Inner Alignment is the problem of ensuring mesa-optimizers (i.e. when a trained ML system is itself an optimizer) are aligned with the objective function of the training process.
Inner alignment asks the question: How can we robustly aim our AI optimizers at any objective function at all?
As an example, evolution is an optimization force that itself ‘designed’ optimizers (humans) to achieve its goals. However, humans do not primarily maximize reproductive success, they instead use birth control while still attaining the pleasure that evolution meant as a reward for attempts at reproduction. This is a failure of inner alignment.
The term was first given a definition in the Hubinger et al paper Risk from Learned Optimization:
We refer to this problem of aligning mesa-optimizers with the base objective as the inner alignment problem. This is distinct from the outer alignment problem, which is the traditional problem of ensuring that the base objective captures the intended goal of the programmers.
Goal misgeneralization due to distribution shift is another example of an inner alignment failure. It is when the mesa-objective appears to pursue the base objective during training but does not pursue it during deployment. We mistakenly think that good performance on the training distribution means that the mesa-optimizer is pursuing the base objective. However, this might have occurred only because there were some correlations in the training distribution resulting in good performance on both the base and mesa objectives. When we had a distribution shift from training to deployment it caused the correlation to be broken and the mesa-objective failed to generalize. This is especially problematic when the capabilities successfully generalize to the deployment distribution while the objectives/goals don’t. Since now we have a capable system that is optimizing for a misaligned goal.
To solve the inner alignment problem, some sub-problems that we would have to make progress on include things such as deceptive alignment, distribution shifts, and gradient hacking.
Inner Alignment Vs. Outer Alignment
Inner alignment is often talked about as being separate from outer alignment. The former deals with working on guaranteeing that we are robustly aiming at something, and the latter deals with the problem of what exactly are we aiming at. For more information see the corresponding tag.
It should be kept in mind that you can have both inner and outer alignment failures together. It is not a dichotomy and often even experienced alignment researchers are unable to tell them apart. This indicates that the classifications of failures according to these terms are fuzzy. Ideally, we don’t think of a binary dichotomy of inner and outer alignment that can be tackled individually but of a more holistic alignment picture that includes the interplay between both inner and outer alignment approaches.
Related Pages:
Mesa-Optimization, Treacherous Turn, Eliciting Latent Knowledge, Deceptive Alignment, Deception
This seems inaccurate to me. An AI can be inner aligned and still not aligned if we solve inner aliment but mess up outer alignment.
This text also shows up in the outer alignment tag: Outer Alignment—LessWrong
I’ve made an edit to remove this part.
I think the better phrasing would be “is the model going to do what the humans trained (or told) it to do?” (specifying a goal you really want is outer alignment).
I’m not actually sure about the difference here between this tag and Mesaoptimizers
I’m guessing the distinction was intended to be:
Mesa-Optimizers: Under what condition do mesa-optimizers arise, and how can we detect or prevent them (if we want to, and if that’s possible)?
Inner Alignment: How do you cause mesa-optimizers to have the same goal as the base optimizer? (Or maybe, more generally, how do you cause mesa-optimizers to have good desired properties?)
Or ‘Inner Alignment’ is meant to be a subcategory of ‘Mesa-Optimizers’?