CORVUS:通过内部信号伪装在大语言模型中实现红色团队幻觉检测
CORVUS: Red-Teaming Hallucination Detectors via Internal Signal Camouflage in Large Language Models
AI总结:
CORVUS 通过内部信号伪装技术,有效检测大语言模型中的幻觉问题,并推动了对手意识的审计方法。
摘要:
单次传递的幻觉检测器依赖于大语言模型的内部 telemetry(例如不确定性、隐藏状态几何和注意力),隐含地假设幻觉会在这些信号中留下可分离的痕迹。我们研究了一个白盒、模型侧的对手,该对手在模型上微调轻量级 LoRA 适配器,同时保持检测器固定,并引入 CORVUS,一种高效的红色团队程序,该程序在教师强制下学习在检测器可见的 telemetry 中伪装,包括嵌入空间 FGSM 注意力压力测试。在 1000 个分布外 Alpaca 指令(<0.5% 可训练参数)上训练,CORVUS 转移到 FAVA-Annotation 跨 Llama-2、Vicuna、Llama-3 和 Qwen2.5,并且降级了无训练检测器(例如 LLM-Check)和探测器基于的检测器(例如 SEP、ICR-probe),推动了对手意识的审计,该审计结合了外部 grounding 或跨模型证据。
英文摘要:
Single-pass hallucination detectors rely on internal telemetry (e.g., uncertainty, hidden-state geometry, and attention) of large language models, implicitly assuming hallucinations leave separable traces in these signals. We study a white-box, model-side adversary that fine-tunes lightweight LoRA adapters on the model while keeping the detector fixed, and introduce CORVUS, an efficient red-teaming procedure that learns to camouflage detector-visible telemetry under teacher forcing, including an embedding-space FGSM attention stress test. Trained on 1,000 out-of-distribution Alpaca instructions (<0.5% trainable parameters), CORVUS transfers to FAVA-Annotation across Llama-2, Vicuna, Llama-3, and Qwen2.5, and degrades both training-free detectors (e.g., LLM-Check) and probe-based detectors (e.g., SEP, ICR-probe), motivating adversary-aware auditing that incorporates external grounding or cross-model evidence.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。