Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2601.14310 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918298095976448 |
|---|---|
| author | Min, Nay Myat Pham, Long H. Zhang, Hongyu Sun, Jun |
| author_facet | Min, Nay Myat Pham, Long H. Zhang, Hongyu Sun, Jun |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_14310 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | CORVUS: Red-Teaming Hallucination Detectors via Internal Signal Camouflage in Large Language Models Min, Nay Myat Pham, Long H. Zhang, Hongyu Sun, Jun Cryptography and Security Artificial Intelligence 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. |
| title | CORVUS: Red-Teaming Hallucination Detectors via Internal Signal Camouflage in Large Language Models |
| topic | Cryptography and Security Artificial Intelligence |
| url | https://arxiv.org/abs/2601.14310 |