A single safety demonstration appended at inference time mitigates many-shot jailbreak attacks by counteracting implicit malicious fine-tuning on harmful examples.
Safety at one shot: Patching fine-tuned llms with a single instance
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 3roles
background 1polarities
background 1representative citing papers
Patcher repairs backdoored LLMs from a single failure case by localizing triggers via response-conditioned gradient saliency and adaptive clustering then applying constrained fine-tuning to break trigger associations.
The paper introduces a dynamical model that decomposes alignment updates in LLM fine-tuning into rebound and driving forces and predicts a rehearsal priming effect.
citing papers explorer
-
Mitigating Many-shot Jailbreak Attacks with One Single Demonstration
A single safety demonstration appended at inference time mitigates many-shot jailbreak attacks by counteracting implicit malicious fine-tuning on harmful examples.
-
Patcher: Post-Hoc Patching of Backdoored Large Language Models
Patcher repairs backdoored LLMs from a single failure case by localizing triggers via response-conditioned gradient saliency and adaptive clustering then applying constrained fine-tuning to break trigger associations.
-
Alignment Dynamics in LLM Fine-Tuning
The paper introduces a dynamical model that decomposes alignment updates in LLM fine-tuning into rebound and driving forces and predicts a rehearsal priming effect.