Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{206734,
        author = {Chandan Kumar},
        title = {Tabular Latent Fusion for Language Models : A Token-Free Framework for Injecting TabFM Representations into Autoregressive Decoders},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {2},
        pages = {2589-2596},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=206734},
        abstract = {Large Language Models (LLMs) are increasingly used to reason over structured data, yet most practical systems still convert tables into textual prompts such as CSV, JSON, Markdown, or natural-language summaries. This text-serialization step increases token cost, weakens numerical fidelity, and obscures the two-dimensional row-column structure that makes tabular data meaningful. This paper presents a research-oriented framework for tabular latent fusion: a token-free mechanism that encodes tables with a Tabular Foundation Model (TabFM), projects the resulting continuous representation into an LLM-compatible hidden space, and fuses it with an autoregressive decoder through cross-modal attention, residual injection, or gated Bayesian correction. The accompanying notebook, research1.ipynb, compares direct table prompting, compact TabFM-vector prompting, local GPT-2 decoding, and TabFM-to-GPT-2 Bayesian/cross-attention injection. This work is positioned as a prototype-level architectural study rather than a full production benchmark. Prototype experiments are intended to test feasibility, token pressure, runtime behavior, and architectural trade-offs.},
        keywords = {tabular foundation models; large language models; TabFM; GPT; multimodal fusion; cross-attention; Bayesian correction; token efficiency; structured data reasoning.},
        month = {July},
        }