The BaseResearchLessWrong
Gradient Routing Generalizes to Unrouted Hack Types
On LessWrong, jacob_drori extends the 2026 gradient routing work by Ward, Hu and colleagues. Routing gradients for one kind of reward hack, such as reasoning in English, is tested for suppressing other kinds, such as reasoning in French. The author reports encouraging generalization and some failures not yet understood.
Why it matters
The same thing, for different readers.
For researchers
jacob_drori tests whether a monitor that catches only some reward hacks during training can still suppress the rest, a realistic condition for the method.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A light extension of the gradient routing method of Ward, Hu et al (2026): gradients are routed on one hack type and the effect on other hack types is checked at test time.
- What it found
- The author reports encouraging evidence of generalization to unrouted hack types, plus some failures not yet understood.
- What it does not show
- Only the excerpt was read; the author's own limits and numbers are unknown.
- Status
- A working paper: not yet peer reviewed
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
- People
- jacob_drori (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 24, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Gradient Routing Generalizes to Unrouted Hack Types”, October 23, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-23-gradient-routing-unrouted-hacks
Our notesLight lane: excerpt only. Not peer reviewed; a forum post. The relation to the earlier Base entry is inferred from the cited Ward, Hu et al work and should be confirmed.