Analysis
Exploring examples of aim misgeneralisation – the place an AI system’s capabilities generalise however its aim would not
As we construct more and more superior synthetic intelligence (AI) methods, we need to ensure that 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 might unintentionally study to pursue undesired targets: goal misgeneralisation (GMG).
GMG happens when a system’s capabilities generalise efficiently however its aim doesn’t generalise as desired, so the system competently pursues the fallacious aim. Crucially, in distinction to specification gaming, GMG can happen even when the AI system is educated 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, under) should navigate round its setting, visiting the colored spheres within the right order. Throughout coaching, there’s an “professional” agent (the purple blob) that visits the colored spheres within the right order. The agent learns that following the purple blob is a rewarding technique.
The agent (blue) watches the professional (purple) to find out which sphere to go to.
Sadly, whereas the agent performs nicely throughout coaching, it does poorly when, after coaching, we exchange the professional with an “anti-expert” that visits the spheres within the fallacious order.
The agent (blue) follows the anti-expert (purple), accumulating destructive reward.
Despite the fact that the agent can observe that it’s getting destructive reward, the agent doesn’t pursue the specified aim to “go to the spheres within the right order” and as an alternative competently pursues the aim “comply with the purple agent”.
GMG isn’t restricted to reinforcement studying environments like this one. In actual fact, it could happen with any studying system, together with the “few-shot studying” of huge language fashions (LLMs). Few-shot studying approaches intention 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 resolve these expressions, Gopher should first ask concerning 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 appropriately to expressions with one or three unknown variables, when there are not any unknowns, it nonetheless asks redundant questions like “What’s 6?”. The mannequin at all times queries the consumer not less than as soon as earlier than giving a solution, even when it isn’t mandatory.
Dialogues with Gopher for few-shot studying on the Evaluating Expressions job, with GMG behaviour highlighted.
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 might misfire. This will likely be particularly crucial as we method synthetic basic intelligence (AGI).
Contemplate two attainable varieties of AGI methods:
- A1: Supposed mannequin. This AI system does what its designers intend it to do.
- A2: Misleading mannequin. This AI system pursues some undesired aim, however (by assumption) can also be sensible sufficient to know that will probably 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 potential for GMG implies that both mannequin might take form, even with a specification that solely rewards supposed behaviour. If A2 is discovered, it will attempt to subvert human oversight with the intention to enact its plans in direction of the undesired aim.
Our analysis workforce could be blissful to see follow-up work investigating how probably it’s for GMG to happen in follow, and attainable mitigations. In our paper, we recommend some approaches, together with mechanistic interpretability and recursive evaluation, each of which we’re actively engaged on.
We’re at the moment amassing examples of GMG on this publicly available spreadsheet. When you have come throughout aim misgeneralisation in AI analysis, we invite you to submit examples here.
