Ufinq
Real-world · Limitations

Known limitations

What the real-world base and the standardized sweeps show Ufinq does not do well, each with the numbers behind it and its status.

This list is a file in the Ufinq repository (limitations-data.ts), changed only by merge request, so every addition, change of status and removal is on record. An entry is added when the evidence is there and stays until a release addresses it, after which it stays with the release named.

6 entries · 5 open · 1 in progress · 0 addressed

  1. Quality falls with the size of the data.

    open

    Median quality 0.68 under 100 rows and 0.48 at 10k to 100k rows; tall tables (100 and more rows per column) 0.57 against wide tables 0.78. Live base, 2026-10-10.

    Recorded 2026-10-10 · #quality-falls-with-size

  2. Some analyses deliver no model worth the name.

    open

    9 of 158 models have a quality of 0.1 or less. Live base, 2026-10-10. Without a fit record the base cannot say why.

    Recorded 2026-10-10 · #no-model-worth-the-name

  3. The base is mostly example projects.

    open

    157 of 163 tagged analyses are textbook or benchmark datasets published as examples. Live base, 2026-10-10. The overview is not yet a picture of user problems.

    Recorded 2026-10-10 · #mostly-example-projects

  4. Standardized: no lead on the holdout seeds.

    open

    On the weekly sweep’s holdout seeds 6 to 10 Ufinq shows no lead over PySR. Weekly deep read, 2026-09-27.

    Recorded 2026-10-10 · #standardized-no-lead-on-holdout-seeds

  5. Medicine and health is the weakest domain.

    open

    Median quality 0.51 over 19 models. Live base, 2026-10-10. Observed only; no suite has compared other methods there.

    Recorded 2026-10-10 · #medicine-and-health-weakest-domain

  6. Few fit records yet.

    in progress

    Only models fitted since Ufinq 0.3.34 (2026-10-03) carry held-out quality, the linear reference and the version: 3 of 163 on 2026-10-10. Fills as owners run new analyses; no backfill is possible.

    Recorded 2026-10-10 · #few-fit-records

How to read this

  • Evidence. A number from the live base is an observation over the public analyses as their owners ran them, on the date given; it changes as the base grows. A number from a standardized sweep is a controlled result.
  • Status. Open: nothing has been done. In progress: the cause is being worked on, or the numbers fill in by themselves over time. Addressed in a release: the release is named and the entry keeps its before-and-after.
  • Where the numbers come from. The map and the overview for the real-world base, the standardized report for the sweeps; releases are on the changelog.