When is Goodhart catastrophic?
Thanks to Aryan Bhatt, Eric Neyman, and Vivek Hebbar for feedback.
This post gets more math-heavy over time; we convey some intuitions and overall takeaways first, and then get more detailed. Read for as long as you’re getting value out of things!
How much should you optimize for a flawed measurement? If you model optimization as selecting for high values of your goal plus an independent error , then the answer ends up being very sensitive to the distribution of the error : if it’s heavy-tailed you shouldn’t optimize too hard, but if it’s light-tailed you can go full speed ahead.
Why the tails come apart by Thrasymachus discusses a sort of “weak Goodhart” effect, where extremal proxy measurements won’t have extremal values of your goal (even if they’re still pretty good). It implicitly looks at cases similar to a normal distribution.
Scott Garrabrant’s taxonomy of Goodhart’s Law discusses several ways that the law can manifest. This post is about the “Regressional Goodhart” case.
Scaling Laws for Reward Model Overoptimization (Gao et al., 2022) considers very similar conditioning dynamics in real-world RLHF reward models. In their Appendix A, they show a special case of this phenomenon for light-tailed error, which we’ll prove a generalization of in the next post.
Defining and Characterizing Reward Hacking (Skalse et al., 2022) shows that under certain conditions, leaving any terms out of a reward function makes it possible to increase expected proxy return while decreasing expected true return.
How much do you believe your results? by Eric Neyman tackles very similar phenomena to the ones discussed here, particularly in section IV; in this post we’re interested in characterizing that sort of behavior and when it occurs. We strongly recommend reading it first if you’d like better intuitions behind some of the math presented here—though our post was written independently, it’s something of a sequel to Eric’s.
An Arbital page defines Goodhart’s Curse and notes
The exact conditions for Goodhart’s Curse applying between and a point estimate or probability distribution over [a proxy measure that an AI is optimizing], have not yet been written out in a convincing way.
To the extent this post adopts a reasonable frame, we think it makes progress towards this goal.
Goodhart’s Law says
When a measure becomes a target, it ceases to be a good measure.
When I (Drake) first heard about Goodhart’s Law, I internalized something like “if you have a goal, and you optimize for a proxy that is less than perfectly correlated with the goal, hard enough optimization for the proxy won’t get you what you wanted.” This was a useful frame to have in my toolbox, but it wasn’t very detailed—I mostly had vague intuitions and some idealized fables from real life.
Much later, I saw some objections to this frame on Goodhart that actually used math. The objection went something like:
Let’s try to sketch out an actual formal model here. What’s the simplest setup of “two correlated measurements”? We could have a joint normal distribution over two random variables, and , with zero mean and positive covariance. You actually value , but you measure a proxy . Then we can just do the math: if I optimize really hard for , and give you a random datapoint with or something, how much do you expect to get?
If we look at the joint distribution of and , we’ll see a distribution with elliptical contour lines, like so:
Now, the naïve hope is that expected as a function of observed would go along the semi-major axis, shown in red below:
But actually we’ll get the blue line, passing through the points at which the ellipses are tangent to the -axis.
Importantly, though, we’re still getting a line: we get linearly more value for every additional unit of we select for! Applying percentile selection on isn’t going to be as good as percentile selection on , but it’s still going to give us more than any lower percentile selection on . The proxy is inefficient, but it’s not doomed.
Lately, however, I’ve come to think that this story is a little too rosy. One thing that’s going on here is that we’re just thinking about a “regressional Goodhart” problem, which is only one of several ways something Goodhart-like can come into play—see Scott Garrabrant’s taxonomy. But even in this setting, I think things can be much thornier.
In the story above, we can think of our measurement as being some multiple of plus an independent normally-distributed source of error, . When we ask for an outcome with a really high value of , we’re asking for a datapoint where is very high.
Because normal distributions drop off in probability very fast, it gets harder and harder to select for high values of either component: given that a datapoint is at least 4 standard deviations above the mean, the odds that it’s at least 5 standard deviations above are less than 1%. So the least-rare outcomes with high are going to look like a compromise between the noise and value , where we have a medium amount of each piece (because going to the extremes for either one is disproportionately costly in terms of improbability).
To see this more visually, here are some plots of possible pairs, restricted to the triangle of values where . Points are brighter if that outcome is more probable, and the black contour lines show regions of equal probability density. On the right, we have the expected value of as a function of our proxy threshold .
