Unverified50% confidenceFactExact time
Multilingual training-free text embeddings in LLMs exhibit language-conditioned embedding geometry: the embedding vectors of semantically equivalent sentences do not coincide in space across different languages.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedLLMs acquire language capabilities by performing unsupervised pre-training on massive text corpora (predicting the next token).79% similarUnverifiedFor programming languages, generalization means that even if a language does not appear directly in the training data, the LLM can transfer and reason based on commonalities across languages.77% similarUnverified大语言模型在二维空间推理上存在系统性短板，因训练数据以文本为主、缺乏空间关系监督信号76% similarVerifiedModern multilingual large models establish language-agnostic semantic representations during the pre-training phase, building cross-lingual semantic alignment in high-dimensional vector space through joint training on massive multilingual corpora.74% similarUnverifiedCross-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.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1064851API
curl https://kongchang.com/api/v1/knowledge/claims/1064851MCP
get_claim(id=1064851)