LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
← Volver a la lectura
NoticiasResumen rápido1 min de lecturaEditorial de DisconfirmAI

New analysis says defaulting J-lens to the final layer injects a language bias

Self-tested on DeepSeek-V3: final-layer targets push about half of mid-depth readouts to Chinese tokens; the penultimate layer cuts that to 3%.

2026-09-25 14:32 GMT+8

A researcher argues that J-lens, an interpretability tool that translates a model's internal states into words, is polluted by its default choice of the final layer as target.

The author, Kaley Brauer, checked 78 publicly released J-lenses and found 80% target the final layer. In her own tests on DeepSeek-V3, a final-layer target made about half of mid-depth readouts come out as mostly Chinese tokens; switching to the penultimate layer dropped that to 3%.

The cause, she says, is that DeepSeek-V3's last block strongly suppresses Chinese tokens on English text, and that amplified direction comes to dominate the lens. She recommends checking whether the final block injects a dominant direction when fitting or using a J-lens. The result is self-tested, covers only DeepSeek-V3, and awaits independent reproduction.

Fuentes

  • lesswrong.com↗
Continuar investigando →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión