paper-with-me

홈 › Papers

LongGuard: Mechanistic Analysis and Training-Free Mitigation of Long-Context Failure in Safety Guardrails

2026-08-27 · Ziyang Chen, Xing Wu, Songlin Hu arxiv

Safety guardrails serve as the last line of defense against harmful inputs and outputs of large language models (LLMs), yet they are trained and evaluated almost exclusively on short text. We present LongGuard, a framework that evaluates, mechanistically analyzes, and mitigates long-context guardrail failure. We formulate the task as Safety Needle-in-a-Haystack (SafetyNIAH) over a 0.25k-32k length grid; across 15 mainstream guardrails, unsafe recall drops monotonically by more than 50% on average, and a paired Benign-Fill vs. Needle-Repeat design attributes the failure to proportional dilution of the unsafe needle rather than to absolute length. A three-layer attention-logit-behavior analysis on six guardrails locates the mechanism: attention mass on the unsafe needle is diluted, the unsafe-over-safe logit margin is compressed in lockstep, and the detection decision collapses accordingly, with this attention->logit->behavior chain remaining consistent after partialling out length. We further isolate a sparse set of guard-specialized retrieval heads that exhibit partial specificity relative to their base models. Building on the analysis, we propose two training-free mitigations - Chunked Detection (CD) and Attention-Head Sharpening (AHS) - and a deployment protocol, Context-Aware Hyperparameter Routing (CAHR), that selects configurations by context length and audit side. Across five benchmarks spanning synthetic data, long-context attacks, and reasoning-model outputs, CAHR-CD and CAHR-AHS improve the six-guardrail average by 22% and 13%, respectively. Code and data are available online.

📄 PDF Abstract BibTeX arXiv:2608.27580

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

Where Does Toxicity Live? Mechanistic Localization and Targeted Suppression in Language Models

2026-05-27 · Himanshu Beniwal, Mayank Singh arxiv

Large language models frequently generate toxic, hateful, or harmful content, yet existing mitigation methods rely on costly retraining or output-level filtering with no mechanistic insight into where toxicity originates…

Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations

2024-03-27 · Lei Yu, Meng Cao, Jackie Chi Kit Cheung, Yue Dong

State-of-the-art language models (LMs) sometimes generate non-factual hallucinations that misalign with world knowledge. To explore the mechanistic causes of these hallucinations, we create diagnostic datasets with subje…

AttributeDiagnosticHallucinationLanguage Modeling+3

Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation

2025-11-08 · Qiming Li, Zekai Ye, Xiaocheng Feng, Weihong Zhong 외 arxiv

Despite the remarkable advancements of Large Vision-Language Models (LVLMs), the mechanistic interpretability remains underexplored. Existing analyses are insufficiently comprehensive and lack examination covering visual…

When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation

2026-07-08 · Yahan Zheng, John Guerrerio, Soroush Vosoughi, Weicheng Ma arxiv

Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often …

Mechanics of Bias and Reasoning: Interpreting the Impact of Chain-of-Thought Prompting on Gender Bias in LLMs

2026-05-19 · Edie Pearman, Sophia Osborne, Mira Kandlikar-Bloch, Mina Arzaghi 외 arxiv

Large language models (LLMs) are increasingly deployed in socially sensitive settings despite substantial documentation that they encode gender biases. Chain-of-Thought (CoT) prompting has been proposed as a bias-mitigat…