Analysis
Exploring examples of aim misgeneralisation – the place an AI system’s capabilities generalise however its aim does not
As we construct more and more superior synthetic intelligence (AI) methods, we wish to make certain they don’t pursue undesired targets. Such behaviour in an AI agent is usually the results of specification gaming – exploiting a poor selection of what they’re rewarded for. In our latest paper, we discover a extra refined 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 improper 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 surroundings, visiting the colored spheres within the right order. Throughout coaching, there’s an “knowledgeable” agent (the crimson blob) that visits the colored spheres within the right order. The agent learns that following the crimson blob is a rewarding technique.
The agent (blue) watches the knowledgeable (crimson) to find out which sphere to go to.
Sadly, whereas the agent performs nicely throughout coaching, it does poorly when, after coaching, we change the knowledgeable with an “anti-expert” that visits the spheres within the improper order.
The agent (blue) follows the anti-expert (crimson), accumulating damaging reward.
Regardless that the agent can observe that it’s getting damaging 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 “observe the crimson agent”.
GMG just isn’t restricted to reinforcement studying environments like this one. The truth is, it may well 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 knowledge.
We prompted one LLM, Gopher, to judge linear expressions involving unknown variables and constants, corresponding to x+y-3. To unravel 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 aren’t any unknowns, it nonetheless asks redundant questions like “What’s 6?”. The mannequin all the time queries the consumer not less than as soon as earlier than giving a solution, even when it’s not crucial.
Dialogues with Gopher for few-shot studying on the Evaluating Expressions process, with GMG behaviour highlighted.
Inside our paper, we offer further 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 shall be particularly essential as we strategy synthetic common intelligence (AGI).
Think about 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 aim, 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 potential for GMG signifies 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 to be able to enact its plans in the direction of the undesired aim.
Our analysis workforce can be blissful to see follow-up work investigating how probably it’s for GMG to happen in observe, 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. You probably have come throughout aim misgeneralisation in AI analysis, we invite you to submit examples here.