We can see that the most likely outcomes skew towards one side or the other depending on which of and has more variance, but because these contour lines are convex, we still expect to see outcomes that have some of each component.
But now let’s look at a case where and are heavy-tailed, such that each additional unit of or requires fewer bits of optimization power. Say that the probability density functions (PDFs) of and are proportional to , instead of like before. Then we’ll see something more like
The resulting distribution is symmetric about and , of course, but unlike in the normal case, that doesn’t manifest as ” and will be about the same”, but instead as “the outcome will be almost entirely or almost entirely with even odds”.
In this heavy-tailed regime, though, we care a lot about which of or has the edge here. For instance, suppose that optimizing a given amount for only gets us half as far as it would for (so e.g. the 99th percentile value is half as large as the 99th percentile value). Our plot now looks like
and in the limit for large we won’t get any expected at all by optimizing for the sum—all that optimization power goes towards producing high values. We call this catastrophic Goodhart because the end result, in terms of , is as bad as if we hadn’t conditioned at all.
(In general, if the right-hand tails of and are each on the order of , we’ll switch between the two regimes right at - that’s when these contour lines switch from being convex to being concave.)
To help visualize this behavior, let’s zoom in closer on a concrete example where we get catastrophic Goodhart. See below for plots of the PDFs of and :
On the left is a standard plot of the two PDFs; on the right is a plot of their negative logarithms. The right-hand plot makes it apparent that has heavier right tails, because the green line gets arbitrarily far below the orange line in the limit.
Here is a GIF of the conditional distribution on as goes from up to , with a dashed blue line indicating the conditional expectation:
Note the spike in the conditional PDF around , corresponding to outcomes where is small and is large; because of the heavier tails on , this spike gets smaller and smaller with larger . (We recommend staring at this GIF until you feel like you have a good understanding of why it looks the way it does.)
The expected value initially goes up when we apply a little selection pressure to our proxy, but as we optimize harder, that optimization pressure gets shunted more and more into optimization for , and less and less for , even in absolute terms. (This is the same dynamic that Eric Neyman recently discussed in section IV of How much do you believe your results?, put in a slightly different framing.)
In the next post, we’re going to prove some results about when this effect happens; this will be pretty technical, so we’ll talk a bit about the results in broad strokes here.
Suppose that and are independent real-valued random variables. We’ll show, roughly, that if
is subexponential (a slightly stronger property than being heavy-tailed).
has lighter tails than by more than a linear factor, meaning that the ratio of the tails of and the tails of grows superlinearly.
Less formally, we’re saying something like “if it requires relatively little selection pressure on to get more of and asymptotically more selection pressure on to get more of , then applying very strong optimization towards will not get you even a little bit of optimization towards - all the optimization power will go towards , where it has the best return on investment.”
We’ll also show a sort of inverse to this: if has right tails that are lighter than an exponential (for instance, if is normal or bounded), then we’ll get infinitely much in the limit no matter what kind of tail distribution has.
(What if is heavy-tailed but has even heavier tails than ? Then we can exchange their places in the first theorem, and conclude that we get zero in the limit—which means that all of that optimization is going towards .)
In the next post, we’ll prove these claims.
Application to alignment
We might want to use unaligned AI to generate alignment research for us. One model for this is sampling a random document from the space of 10000-bit strings, then conditioning on a high human rating. If evaluation of alignment proposals is substantially easier than generating good alignment proposals, these plans will be useful. If not, we’ll have a hard time getting research out of the AI. This is a crux between John Wentworth and Paul Christiano + Jan Leike that informs their differing approaches to alignment.
We can frame the problem of evaluation in terms of Goodhart’s Law. Let be the true quality of an alignment plan (say in utility contributed to the future), and be the human rating, so that is the human’s rating error. If and are independent, and we have access to arbitrarily strong optimization for , then our result implies that to implement an alignment plan better than random…
… if V is light-tailed, X must not be heavy-tailed.
… if V is heavy-tailed, X must not be much heavier-tailed than V.
We don’t know whether V is heavy- or light-tailed in real life, so to be safe, we should make X light-tailed. To the extent this model is accurate, a large part of alignment reduces to the problem of finding a classifier with light-tailed errors, which is able to operate in the exceptionally complicated domain of evaluating plans, and is not itself dangerous.
