Unverified50% confidenceFactExact time
The study proposes a "paired reversal" method that separates the effects of component language identity from text quantity by controlling behavior, English framing, and the number of translated spans.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe study recommends that evaluations record component language and the amount of translated text, compare matched behaviors at equal numbers of translated spans, and interpret refusal and unsafe compliance separately.77% similarUnverifiedThe study provides behavioral-level evidence that "bilingual models capture some aspects of cross-linguistic influence."74% similarUnverified该论文的核心思想是通过改变提示词的表述方式让大语言模型产生更加多样化的输出73% similarUnverifiedThe source, quality, and quantity of monolingual data affect the effectiveness of back-translation in generating synthetic corpora.73% similarUnverifiedThe study finds that the translation process actually operates through sparse, structured circuits on a shared semantic subspace, rather than relying on English-centric pivot representations.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1114724API
curl https://kongchang.com/api/v1/knowledge/claims/1114724MCP
get_claim(id=1114724)