# MoMaTune: Model, Match, Then Tune Language Models for Numerical Records

arXiv: https://arxiv.org/abs/2610.15235 · HTML: https://mlsanity.com/paper/2610.15235/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 18 of its 18 references that are in the map.
- First posted on arXiv 11 Oct 2026.

## Builds on (most cited first, 18 of 18)

- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Qwen Technical Report](https://mlsanity.com/paper/2309.16609/) (2023)
- [OpenML: networked science in machine learning](https://mlsanity.com/paper/1407.7722/) (2014)
- [NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models](https://mlsanity.com/paper/2405.17428/) (2024)
- [LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders](https://mlsanity.com/paper/2404.05961/) (2024)
- [Improving Text Embeddings with Large Language Models](https://mlsanity.com/paper/2401.00368/) (2023)
- [Language Models are Realistic Tabular Data Generators](https://mlsanity.com/paper/2210.06280/) (2022)
- [TabLLM: Few-shot Classification of Tabular Data with Large Language Models](https://mlsanity.com/paper/2210.10723/) (2022)
- [Generative Representational Instruction Tuning](https://mlsanity.com/paper/2402.09906/) (2024)
- [Do NLP Models Know Numbers? Probing Numeracy in Embeddings](https://mlsanity.com/paper/1909.07940/) (2019)
- [SubTab: Subsetting Features of Tabular Data for Self-Supervised Representation Learning](https://mlsanity.com/paper/2110.04361/) (2021)
- [Rank-N-Contrast: Learning Continuous Representations for Regression](https://mlsanity.com/paper/2210.01189/) (2022)
- [Making Pre-trained Language Models Great on Tabular Prediction](https://mlsanity.com/paper/2403.01841/) (2024)
- [xVal: A Continuous Numerical Tokenization for Scientific Language Models](https://mlsanity.com/paper/2310.02989/) (2023)
- [From Supervised to Generative: A Novel Paradigm for Tabular Deep Learning with Large Language Models](https://mlsanity.com/paper/2310.07338/) (2023)
- [TabGemma: Text-Based Tabular ICL via LLM using Continued Pretraining and Retrieval](https://mlsanity.com/paper/2511.03570/) (2025)
- [Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models](https://mlsanity.com/paper/2409.17836/) (2024)
- [Revealing the Numeracy Gap: An Empirical Investigation of Text Embedding Models](https://mlsanity.com/paper/2509.05691/) (2025)

## Abstract (arXiv)

Despite remarkable progress in language understanding and generation, decoder language models continue to exhibit striking failures on numerical data. Numerical examples are typically serialized as text by concatenating feature names with their values (e.g., fat:3.4), while conventional task adaptation applies loss only to the downstream prediction target. This leaves two structures weakly supervised: within-value sensitivity, where digit position determines numerical magnitude, because token cross-entropy contains no explicit numerical metric; and across-field structure, where measurements in the same record are statistically dependent, because target-only tuning provides no direct loss on the measurements or their dependencies. We introduce MoMaTune (Model, Match, then Tune), a two-stage post-training strategy organized around these supervision gaps. Stage 1 applies Model, ordinary next-token prediction densely over label-free serialized records, and optionally adds Match, a same-record consistency objective that aligns representations of different partial views of the same example. Stage 2 applies Tune, supervised fine-tuning for downstream target prediction. We evaluate MoMaTune on 169,285 records from eight public datasets, with 339 of 340 input fields numerical. At 1.5B parameters, Model raises the eight-dataset masked-value imputation aggregate from 0.106 to 0.401, while the three-dataset correct-context gain rises from 0.058 to 0.167. On diagnostics, Model also increases sensitivity to numerically consequential digit changes, while Match improves robustness to informative partial views. Adding Tune yields higher target-supervised embedding aggregates for each tested Stage-1 parent, and the two-stage recipes obtain higher aggregate scores than Tune alone. The same pattern is observed at 4B, 9.6B, and 27B parameters when model embeddings are used for downstream prediction.