This model makes two really strong assumptions: that optimization is like conditioning, and that and are independent. These are violated in real life:
Optimization is not simply conditioning; SGD has too many inductive biases for us to list here, and (Gao et al., 2022) found that for a given level of optimization, RL uses far more KL distance from the prior than best-of-n sampling.
and will not be independent. Among other reasons, we expect that more complicated or optimized plans are more likely to have large impacts on the world (thus having higher variance of ), and harder to evaluate (thus having higher variance of ). However, in some cases, really good plans might be easier to evaluate; for example, formalized proofs can be efficiently checked.
There’s also a sort of implicit assumption in even using a framing that thinks about things as ; the world might be better thought of as naturally containing tuples (with our proxy measurement), and could be a sort of unnatural construction that doesn’t make sense to single out in the real world. (We do think this framing is relatively natural, but won’t get into justifications here.)
Despite these caveats, some takeaways we endorse:
Optimization for imperfect proxies is sometimes fine and sometimes doomed, depending on your distribution.
Goodhart’s law is subtle—even within a given framing of a problem, what happens when you optimize can be very sensitive to the exact numerical details of your measurements.
In particular, reaching for a normally-distributed toy model by default can be super misleading for thinking about a lot of real-world dynamics, because the tails are much lighter than most things in a way that affects the qualitative takeaways.
In an alignment plan involving generation and evaluation, you should either (a) have reason to believe that your classifier’s errors are light-tailed, (b) have a reason why training an AI on human (or AI) feedback will be importantly different from conditioning on high feedback scores, or (c) have a story for why non-independence works in your favor.
Show that when and are independent and , . Conclude that . This means that given independence, optimization always produces a plan that is no worse than random.
When independence is violated, an optimized plan can be worse than random, even if your evaluator is unbiased. Construct a joint distribution for and such that , , and for any , but .
Answers to exercises are at the end of the next post.
Thanks to Eric Neyman for first making this observation clear to me.
One way to see this intuitively is to consider the shear transformation replacing by , where is a constant such that the resulting random variable is uncorrelated with . In that situation we’d have a constant expectation of , so adding the component back in should give us a linear expectation.
To be precise, .
Technically we could have , but we can just rescale until the coefficient is 1 without changing anything.
Most heavy-tailed distributions are also long-tailed, which means that for all . So the optimization needed to get from the event ” is at least ” to ” is at least ” becomes arbitrarily small for large .
Note that this effect doesn’t depend on the behavior of or right around zero, just on their right tails.
We’ll suppose that has a PDF proportional to and has a PDF proportional to , where is an odd function that quickly asymptotes to , so has tails like for large in either direction but is smooth around .
We’ll use something slightly stronger than this; we’d like ’s tails to be larger by a factor of . More precise details in the next post.
I really like that this provides a framework to start thinking about when X is not random but adversarially selected.
Curated. Goodhart’s Law is an old core concept for LessWrong, and I love when someone(s) come along and add more resolution and rigor to our understanding, and all the more so when they start pointing to how this has practical implications. Would be very cool if this leads to articulation of disagreements between people that allow for progress in the discussion there, e.g. John vs Paul, Jan, etc.
And extra bonus points for exercises at the end too. All in all, good stuff, looking forward to seeing more – especially the results as your vary more of the assumptions (e.g. independence) to line up more with scenarios we anticipate in, e.g. Alignment scenarios.
Another piece of related work: Simon Zhuang, Dylan Hadfield-Mennel: Consequences of Misaligned AI.
The authors assume a model where the state of the world is characterized by multiple “features”. There are two key assumptions: (1) our utility is (strictly) increasing in each feature, so—by definition—features are things we care about (I imagine money, QUALYs, chocolate). (2) We have a limited budget, and any increase in any of the features always has a non-zero cost. The paper shows that: (A) if you are only allowed to tell your optimiser about a strict subset of the features, all of the non-specified features get thrown under the buss. (B) However, if you can optimise things gradually, then you can alternate which features you focus on, and somehow things will end up being pretty okay.
Personal note: Because of the assumption (2), I find the result (A) extremely unsurprising, and perhaps misleading. Yes, it is true that at the Pareto-frontier of resource allocation, there is no space for positive-sum interactions (ie, getting better on some axis must hurt us on some other axis). But the assumption (2) instead claims that positive-sum interactions are literally never possible. This is clearly untrue in the real-world, about things we care about.
