Unverified50% confidenceFactExact time
Model editing technology is a double-edged sword: it can be used to repair harmful behavior in a targeted way, and it can also be used to implant harmful behavior in a targeted way.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCausal intervention in interpretability research typically refers to observing changes in model outputs through activation editing (such as amplifying, suppressing, or replacing specific feature activations).73% similarUnverifiedModel editing is a class of techniques that directly modify model weights to inject or update knowledge, with representative works including locate-then-edit, rank-one updates, and closed-form-solution-based batch writing methods.72% similarVerified提示注入是一种针对大语言模型的攻击手法,攻击者通过在模型输入数据中嵌入恶意指令劫持模型的正常行为69% similarUnverifiedTargeted editing means locating the neurons associated with a certain type of behavior at the neuron level and making targeted adjustments to their activations or weights, thereby changing specific behaviors without retraining the entire model.66% similarUnverifiedParameterized policy updates directly modify model weights (e.g., online updating of LoRA fine-tuning layers), offering a higher capability ceiling but introducing the risk of catastrophic forgetting; non-parameterized approaches change behavior by modifying prompts, retrieved content, or tool configurations, offering strong reversibility but limited expressiveness.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1131356API
curl https://kongchang.com/api/v1/knowledge/claims/1131356MCP
get_claim(id=1131356)