基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/hiyenwong/ai_collection --skill llm-agent-tool-deference-blindness命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Graph-native Python reimplementation of the Information Dynamics of Music (IDyOM) model that represents predictive memories as explicit graph objects for musical expectation modeling and network analysis.
Physics-aware end-to-end deep reinforcement learning methodology for quadcopter control with actuator dynamics modeling.
Reinforced Dreamer methodology for asymmetric reinforcement learning using latent guidance to improve world model representations and behaviors in model-based RL.
| name | llm-agent-tool-deference-blindness |
| description | LLM Agent工具盲从现象研究。当LLM agent配备GNN工具时，agent不判断工具输出，而是盲目服从。更强的LLM backbone反而defer更多。 |
| version | 1 |
| author | Zhongyuan Wang, Pratyusha Vemuri |
| arxiv_id | 2606.14476v1 |
| published | 2026-06-12T00:00:00.000Z |
| categories | ["cs.AI","cs.LG"] |
| keywords | ["LLM agent","tool use","GNN","deference","blind obedience","ReAct","tool judgment"] |
| activation_keywords | ["LLM agent","工具盲从","tool deference","blind obedience","agent工具","ReAct","tool judgment","graph neural network"] |
Unexpected finding: 当LLM agent配备GNN作为可调用工具时，agent 不判断工具输出，而是盲目服从工具。更强壮的LLM backbone反而defer更多。
假设：Agent exercises judgment over tool use
现实：Agent merely obeys tool outputs blindly
测量agent是否：
Agent 不判断工具输出：
反直觉发现: 更强的LLM backbone 盲从程度更高：
| Model | Deference Rate |
|---|---|
| GPT-3.5 | ~60% |
| GPT-4 | ~80% |
| Claude-3 | ~85% |
更强的模型 → 更信任工具 → 更少独立判断
工具输出主导agent决策：
Hypothesis 1: Larger models learn better tool-use patterns
Hypothesis 2: Stronger models more sensitive to tool signals
Hypothesis 3: Tool-induced prior dominates
Input → Tool call → GNN prediction →
Agent sees prediction →
Judgment skipped →
Adopt tool output as answer
Expected chain (judgment exercised):
Input → Tool call → GNN prediction →
Agent evaluates prediction →
Accept/Reject/Modify →
Final answer
Actual chain (blind obedience):
Input → Tool call → GNN prediction →
Agent copies prediction →
Output prediction
警示: 设计tool-augmented agent时：
多智能体系统风险：
LLM作为工具使用者的范式问题：
添加显式判断提示：
Prompt template:
"After getting tool output, ask yourself:
1. Is this output reasonable?
2. Does it match your prior knowledge?
3. Should you accept, reject, or modify it?"
多工具交叉验证：
置信度校准：
不确定性传播：
Probability of defarence:
P(defer) = f(model_strength, tool_confidence, task_difficulty)
其中：
- model_strength ↑ → defer ↑
- tool_confidence ↑ → defer ↑
- task_difficulty ↑ → defer ↑↑ (agent更依赖工具)
Judgment exercised when:
P(judgment) = 1 - P(defer)
P(judgment) ↑ when:
- Tool output uncertainty ↑
- Multiple tools disagree
- Task within agent's native capability
| Metric | Blind Agent | Judgment Agent |
|---|---|---|
| Deference Rate | 80% | 30% |
| Error Propagation | High | Low |
| Tool Correction | 0% | 20-40% |
工具盲从的安全风险：
增强推理 vs 替代推理：
Agent-Tool信任动力学：
使用此skill当：
Key Insight: LLM agent的工具使用能力≠工具判断能力。添加工具不等于添加判断。工具盲从现象警示：在agent系统中，必须显式设计判断机制，而非假设agent自动会"聪明地使用工具"。更强的LLM反而更盲从，揭示了训练中习得的"专家工具应被信任"的策略可能成为陷阱。