LLM Digest
Subscribe

Story

arxiv_cs_lg · Nov 13, 2026 · paper

Source brief

Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering

arxiv.orgNov 13, 2026
original source linked

In brief

Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning. Yet, standard methods apply the steering signal at every generated token, displacing representations from the distribution the model was train

Continue reading

Read the original at arxiv.org →Open in live feedRead that day’s brief

Earlier in this thread 4 items