Recently, when looking at AI model performance reports, I always come across a term: Pareto frontier.
It looks like an advanced mathematical concept. In fact, it answers a very practical question: Models vary in intelligence, cost, and speed—how should we compare them?
This is where Pareto optimality comes into play: You cannot find another solution that is no worse on all metrics and strictly better on at least one.
Understanding It Through Dividing Apples
Suppose two people are dividing 10 apples. The apples are identical, and each person only cares about the quantity they receive—the more, the better.
Currently, each has taken 4, leaving 2 undistributed. If you give one more apple to either person without causing any loss to the other, this is called a “Pareto improvement”: At least one person gets better off, and no one gets worse off.
If the division is already 5 and 5, for one person to get more, the other must get less. This allocation achieves Pareto optimality.
However, a split of 9 and 1 also meets this condition.
Therefore, Pareto optimality does not guarantee fairness, nor does it ensure everyone is satisfied. It simply indicates that there is no room for improvement without harming anyone.
The judgment here relies on the assumptions mentioned earlier. If someone dislikes apples or if the apple qualities differ, the conclusion would need to be re-analyzed.
Why AI Performance Reports Favor This Concept
When comparing AI models, we often care simultaneously about answer quality, price, and response speed.
If you look only at quality, you might select a very expensive and slow model; if you look only at price, you might find the answers unusable. This is a multi-objective problem.
Assume we compare only quality and cost, with consistent tasks and test conditions, where cost represents the average expense to complete the same task:
| Model | Quality Score (Higher is Better) | Cost per Task (Lower is Better) |
|---|---|---|
| A | 90 | ¥1 |
| B | 85 | ¥0.2 |
| C | 80 | ¥0.5 |
| D | 90 | ¥1.2 |
- Look at C: B has better quality and is cheaper. Considering only these two metrics, C has no advantage. In technical terms, B dominates C.
- Look at D: A has the same quality but is cheaper. Therefore, A dominates D.
- Remaining are A and B; neither can comprehensively surpass the other. A has higher quality, while B is cheaper. They are both non-dominated solutions among this group of candidate models.
This corresponds to a basic characteristic of multi-objective optimization: The result is often a set of solutions involving trade-offs. Refer to pymoo official documentation on multi-objective optimization.
How Is It Calculated?
The most intuitive algorithm involves pairwise comparison. For Model A, if you can find another Model B that simultaneously satisfies:
- B’s quality is not lower than A’s;
- B’s cost is not higher than A’s;
- At least one metric is strictly better;
Then A is dominated and can be excluded from the candidate frontier.
Perform this check for every model, retaining those not dominated by any other model, and you obtain the set of non-dominated solutions within this batch of candidates. The metric points corresponding to these solutions constitute the Pareto frontier for this dataset.
Adding metrics like latency or VRAM usage works the same way: A solution dominates another only if it is no worse on all metrics and strictly better on at least one.
However, a model eliminated under two metrics might re-enter the frontier upon adding a third metric. For example, although C is more expensive and has lower quality, it might be much faster.
What Does “Pushing the Pareto Frontier” Mean?
Continuing with the previous example, if a new Model E has a quality score of 90 but its cost drops to ¥0.6, it dominates A.
With the same quality, it can now be achieved for less money. This is one scenario where the frontier moves in a more favorable direction. There might also be models with higher quality at the same price, or new intermediate options: stronger than cheap models but cheaper than high-quality ones. New solutions may expand part of the frontier rather than completely replacing old models.
So, when seeing claims about “pushing the frontier,” keep asking: Which specific metrics are being improved, which segment of the frontier is affected, and which competitors were compared against?