Telling a Model to Be Culturally Aware Widens the Gap

The UK and US absorb roughly 35% of all default selections in DiSCo-Bench, a 304-item cultural evaluation covering 12 cultures. The figure comes from DiSCo, an arXiv preprint (2609.10253), which runs six instruction-tuned models through a design that differs from most cultural benchmarks in one specific way: it never asks which answer is correct.

Most cultural evals score accuracy against a single right answer. That measures whether the model knows a fact. It does not measure which culture the model reaches for when several answers are all defensible, and it blends the model's default preference together with its willingness to follow context. DiSCo keeps those apart. Items are forced-choice, the options are all culturally grounded, and the scored object is the distribution over selections rather than a pass rate.

The abstract's argument is that a single-correct benchmark is hard to use for characterising a model's cultural preference prior when several culturally grounded responses are all valid. My inference, and not the paper's: key one option as correct and you are partly measuring the key-writer's culture. Either way the interesting quantity becomes the ratio in which the model picks each option, and that ratio is what this benchmark reports. Forced-choice scoring over options that are all culturally grounded is the piece I would port into other evals first.

The same items then run at four context levels, labelled C0 to C3. The abstract calls this a four-level context gradient and does not spell out the levels. Functionally it is the steerability test: change the available context and see whether the distribution shifts.

Prompt-based steering widened the selection gap between high- and low-resource cultures. The abstract reports the direction and no magnitude, so how far it widens is not checkable from what is public. The direction is the part worth sitting with: asking a model to be culturally appropriate did not spread its selections more evenly across the twelve. It made the skew worse.

Why steering backfires is not something the abstract addresses. My guess, and the abstract gives nothing to check it against: a model told to be culturally appropriate falls back on the cultural default it learned during pretraining rather than the one the prompt names.

Injecting explicit cultural facts into the prompt produced a negligible distributional shift. Again, no effect size.

Both of the moves teams reach for first do little or nothing. Prompt-based steering pushes the distribution the wrong way. Injected cultural facts leave it roughly where it was. DiSCo's own conclusion is that cultural preference bias cannot be resolved through prompt-based personalisation alone.

If you ship to more than one market, your localisation test is probably measuring the wrong thing. If your harness asks whether the model returns the correct culturally specific answer after you tell it which culture you want, you are testing instruction following on a known context, which is close to the easiest case. The harder case is a prompt with no cultural signal at all, where the model picks a default and everyone outside the preferred markets gets whatever the prior prefers. DiSCo's argument is that the second case is where the bias lives, and that a single-answer key cannot see it.

Score distributions, not pass rates. Build the item as a set of plausible culturally grounded options, record which one the model picks, then repeat it at several context levels, keeping the item fixed across them. Report the deltas between levels rather than an absolute score at any one of them. Run it per locale, since a model that averages well across 12 cultures can still be badly skewed on any one of them. If the distribution barely moves as context changes, you have learned something an accuracy score would have hidden.

Now the arithmetic, which is mine and not the paper's. Twelve cultures selected uniformly would be about 8.3% each, so the UK and US pair would sit near 16.7%. The reported figure is a little more than double that. Put another way, if two cultures take 35% of the selections, the remaining ten split 65% between them at about 6.5% each, against roughly 17.5% for each of the top two. Which ten sit at the bottom is not in the abstract, and the bottom of the distribution is the number a localisation team would actually want.

Resolution is the other thing to settle before trusting any per-culture result. The abstract reports no replication and no per-culture item counts. If the 304 items were split evenly, that works out to about 25 per culture, and the abstract does not say they are. Estimate a rate near 50% from 25 observations and the 95% interval is roughly 20 points either way. Any per-culture ranking built on that few items is closer to a coin toss than an ordering. That is an argument for the distributional measure, which pools all 304 items, and an argument against reading much into any single culture's row.

The abstract names neither the six models nor any per-model result, so there is no way to tell whether this is one family's problem or a general property of instruction tuning. The C0 to C3 levels are not described, so the steerability finding is hard to reproduce from the abstract alone. And the per-culture tables are absent, so the shape of the distribution is asserted at the top and unspecified below it.

What would change my mind: per-model and per-culture tables showing the concentration is driven by one or two models rather than the group, or a non-prompt intervention, a fine-tune or a decoding constraint, that actually moves the distribution.