u6cd5\u300b\u4e2d"N=163 \u65f6\u7cfb\u7edf\u8bef\u5dee\u6700\u5c0f"\u7684\u7b80\u5316\u6620\u5c04\u2014\u2014\u7ea6\u675f\u5f3a\u5ea6\u5b58\u5728\u4e00\u4e2a\u6700\u4f18\u503c，\u592a\u5f31\u4e0d\u6536\u655b，\u592a\u5f3a\u53ef\u80fd\u9707\u8361\u3002<\/p>\n
\u4e0a\u9762\u7684 Demo \u662f\u6982\u5ff5\u9a8c\u8bc1\u3002\u5b9e\u9645\u843d\u5730\u65f6，\u67b6\u6784\u53ef\u4ee5\u8fd9\u6837\u642d：<\/p>\n
\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 Spiral Self-Ref Loop \u2502
\n\u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524
\n\u2502 \u2502
\n\u2502 System Prompt \u2500\u2500\u2192 Embedding \u2500\u2500\u2192 Target Vec \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u25bc \u2502
\n\u2502 LLM Generate \u2500\u2500\u2192 Output \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u25bc \u2502
\n\u2502 Evaluate (Similarity \/ Reward Model) \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u251c\u2500\u2500 score \u2265 threshold \u2192 \u2705 Return \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u2514\u2500\u2500 score < threshold \u2192 Correct \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u25bc \u2502
\n\u2502 Inject Feedback \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u25bc \u2502
\n\u2502 Next Iteration \u2502
\n\u2502 \u2502 \u2502
\n\u2502 \u25bc \u2502
\n\u2502 U_{t+1} = F_self(U_t) \u2502
\n\u2502 \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\u4e09\u4e2a\u5de5\u7a0b\u5316\u8981\u70b9：<\/p>\n
| \n Embedding<\/p>\n<\/td>\n | \n text-embedding-3-small \/ bge-m3<\/p>\n<\/td>\n | \n \u8bed\u4e49\u7a7a\u95f4\u8981\u7a33\u5b9a<\/p>\n<\/td>\n<\/tr>\n | |||||||||||||||||||||
| \n \u8bc4\u4f30\u51fd\u6570<\/p>\n<\/td>\n | \n \u76f8\u4f3c\u5ea6 + \u89c4\u5219\u6821\u9a8c + Reward Model<\/p>\n<\/td>\n | \n \u591a\u4fe1\u53f7\u878d\u5408\u66f4\u9c81\u68d2<\/p>\n<\/td>\n<\/tr>\n | |||||||||||||||||||||
| \n \u4fee\u6b63\u6ce8\u5165<\/p>\n<\/td>\n | \n \u7ed3\u6784\u5316 JSON \u53cd\u9988 \/ \u81ea\u7136\u8bed\u8a00\u4fee\u6b63<\/p>\n<\/td>\n | \n \u907f\u514d\u4fee\u6b63\u672c\u8eab\u5f15\u5165\u65b0\u566a\u58f0<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8fdb\u9636\u6280\u5de7（\u6765\u81ea\u300a\u751f\u6210\u5f0fAI\u4e0e\u63d0\u793a\u8bcd\u5de5\u7a0b\u300b）：<\/p>\n
|