That said, I find the result (B) quite interesting, and I don’t mean to hate on the paper :-).
I wonder if the brainstem is limiting optimization is some way like this. So far my assumption was that the brainstem uses some saturation and temporal decay for the multiple reward components to prevent Goodhardting. But maybe something closer to the t-limiting here.
Great post! I especially enjoyed the intuitive visualizations for how the heavy-tailed distributions affect the degree of overoptimization of X.
As a possibly interesting connection, your set of criteria for an alignment plan can also be thought of as criteria for selecting a model specification that approximates the ideal specification well, especially trying to ensure that the approximation error is light-tailed.
This was a fantastic read. Among my top three (at least) on Goodhart!
Stupid simple observation: if you could get enough independent evaluations of X you could smooth out heavy tails by ensembling (by central limit theorem).
actually independent, not like asking lots of humans to ‘independently’ rate something, which is obviously correlated in important ways—I think this condition is very hard to achieve in reality
I think these are generally not the same, and I object to any implied privileging of this hypothesis. But above you say that SGD has a ton of inductive biases in general, so why do you seem to endorse a takeaway like (my words) “you need to have a reason why SGD has the relevant inductive biases”?
SGD has inductive biases, but we’d have to actually engineer them to get high V rather than high X when only trained on U=V+X. In the Gao et al paper, optimization and overoptimization happened at the same relative rate in RL as in conditioning, so I think the null hypothesis is that training does about as well as conditioning. I’m pretty excited about work that improves on that paper to get higher gold reward while only having access to the proxy reward model.
I think the point still holds in mainline shard theory world, which in my understanding is using reward shaping + interp to get an agent composed of shards that value proxies that more often correlate with high V rather than higher X, where we are selecting on something other than U=V+X. When the AI ultimately outputs a plan for alignment, why would it inherently value having the accurate plan, rather than inherently value misleading humans? I think we agree that it’s because SGD has inductive biases and we understand them well enough to do directionally better than conditioning at constructing an AI that does what we want.
I think there is an additional effect related to “optimization is not conditioning” that stems from the fact that causation is not correlation. Suppose for argument’s sake that people evaluate alignment research partly based on where it’s come from (which the machine cannot control). Then producing good alignment research by regular standards is not enough to get high ratings. If a system manages to get good ratings anyway, then the actual papers it’s producing must be quite different to typical highly rated alignment papers, because they are somehow compensating for the penalty incurred by coming from the wrong source. In such a situation, I think it would not be surprising if the previously observed relationship between ratings and quality did not continue to hold.
This is similar to “causal Goodhart” in Garrabrant’s taxonomy, but I don’t think it’s quite identical. It’s ambiguous whether ratings are being “intervened on” in this situation, and actual quality is probably going to be affected somewhat. I could see it as a generalised version of causal Goodhart, where intervening on the proxy is what happens when this effect is particularly extreme.
I think this is more like Extremal Goodhart in Garrabrant’s taxonomy: there’s a distributional shift inherent to high U.
Maybe it’s similar, but high U is not necessary
Will you get into justifications in the next post? Because otherwise the following advice, which I consider literally correct:
in practice reduces just to the part “have a story for why inductive bias and/or non-independence work in your favor”, because I currently think Normality + additivity + independence are bad assumptions, and I see that as almost a null advice.
I think that Normality + additivity + independence come out together if you have a complex system subject to small perturbations, because you can write any dynamic as linear relationships over many variables. This gets you the three perks with:
Normality: complex system means many variables with nontrivial roles, and so the linearization tends to produce Normal distributions, it behaves like a sum with not too much concentrated weights.
Additivity: due to the small perturbations that allow you to linearize any relationship as approximation.
Independence: a linear system should be easy enough to analyze that you expect, if you spend effort, to get to a situation where the error is independent, and all the rest has been accounted for in some way.
Since we want to study the situation in which we apply a lot of optimization pressure, I think this scenario gets thrown out the window.
Do you have a more general reason to expect these assumptions? Possibly each one or subsets separately? First raw ideas that come to my mind:
Normality because the number of variables involved grows in a balanced way with nonlinearity such that you get Normality
Additivity because scenario we can realistically study are limited enough that the kind of errors you can make stay the same, and we have to deliberately put ourselves in that situation
Independence because a human manages to get as much information as possible until some hard boundary of chaos
Do you have some clever trick such that it is always possible to always see the problem in this light? I expect not because utilities can only be affinely transformed.
my gut instinct tells me to look at elliptical distributions like exp(−(ax2+by2)c), which will not show this specific split-tail behavior. My gut instinct is not particularly justified, but seems to be making weaker assumptions.
