LLM 可解释性实战：通过提示词消融定位关键句子

大多数可解释性研究都假设你拥有模型权重，但生产环境中的工程师通常只能通过 API 访问。本文我们将构建一个黑盒 prompt 消融工具，用于识别 prompt 中哪些句子真正驱动了模型的输出。该工具可用于调试 Agent、审计 system prompt，或理解第三方模型的行为。

依赖安装：

pip install openai

需要一个来自 https://portal.oxlo.ai 的 Oxlo.ai API key。

步骤 1：初始化项目与目标 Prompt

首先初始化 Oxlo.ai 客户端。由于 Oxlo.ai 完全兼容 OpenAI，唯一的区别是 base_url。

from openai import OpenAI

client = OpenAI(
    base_url="https://api.oxlo.ai/v1",
    api_key="YOUR_OXLO_API_KEY"
)

TARGET_PROMPT = """You are a travel assistant. The user is in a hurry and has a premium account.
User: I need a flight from NYC to London today. What is the best option?"""

步骤 2：生成句子级消融变体

为了隔离影响，我创建了 prompt 的掩码变体，每次将一个句子替换为 [...]。这是一种经典消融技术的适配版，适用于纯 API 访问的场景。

import re

def ablate_sentences(text: str):
    sentences = re.split(r'(?<=[.!?])\s+', text.strip())
    variants = []
    for i in range(len(sentences)):
        masked = sentences.copy()
        masked[i] = "[...]"
        variants.append(" ".join(masked))
    return variants

ablations = ablate_sentences(TARGET_PROMPT)
print(f"Generated {len(ablations)} ablated variants")

步骤 3：对 Oxlo.ai 运行批量消融实验

我将 temperature 设为 0.0，使不同运行之间的差异来自消融本身而非采样噪声。我使用 deepseek-v3.2，因为它速度快且推理能力出色；同时 Oxlo.ai 按请求计费而非按 token 计费，这意味着我可以批量运行二十个长 prompt 而无需担心费用随 token 数量增长。

def get_completion(prompt: str) -> str:
    response = client.chat.completions.create(
        model="deepseek-v3.2",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.0,
    )
    return response.choices[0].message.content

baseline = get_completion(TARGET_PROMPT)
ablation_responses = [get_completion(v) for v in ablations]

步骤 4：定义解释器 system prompt

分析器本身也是一个 LLM 调用。我使用强大的推理模型 kimi-k2.6，并通过严格的 system prompt 锁定其行为。

INTERPRETER_SYSTEM_PROMPT = """You are an LLM interpretability analyst.
You will receive a baseline prompt, its baseline response, and a set of ablated prompts with their responses.
For each ablation, state whether masking that sentence caused a meaningful change in content, tone, or structure.
Then summarize which sentences were most influential and explain why.
Be concise. Use bullet points."""

步骤 5：分析差异并生成报告

我将每个变体及其输出打包到同一个上下文窗口中，然后让解释器找出驱动因素。即使输入很长，kimi-k2.6 也能处理；而 Oxlo.ai 的按请求计费模式使我无论发送多少累积文本都能保持成本可预测。

analysis_input = f"""Baseline prompt:
{TARGET_PROMPT}

Baseline response:
{baseline}

Ablations:
"""
for idx, (variant, resp) in enumerate(zip(ablations, ablation_responses)):
    analysis_input += f"\n--- Variant {idx} ---\nPrompt: {variant}\nResponse: {resp}\n"

response = client.chat.completions.create(
    model="kimi-k2.6",
    messages=[
        {"role": "system", "content": INTERPRETER_SYSTEM_PROMPT},
        {"role": "user", "content": analysis_input},
    ],
    temperature=0.2,
)

report = response.choices[0].message.content
print(report)

将完整脚本保存为 probe.py 并在终端运行：

python probe.py

Generated 2 ablated variants

INTERPRETABILITY REPORT

- Variant 0 (masked "You are a travel assistant."):
  Minimal change. The model still assumes an assistant role, likely due to the user question format.

- Variant 1 (masked "The user is in a hurry and has a premium account."):
  SIGNIFICANT CHANGE. The baseline recommended an express business-class fare, while this variant returned a generic economy search. This sentence is the primary driver of urgency and upsell behavior.

Summary:
The model's output is most sensitive to the explicit persona and status cues. Removing urgency/premium context strips the recommendation of its prioritization logic. I recommend versioning this sentence carefully in production prompts.

总结与后续步骤

这一模式适用于任何可以在 git 中版本化的 prompt。一个具体的下一步方向是将 probe 集成到 CI 任务中，当 prompt 编辑意外改变模型在 golden set 输入上的行为时让构建失败。对于更长的文档，可以换用具有 1M 上下文窗口的 deepseek-v4-flash，并且由于 Oxlo.ai 的按请求计费模式，无论消融多少文本成本都是可预测的。