gradient-routing-adapter TEC-0237Route flagged updates into a removable adapter
During post-training, send the gradient updates from samples an imperfect monitor flags as unwanted into a separate set of new parameters, then remove those parameters before deployment.
The GRAFT method (Gradient Routing Adapter Fine Tuning). Instead of penalising or filtering the samples a monitor flags, their updates go into a sacrificial adapter, and everything else trains as normal. Two consequences follow. The monitor never enters the reward, so the model has no incentive to learn to hide what the monitor looks for. And deleting the adapter afterwards removes the unwanted behaviour while keeping the task gains. It is a training-time technique: it needs access to the weights and a monitor that catches at least some of the misbehaviour.
Requires: Access to train the weights, and a monitor that flags at least some instances of the unwanted behaviour.
Contexts: Coding agent, Autonomous agent
Does it work?
nothing measured0 supporting · 0 contesting sources
Efficacy claims — what this technique actually moves, under which conditions, and whether that has been contested.
1 more condition from claims reviewed by AI, listed with the claims below.
No efficacy claim filed yet. The technique is catalogued; whether it moves the capability, and when, is a separate assertion that needs its own sources.
No search recorded either, so this says nothing about the literature — only that nobody has looked here yet.
Code
No repository linked yet. Contribute one.
Sources
Whether a technique works is a claim. What you file here becomes a claim about this technique under the capability it addresses, with your evidence as its source, and moves the standing above the same way a paper would.