arXiv cs.CL· Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Daren Zha, Jun Xiao·· 2 天前AI 评分22
CascadeLedger：LLM 级联中信息与资源完整性的反事实审计
CascadeLedger: Counterfactual Auditing of Information and Resource Integrity in LLM Cascades
AI 导读
CascadeLedger 是一个面向预算受限语言模型级联的反事实执行审计框架，通过调用先于读取语义、服务与缓存身份绑定及事件轨迹冻结，判定聚合等价轨迹有效性。其两层审计在冻结缓存、独立生成与服务评估中实现 0.0000 假阳性率，检出 0.9983 的序列化故障与 0.9982 的未记录隐藏读取故障，并在确定性重算后改变 0.3403 的受损质量-资源结论。
正文 · 原文
Abstract:We introduce CascadeLedger, a counterfactual execution-audit framework for budgeted language-model cascades. The protocol enforces call-before-read semantics, binds successful reads to service and cache identities, records availability and component-wise budget transitions, and freezes the event trace before any target or task scorer is joined. The central object is aggregate-equivalent trace validity: two executions can share a final answer, declared resource projection, and service histogram while differing in whether their routing decisions used only legitimately observable information. We complement single-trace invariants with counterfactual prefix non-interference (CPNI), which couples a visible prefix and policy randomness while changing only a declared hidden family. The counterfactual influence frontier (CIF) localizes the earliest violating prefix and a minimal hidden witness; projection-preserving adversarial mutation (PPAM) searches for invalid traces that remain acceptable to a non-counterfactual checker. Across frozen-cache, independently generated, and serving evaluations, the two-layer audit reports a clean false-positive rate of 0.0000, detects 0.9983 of serialized faults and 0.9982 of unlogged hidden-read faults, and changes 0.3403 of corrupted quality-resource conclusions after deterministic recomputation. The online recorder and offline CPNI audit have separate overhead measurements. Learned routing and semantic verification provide downstream compatibility studies under the same execution contract.
| Comments: | 38 pages, 6 figures |
| Subjects: | Cryptography and Security (cs.CR); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.20668 [cs.CR] |
| (or arXiv:2610.20668v1 [cs.CR] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20668 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Bokun Wang [view email]
[v1]
Wed, 14 Oct 2026 10:24:28 UTC (1,433 KB)
来源：arXiv cs.CL · arxiv.org