I’m not sure what you mean formally by these assumptions, but I don’t think we’re making all of them. Certainly we aren’t assuming things are normally distributed—the post is in large part about how things change when we stop assuming normality! I also don’t think we’re making any assumptions with respect to additivity; X=U−V is more of a notational or definitional choice, though as we’ve noted in the post it’s a framing that one could think doesn’t carve reality at the joints. (Perhaps you meant something different by additivity, though—feel free to clarify if I’ve misunderstood.)
Independence is absolutely a strong assumption here, and I’m interested in further explorations of how things play out in different non-independent regimes—in particular we’d be excited about theorems that could classify these dynamics under a moderately large space of non-independent distributions. But I do expect that there are pretty similar-looking results where the independence assumption is substantially relaxed. If that’s false, that would be interesting!
I wasn’t saying you made all those assumption, I was trying to imagine an empirical scenario to get your assumptions, and the first thing to come to my mind produced even stricter ones.
I do realize now that I messed up my comment when I wrote
Here there should not be Normality, just additivity and independence, in the sense of U−V⊥V. Sorry.
I do agree you could probably obtain similar-looking results with relaxed versions of the assumptions.
However, the same way U−V⊥V seems quite specific to me, and you would need to make a convincing case that this is what you get in some realistic cases to make your theorem look useful, I expect this will continue to apply for whatever relaxed condition you can find that allows you to make a theorem.
Example: if you said “I made a version of the theorem assuming there exists f such that f(U,V)⊥V for f in some class of functions”, I’d still ask “and in what realistic situations does such a setup arise, and why?”
In my frame, U is not just some variable correlated with V, it’s some estimator’s best estimate, and so it makes sense that residuals X=U−V would have various properties, for the same reason we consider residuals in statistics, returns in finance, etc.
The basic idea why we might get U−V⊥V is that there are some properties that increase the overseer’s rating and actually make the plan good (say, the plan includes a solution to the shutdown problem, interpretability, or whatever) and different properties that increase the overseer’s rating for no good reason (e.g. the plan uses really sophisticated words and an optimistic tone). I think assuming these are independent and additive is reasonable as a toy model, though as we said they’re probably violated in real life and we’re interested in weakening these assumptions.
I guess you could get an elliptical distribution through something like this: all properties contribute to both X and V to some degree, and distribution of the angle is roughly uniform while the magnitudes are heavy-tailed. I’m not sure whether this is as natural as independence: if some property of the AI’s output makes the human irrationally approve of it (high X), then it seems likely to be optimized for that, rather than also having huge impacts on V one way or the other.
Are you saying that your (rough, preliminary) justification for independence is that it’s what gets you Goodhart, so you use it? Isn’t this circular? Ok so maybe I misinterpreted your intentions: I thought you wanted to “prove” that Goodhart happens, while possibly you wanted to “show an example” of Goodhart happening?
It doesn’t look circular to me? I’m not assuming that we get Goodhart, just that properties that result in very high X seem like they would be things like “very rhetorically persuasive” or “tricks the human into typing a very large number into the rating box” that won’t affect V much, rather than properties with very high magnitude towards both X and V. I believe this less for V, so we’ll probably have to replace independence with this.
I think you’re splitting hairs. We prove Goodhart follows from certain assumptions, and I’ve given some justification for the assumptions as well as their limitations, so you could equally say that we “prove” or “show an example”. If by circular you mean we proved something about independent X and V because this was easier than more realistic assumptions, we’re guilty! The proof was a huge pain and we wanted to publish rather than overcomplicating it more, partly to get feedback like yours. But I do have some intuition that the result is useful, partly because things are sometimes approximately independent, and partly because the basic reasons behind the proof extend to other cases.
An example of the sort of strengthening I wouldn’t be surprised to see is something like “If V is not too badly behaved in the following ways, and for all v∈R we have [some light-tailedness condition] on the conditional distribution (X|V=v), then catastrophic Goodhart doesn’t happen.” This seems relaxed enough that you could actually encounter it in practice.