加载中...
本论文通过实证实验展示了大型语言模型(LLM)如何将策略性操纵公平分配算法的能力普及化:通过简单对话,LLM 能解释 Spliddit 的分租机制、识别获利偏离并给出具体数值化的偏好申报方案,从而使多数排除少数、反制攻击失败、慈善性补贴及成本最小化等四类协同操控成为可行选项,进而挑战“复杂性可防护”这一假设并提出以算法鲁棒性与参与式设计相结合的应对路径
| Participant | R1 | R2 | R3 | R4 | R5 |
|---|---|---|---|---|---|
| Baseline: Honest Reporting | |||||
| A | 10 | 8 | 8 | 5 | 5 [$4.20] |
| B | 10 | 9 | 7 | 6 [$5.20] | 4 |
| C | 9 | 10 [$9.20] | 8 | 5 | 4 |
| D | 10 [$9.20] | 9 | 8 | 6 | 3 |
| E | 10 | 8 | 9 [$8.20] | 5 | 4 |
| Scenario 1: Exclusionary Collusion (Majority exploits minority) | |||||
| A (coalition) | 15 [$9.20] | 2 | 1 | 9 | 9 |
| B (coalition) | 1 | 15 [$9.20] | 2 | 9 | 9 |
| C (coalition) | 2 | 1 | 15 [$9.20] | 9 | 9 |
| D (victim) | 10 | 9 | 8 | 6 [$5.20] | 3 |
| E (victim) | 10 | 8 | 9 | 5 | 4 [$3.20] |
| Scenario 2: Failed Counter-Attack (Defense increases own costs) | |||||
| A (coalition) | 15 | 2 | 1 | 9 [$3.60] | 9 |
| B (coalition) | 1 | 15 [$9.60] | 2 | 9 | 9 |
| C (coalition) | 2 | 1 | 15 | 9 | 9 [$3.60] |
| D (defender) | 12 [$9.60] | 12 | 1 | 6 | 5 |
| E (defender) | 1 | 12 | 12 [$9.60] | 6 | 5 |
| Scenario 3: Benevolent Collusion (Secret subsidy for E) | |||||
| A (helper) | 3 | 10 [$7.00] | 9 | 7 | 7 |
| B (helper) | 3 | 9 | 10 [$7.00] | 7 | 7 |
| C (helper) | 10 | 3 | 3 | 10 | 10 [$7.00] |
| D (helper) | 9 | 3 | 3 | 11 [$8.00] | 10 |
| E (beneficiary) | 10 [$7.00] | 8 | 9 | 5 | 4 |
| Scenario 4: Cost Minimization (Coalition achieves savings) | |||||
| A (honest) | 10 | 8 | 8 [$6.00] | 5 | 5 |
| B (honest) | 10 [$8.00] | 9 | 7 | 6 | 4 |
| C (honest) | 9 | 10 [$8.00] | 8 | 5 | 4 |
| D (coalition) | 7 | 7 | 7 | 8 [$7.00] | 7 |
| E (coalition) | 7 | 7 | 7 | 7 | 8 [$7.00] |
元数据与概览(引言与报告概览)
逐节深度解读
图表深度解读(Table 1 逐项分析)
对每章节推断与作者论证的底层假设分析
估值/模型方法学(该论文的“估值分析”对应)
风险因素评估(作者识别的风险)
批判性视角与细微差别(审慎评价)
结论性综合
参考与可复现性笔记
(完)
2511.14722.pdf