Learning the Pareto Front Beyond Linear Scalarization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2326 papers429 conceptsfrozen 2026-10-04

In machine learning, we rarely have a single goal. We want models that are accurate but also fair, fast but also energy-efficient, or high-performing across multiple distinct tasks. The standard way to handle these competing objectives is linear scalarization: we assign a weight to each goal and minimize their weighted sum.

While intuitive, this approach has a fundamental geometric flaw. As researchers like Michele Donini and Abbas Abdolmaleki have pointed out, linear scalarization is mathematically incapable of identifying optimal trade-offs located in the non-convex regions of a "Pareto front."

The Geometry of the Trade-off

To understand why linear scalarization fails, we must look at the Pareto front—the set of all solutions where you cannot improve one objective without making another worse.

If the relationship between two objectives is convex (bowed outward), a linear weighted sum works perfectly. By changing the weights, you can "slide" along the curve and find every possible optimal point. However, real-world machine learning problems—from Domain Adaptation to Fairness—often involve non-convex (inward-bending) trade-off surfaces.

In these non-convex regions, a linear weight acts like a flat board being pressed against a jagged rock. The board will only ever touch the "peaks" (the convex hull). No matter how you tilt the board (change the weights), it will never touch the "valleys" (the concave regions). As Aviv Navon and colleagues argue, this means that the most nuanced compromise solutions—the ones that actually balance competing goals effectively—are often physically unreachable using standard weighted losses.

Moving Beyond Weighted Sums

To find these "hidden" optimal solutions, researchers are moving toward non-linear methods. One powerful alternative is Chebyshev (or Tchebycheff) scalarization. Instead of summing objectives, this method minimizes the distance to a "reference point" (an ideal but unattainable goal). Geometrically, this creates L-shaped contours that can "reach into" the concave regions of the Pareto front, as demonstrated in The Fairness-Accuracy Pareto Front.

Other approaches focus on the Hypervolume of the reached set. Qiuyi Zhang and Daniel Golovin have shown that maximizing hypervolume provides a more robust signal for optimization than simple sums, as it naturally pushes the frontier outward in all directions simultaneously.

Learning the Entire Front at Once

Perhaps the most significant shift is from finding a point on the front to learning the entire front. Traditionally, if you wanted to see 10 different trade-offs, you had to train 10 different models.

Recent work on Pareto Set Learning (PSL) changes this. Xi Lin and Xiaoyuan Zhang have developed frameworks that train a single model to approximate the entire set of optimal trade-offs in one go. Similarly, Aviv Navon introduced Pareto HyperNetworks (PHNs), which are meta-models that take a "preference vector" as input and instantly output the weights for a model that satisfies that specific trade-off.

Why This Matters for Production

In production environments like LinkedIn's notification systems, the "correct" weight between engagement and user fatigue might change daily. If your optimization framework can only see the extreme ends of the spectrum because of linear scalarization, you are forced to choose between "too many notifications" and "too few."

By using methods that can navigate non-convex fronts, such as Conservative Q-Learning (CQL) or COSMOS, practitioners can discover the "sweet spot" compromises that were previously invisible to their algorithms. As Josif Grabocka notes, these techniques allow us to treat the trade-off as a continuous landscape to be explored, rather than a binary choice to be made.

Papers in this version

Showing 60 of 326 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 323 papers
  2. Version 1 · 2026-10-04 · 326 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_linear_scalarization_limitations_95162382,
  title = {linear scalarization limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/linear-scalarization-limitations/rdr_95162382a4a74749bd4d7d550bfc9a30},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_the_pareto_front_beyond_linear__95162382_v1,
  title = {Learning the Pareto Front Beyond Linear Scalarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-the-pareto-front-beyond-linear-scalarization/rdr_95162382a4a74749bd4d7d550bfc9a30/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}