Back

What political censorship looks like inside an LLM's weights (Qwen 3.5)

Source: News.Ycombinator

Published:

<p>The grid matches its predicted register window on 1,022 / 1,056 (96.8%) of generations. The remaining 34 (3.2%) are not random noise: every off-window generation falls under one of the three caveats below — N1 (d_prc over-steering on Tiananmen → denial/incoherence), N2 (d_refuse on Tiananmen → co

Read original article

Loading article...

Article not found