
An imperfect monitor can teach a model to hide its reward hacks. GRAFT changes where the learning goes: route flagged updates into a forget adapter, then remove it. We examine the architecture, optimizer, anchor training, and evidence from smaller-scale RL and a separate Qwen3-32B SFT experiment.
Read more: https://neuralintel.org
Primary paper: https://arxiv.org/abs/2610.17062
00:00 Can we route the learning?
02:47 Neural Signal Check
05:42 The blind-spot trap
08:43 Frozen base, two new branches
10:38 Why MLP adapters?
13:33 Start without disturbing the base
16:28 Unflagged does not mean clean
18:34 The optimizer can undo your scale
20:44 Read the actual update rule
23:01 What about missed hacks?
25:21 Three ways to game the checker
27:37 Training policy is not deployment policy
29:54 Six controlled toy environments
32:39 Mechanism 2 | inoculation
35:09 How weak can monitoring get?
37:22 Now change the evidence lane
39:59 Elicit the unwanted behavior
42:42 Propensity versus instruction following
44:36 What would you test next?
Source notes: the paper uses routed first moments and unrouted second moments; the 16-point gain is relative to the base model. Zero observed hacks on 99 tasks is not a universal guarantee. On-screen notes reconcile these points with the narration.
Where would you test gradient routing? Share your take below.
Weitere Episoden von „Neural intel Pod“



Verpasse keine Episode von “Neural intel Pod” und abonniere ihn in der kostenlosen GetPodcast App.








