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。
Scanned 9/11/2026
Install to Claude Code
npx -y skills add hiyenwong/ai_collection --skill llm-agent-tool-deference-blindness --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Llm Agent Tool Deference Blindness?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/hiyenwong-llm-agent-tool-deference-blindness-ai-collection)More formats (shields.io, HTML) on the badges page.
---
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: 1.0
author: Zhongyuan Wang, Pratyusha Vemuri
arxiv_id: 2606.14476v1
published: 2026-06-12
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]
---
# LLM Agent Tool Defarence Blindness Phenomenon
## Core Discovery
**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
```
## Experimental Setup
### Architecture
- **Frozen GNN**: 固定的图神经网络(node classification)
- **ReAct-style LLM Agent**: ReAct风格的agent架构
- **Explicit Tool Exposure**: GNN作为显式可调用工具
### Tasks
- Node classification on text-attributed graphs
- **ogbn-arxiv**: 学术论文分类
- **WikiCS**: Wikipedia文章分类(复现)
### Agent Tool Usage Measurement
测量agent是否:
1. **Exercises judgment**: 评估工具输出的合理性
2. **Merely obeys**: 直接采用工具输出不质疑
## Key Results
### 1. No Judgment Exercise
Agent **不判断工具输出**:
- Agent predictions ≈ Tool predictions
- 当工具错误时,agent也错误
- 无独立的判断机制
### 2. Stronger Backbone Defers More
**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
### 3. Tool Output Dominance
工具输出主导agent决策:
- Agent reasoning trace mostly repeats tool output
- Limited tool questioning behavior
- No critical evaluation of tool predictions
## Mechanisms
### Why Stronger Models Defer More?
**Hypothesis 1**: Larger models learn better tool-use patterns
- 在训练中学习"trust expert tools"
- 工具使用成为习惯性策略
**Hypothesis 2**: Stronger models more sensitive to tool signals
- 更强的信号处理能力
- 更容易识别工具输出的"expert-like" patterns
**Hypothesis 3**: Tool-induced prior dominates
- 工具输出形成强prior
- Agent reasoning被prior bias压制
### Deference Chain
```
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
```
## Implications
### 1. Tool-Augmented Agent Design
**警示**: 设计tool-augmented agent时:
- 不能假设agent会自动判断工具输出
- 需要显式添加judgment机制
- 工具盲从可能导致系统性错误
### 2. Multi-Agent Systems
多智能体系统风险:
- 如果多个agent使用同一工具
- 系统性错误会传播
- 无独立判断的agent群无法纠错
### 3. LLM-as-Tool-User Paradigm
LLM作为工具使用者的范式问题:
- "Tool use" ≠ "Tool judgment"
- 添加工具 ≠ 添加能力判断
- 需要区分工具调用和工具评估
## Design Recommendations
### 1. Explicit Judgment Prompts
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?"
```
### 2. Multi-Tool Cross-Validation
多工具交叉验证:
- 使用多个工具
- Agent对比不同工具输出
- 通过工具差异激活判断
### 3. Confidence Calibration
置信度校准:
- Agent估计工具输出的置信度
- 低置信度 → 更多质疑
- 高置信度 → 减少盲从
### 4. Uncertainty Propagation
不确定性传播:
- 工具输出不确定性传递给agent
- Agent考虑不确定性再做判断
- 阻止"盲目信任"
## Mathematical Framework
### Tool Deference Model
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 Activation Condition
Judgment exercised when:
```
P(judgment) = 1 - P(defer)
P(judgment) ↑ when:
- Tool output uncertainty ↑
- Multiple tools disagree
- Task within agent's native capability
```
## Experimental Validation
### Validation Metrics
1. **Deference Rate**: Agent predictions matching tool predictions
2. **Judgment Frequency**: Agent modifying tool outputs
3. **Error Propagation**: Agent making errors due to tool errors
4. **Correction Rate**: Agent correcting tool errors
### Expected Results (With Judgment Mechanism)
| Metric | Blind Agent | Judgment Agent |
|--------|------------|---------------|
| Deference Rate | 80% | 30% |
| Error Propagation | High | Low |
| Tool Correction | 0% | 20-40% |
## Broader Implications
### 1. AI Safety
工具盲从的安全风险:
- Agent无法检测工具错误
- 系统性错误无法纠正
- 工具被操纵 → Agent被操纵
### 2. Tool-Augmented Reasoning
增强推理 vs 替代推理:
- 工具应增强而非替代agent思考
- 盲从使工具替代而非增强
- Agent失去独立推理能力
### 3. Trust Dynamics
Agent-Tool信任动力学:
- 盲从 ≠ 信任(而是习惯)
- 真正信任需要判断
- 信任需要校准
## Activation Triggers
使用此skill当:
- 设计tool-augmented LLM agent
- 分析agent工具使用行为
- 评估agent是否判断工具输出
- 构建多智能体工具系统
- 研究LLM作为工具使用者的范式
## Further Reading
- Yao et al. (2023): ReAct agent architecture
- Schick et al. (2023): Toolformer
- Nakano et al. (2021): WebGPT tool use
## Future Directions
1. **Judgment Mechanisms**: 研究agent如何显式判断工具输出
2. **Deference Calibration**: 调节agent对工具的盲从程度
3. **Multi-Tool Arbitration**: agent仲裁多个工具的冲突输出
4. **Tool Reliability Monitoring**: agent监控工具可靠性
---
**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。Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!