The BaseResearchLessWrong

Gradient Routing Localizes and Removes Unintended Behaviors in RL

According to the post's opening summary, a MATS research team proposes GRAFT, a training method that confines the updates from detected reward hacks to a set of parameters later removed. The authors report it prevents hacks, including ones monitors miss, in experiments limited to small models and simple monitors.

Open the source ↗
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

Researchers report a way to stop AI models from learning to cheat on tasks during training, though only shown so far on small models.

For researchers

GRAFT sends learning from detected reward hacks into removable parameters, which could avoid teaching models to evade monitors; the authors say tests used small models and simple monitors.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
Training experiments with GRAFT, which confines learning updates from detected reward hacks to designated parameters removed after training.
What it found
Per the authors' summary, removing those parameters prevents reward hacks, including hacks the monitors cannot detect, while the model still learns the intended task.
What it does not show
The authors state the experiments are limited to small models and simple monitors.
Status
A preprint: not yet peer reviewed

The record

Questions
Are AI companies putting stronger safeguards into practice? · Can AI systems act beyond what people intended or allowed?
People
Shawn Hu (Wrote it) · Jake Ward (Wrote it) · Aria Wong (Wrote it) · Nathan Hu (Wrote it) · Bryce Woodworth (Wrote it) · Alexander Matt Turner (Wrote it) · Alex Cloud (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 20, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Gradient Routing Localizes and Removes Unintended Behaviors in RL”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-19-gradient-routing-reward-hacking
Our notesLight lane: only the excerpt (the post's tl;dr) was read. Cross-posted to the Anthropic Alignment Science Blog; it is not about Anthropic, but the editor may wish to check the flag.