[Submission] TabNaru: a 28.5M-parameter tabular regression model

## Summary

Hi TabArena team,

We would like to contribute **TabNaru**, a regression-only tabular foundation model with 28,538,682 parameters and a single checkpoint pretrained from random initialization.

- **Architecture:** Three interleaved column/row exchanges carry support-written column summaries between exchanges. Direct two-stream Birkhoff routing is applied around twelve attention and twelve feed-forward blocks.
- **Inference:** The fixed MN8 recipe requests eight views of the checkpoint, combining target Yeo–Johnson transforms, feature coverage, raw-plus-SVD augmentation and quantile integration. An NNLS combiner fits on an internal split of training support, with a fixed 75% fitted / 25% uniform mixture and uniform fallback for small validation sets or incompatible view rules. Neural weights remain frozen.

The implementation builds on TabICL components and DeepSeek-derived leaves. Related inference ideas are credited to TabICL, Nori and EXAONE.

## Evaluation

Historical regression-only comparisons cover **13 TabArena datasets / 222 splits** and **44 BeyondArena core datasets / 138 splits**. The frozen local pools contain 92 and 45 model/configuration entries, including earlier research configurations. Public reference results were combined with our runs, and Elo was recomputed for all entries using the pinned TabArena ranking code. The displayed methods share the same dataset/split coordinates within each benchmark.

| Model | TabArena Elo (position) | BeyondArena core Elo (position) |
|---|---|---|
| **TabNaru (MN8)** | **1747 (2/92)** | **1155.8 (6/45)** |
| Nori-30M (public default) | 1632 (8/92) | — |
| TabICLv2 (public default) | 1562 (25/92) | 1165.5 (5/45) |

These are **local comparison positions, not official leaderboard rankings**. Nori-30M is absent from this frozen BeyondArena reference pool. Ratings belong to separate pools and should be compared within each column.

These are exploratory point estimates, without a claim of statistical superiority. Submission selection followed inspection of benchmark results, and independent development confirmation did not pass. We will disclose the complete experiment history.

**Cost:** On B200 GPUs, attributed preprocessing, combiner-fitting and prediction time totaled **561.3 s** for the 222 TabArena splits and **1507.7 s** for the 138 BeyondArena splits. These are shared-run component totals, not standalone latency measurements. They do not establish a speedup over the reference models; standalone fit/predict timing remains pending.

## Release and integration

We are preparing a pip-installable inference package, ungated checkpoint and model card under **BSD-3-Clause** for owned code and weights, preserving third-party notices. Public URLs are not available yet. We plan to provide the standard Model wrapper and raw results with a PR; the latest upstream integration has not been evaluated.

## Question for maintainers

We propose one fixed configuration without a separate hyperparameter-search loop. Is support-internal NNLS acceptable in a Model entry, following the existing [EXAONE-Tabular wrapper](https://github.com/autogluon/tabarena/blob/903a377430f6f8ec06680a8a4a8f4db9b98f3e26/packages/tabarena/src/tabarena/models/exaone_tabular/model.py), or should this combiner use TabArena's supplied validation split?

Thank you!

Start by reading the referenced packages/tabarena/src/tabarena/models/exaone_tabular/model.py and the existing Model entry conventions. The issue asks maintainers to decide whether support-internal NNLS is acceptable or should use TabArena’s supplied validation split; completion would require an agreed integration plan, public package and checkpoint URLs, and a PR with the promised wrapper and raw results.

Written by the indexing model from the issue text.