Unverified50% confidenceOpinionExact time
For users who only use a single language, multilingual weights are almost dormant during inference but still occupy VRAM and storage space.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCross-lingual jailbreaking exploits the asymmetry between a model's concept space and its language space, as the learning of safety-refusal behavior is highly language-specific.69% similarUnverified在急救、编程语言核心语法、母语表达等场景中，即时的无需查询的掌握仍然至关重要67% similarUnverifiedDue to the stochastic nature of large language model outputs, it is difficult for ordinary users to determine the underlying model version from a single conversation.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/909848API
curl https://kongchang.com/api/v1/knowledge/claims/909848MCP
get_claim(id=909848)