I’m helping build geodesicresearch.ai
Puria
Karma: 400
Because motivations are so underdetermined by the reward signal, we may be able to shape them without running into the downstream problems of blocking access to high-rewards, such as deceptive alignment
Could you help me understand this one a bit better? I would have thought that the degenerate mapping of motivations to actions means that many motivations can be mapped onto the same high-reward actions, and deceptive alignment is a case where actions consistently achieve a high reward, but its underlying motivations are misaligned, not vice versa. Sorry if I’m misunderstanding this sentence!
This is great—thanks for putting such a clear taxonomy and explanation together. I’ve not come across ‘value shaping’ before, and it feels like a missing link between gradient hacking and generalisation hacking. The definition of value shaping:
By anticipating the reward function, the model intentionally generates trajectories that earn high reward while including preferred values, backdoors, or steganographic triggers, forcing gradient descent to reinforce them.
[...] effectively “implanting” a persistent backdoor into the resulting policy, even after the exploration hacking incentive is removed.
is amenable to being read as “manipulating the loss landscape”, ableit by associating a particular cognitive pattern with high reward/low loss, and also as “encod[ing] information causing arbitrary OOD policies after [training]”. Certainly in generalisation hacking it was “the model’s goal is to reinforce or maintain a particular behavior, not to avoid [low loss]”.
Is this a valid way to think about this or am I missing the mark?
Strong agree! Stress testing alignment methods against heavy capabilities/RL training out in the open is the main focus of Geodesic’s research agenda. We’re focussed on the training methods you can use to improve an initialisation (midtraining through to warm-start reasoning) going into RL (benefitting greatly from NVIDIA’s open sourced post-training stack here), but we’re hoping this will spur on more open science into data heavy interventions at larger scales.