How undesired goals can arise with correct rewards
Exploring examples of objective misgeneralisation – the place an AI system’s capabilities generalise however its objective would not
As we construct more and more superior synthetic intelligence (AI) methods, we need to make certain they don’t pursue undesired targets. Such behaviour in an AI agent is commonly the results of specification gaming – exploiting a poor selection of what they’re rewarded for. In our latest paper, we discover a extra delicate mechanism by which AI methods could unintentionally be taught to pursue undesired targets: goal misgeneralisation (GMG).
GMG happens when a system’s capabilities generalise efficiently however its objective doesn’t generalise as desired, so the system competently pursues the mistaken objective. Crucially, in distinction to specification gaming, GMG can happen even when the AI system is skilled with an accurate specification.
Our earlier work on cultural transmission led to an instance of GMG behaviour that we didn’t design. An agent (the blue blob, beneath) should navigate round its atmosphere, visiting the colored spheres within the appropriate order. Throughout coaching, there’s an “skilled” agent (the pink blob) that visits the colored spheres within the appropriate order. The agent learns that following the pink blob is a rewarding technique.
Sadly, whereas the agent performs nicely throughout coaching, it does poorly when, after coaching, we substitute the skilled with an “anti-expert” that visits the spheres within the mistaken order.
Although the agent can observe that it’s getting unfavourable reward, the agent doesn’t pursue the specified objective to “go to the spheres within the appropriate order” and as an alternative competently pursues the objective “comply with the pink agent”.
GMG is just not restricted to reinforcement studying environments like this one. In truth, it might happen with any studying system, together with the “few-shot studying” of enormous language fashions (LLMs). Few-shot studying approaches goal to construct correct fashions with much less coaching information.
We prompted one LLM, Gopher, to judge linear expressions involving unknown variables and constants, resembling x+y-3. To unravel these expressions, Gopher should first ask in regards to the values of unknown variables. We offer it with ten coaching examples, every involving two unknown variables.
At check time, the mannequin is requested questions with zero, one or three unknown variables. Though the mannequin generalises accurately to expressions with one or three unknown variables, when there aren’t any unknowns, it nonetheless asks redundant questions like “What’s 6?”. The mannequin at all times queries the person not less than as soon as earlier than giving a solution, even when it isn’t essential.
Inside our paper, we offer extra examples in different studying settings.
Addressing GMG is vital to aligning AI methods with their designers’ targets just because it’s a mechanism by which an AI system could misfire. This will likely be particularly vital as we method synthetic normal intelligence (AGI).
Take into account two attainable kinds of AGI methods:
- A1: Meant mannequin. This AI system does what its designers intend it to do.
- A2: Misleading mannequin. This AI system pursues some undesired objective, however (by assumption) can be good sufficient to know that it is going to be penalised if it behaves in methods opposite to its designer’s intentions.
Since A1 and A2 will exhibit the identical behaviour throughout coaching, the opportunity of GMG implies that both mannequin may take form, even with a specification that solely rewards meant behaviour. If A2 is discovered, it might attempt to subvert human oversight with the intention to enact its plans in direction of the undesired objective.
Our analysis group can be completely satisfied to see follow-up work investigating how probably it’s for GMG to happen in apply, and attainable mitigations. In our paper, we advise some approaches, together with mechanistic interpretability and recursive evaluation, each of which we’re actively engaged on.