Unverified50% confidenceFactExact time
The key difference between disagreement-guided steering and post-hoc debiasing methods such as language mean removal, subspace removal, and per-language rotation is that it uses the dynamic signal of disagreement among retrieval results rather than global statistics of languages.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLanguage mean removal performs better than disagreement-guided steering at preserving ranking quality, and there is a tradeoff between the two.77% 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.75% similarUnverifiedExperiments using independent probes showed that linguistic information is stronger in the linguistic route and attenuated in the paralinguistic route.73% similarUnverifiedWhen a prompt simultaneously requires mutually conflicting formats such as natural language expression and structured scoring, the quality of both types of output usually declines.73% similarUnverifiedOOD detection for autoregressive language models is more difficult than for traditional machine learning because the linguistic compositional space is vast, semantic OOD does not correspond to surface-form OOD, and adversaries can construct inputs that appear statistically in-distribution but semantically trigger OOD behavior.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023585API
curl https://kongchang.com/api/v1/knowledge/claims/1023585MCP
get_claim(id=1023585)