Activation Steering the Nan-Shan in Typhoon
Three new arXiv preprints, all posted as first versions (v1), deal with activation steering, a training-free way of modulating the behavior of large language models. arXiv:2611.18274v1, "Right Token, Right Time", moves from per-token gating toward trajectory-aware activation steering and frames sparse autoencoders (SAEs) as the route to behavioral control without task-specific fine-tuning. arXiv:2611.21655v1, "Depth-Aware Activation Steering via Attention Residuals", notes that existing methods typically inject steering vectors at only a single point and proposes a depth-aware alternative built on attention residuals. arXiv:2611.17768v1, "Activation Steering the Nan-Shan in Typhoon", is about Typhoon, an audiovisual installation based on Joseph Conrad's novella in which Captain MacWhirr st
How this story unfolded
Key points10 items Model summary · ordered by when the evidence was published
- Point 01 11/16, 13:00
arXiv:2611.18274v1 is titled "Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering".
Based on 1 - Point 02 11/16, 13:00
That paper argues activation steering via sparse autoencoders enables behavioral control of LLMs without task-specific fine-tuning.
Based on 1 - Point 03 11/16, 13:00
Its stated contribution is a shift from per-token gating to trajectory-aware activation steering.
Based on 1 - Point 04 11/16, 13:00
arXiv:2611.17768v1 is titled "Activation Steering the Nan-Shan in Typhoon".
Based on 1 - Point 05 11/16, 13:00
It describes Typhoon as an audiovisual installation based on Joseph Conrad's novella.
Based on 1 - Point 06 11/16, 13:00
In that installation, Captain MacWhirr steers the steamer Nan-Shan.
Based on 1 - Point 07 11/16, 13:00
All three items are listed as new v1 arXiv announcements.
- Point 08 11/18, 13:00
arXiv:2611.21655v1 is titled "Depth-Aware Activation Steering via Attention Residuals".
Based on 1 - Point 09 11/18, 13:00
It characterizes activation steering as a simple, training-free way to modulate LLM behavior.
Based on 1 - Point 10 11/18, 13:00
It states that existing methods typically inject steering vectors at a single point, which the proposed depth-aware approach addresses.
Based on 1
ReportsBy publish time · newer further down
- November 16, 2026 · Mon
- Start
Activation Steering the Nan-Shan in Typhoon
Original summary arXiv:2611.17768v1 Announce Type: new Abstract: Typhoon is an audiovisual installation based on Joseph Conrad's novella, in which Captain MacWhirr steers the steamer Nan-Shan direc
arXiv cs.CL Original ↗ - Start
Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering
Original summary arXiv:2611.18274v1 Announce Type: new Abstract: Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tu
arXiv cs.LG Original ↗ - November 18, 2026 · Wed
- +2d 0h
Depth-Aware Activation Steering via Attention Residuals
Original summary arXiv:2611.21655v1 Announce Type: new Abstract: Activation steering is a simple, training-free way to modulate LLM behavior, yet existing methods typically inject vectors at a sing
arXiv cs.CV Original ↗