🧬 biology

Mind the Gap No More: Achieving Zero-Gap Multimodal Integration via One Tokenizer

This paper proposes "One Tokenizer," a native multimodal architecture that eliminates the geometric modality gap by mapping all inputs into a unified token space, thereby enabling superior deep reasoning and outperforming traditional modular approaches in DNA-text tasks.

Original authors: Yanan Li, Christina Yi Jin, Yuan Jin, Manli Luo, Tie Xu, Shuai Jiao, Wei He, Qing Zhang

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yanan Li, Christina Yi Jin, Yuan Jin, Manli Luo, Tie Xu, Shuai Jiao, Wei He, Qing Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: The "Language Barrier" Between Data Types

Imagine you have a brilliant translator (a Large Language Model, or LLM) who speaks perfect English. Now, you want this translator to understand two very different things at the same time:

Text: A story written in English.
DNA: A biological code made of letters (A, C, T, G).

The Old Way (Modular Architecture):
Currently, most AI systems handle this by hiring two separate specialists.

Specialist A reads the DNA and writes a summary in a secret code.
Specialist B reads the English story.
They both hand their notes to the Translator.

The problem? Specialist A and Specialist B speak different "languages" and write in different styles. When they hand their notes to the Translator, the Translator has to spend a huge amount of brainpower just trying to figure out how to make those two different notes fit together. It's like trying to glue a square peg into a round hole. The AI has to constantly "reconcile" the differences, which wastes energy and often leads to misunderstandings. This difference between the two notes is what the authors call the "Modality Gap."

The New Solution: "One Tokenizer"

The authors propose a radical new idea: Stop hiring specialists. Just teach the Translator to read everything in the same language from the start.

They created a system called One Tokenizer. Instead of using a separate specialist to translate DNA, they simply added DNA "words" directly into the Translator's own dictionary.

Now, when the AI sees a DNA sequence, it doesn't see a foreign code needing translation. It sees a native word, just like it sees the word "cat" or "run."
Both the DNA and the English text are fed into the AI's brain as the exact same type of "token" (a digital word).

The "Zero-Gap" Analogy: The Great Hall vs. The Two Rooms

To understand why this is better, imagine the AI's brain as a building with many floors (layers).

The Old Way (Two Rooms): On the first floor, DNA lives in Room A and Text lives in Room B. There is a thick, impenetrable wall between them. As the information travels up the floors, the AI has to keep building bridges to connect the rooms. Even on the top floor, the two groups are still slightly separated, and the AI is exhausted from trying to build those bridges.
The New Way (One Great Hall): With One Tokenizer, there are no separate rooms. DNA and Text are all in one giant Great Hall from the very first step. They are mixed together naturally. Because they started in the same space, they stay together all the way to the top floor. There is no wall to bridge, so the AI can focus entirely on understanding the meaning of the story and the DNA, rather than wasting energy on geometry and translation.

What Did They Prove?

The paper makes two main claims, backed by math and experiments:

The Math Proof: The authors used geometry to prove that if you start with two separate vocabularies (the old way), there will always be a gap between them that is hard to close. But if you use one shared vocabulary (the new way), the gap is mathematically zero from the start and stays zero all the way through the deep layers of the AI.
The Experiment: They tested this using DNA and Text (because DNA is made of discrete letters, just like text, making it a perfect test case).
- They compared their "One Tokenizer" against the old "Specialist" methods.
- The Result: The "One Tokenizer" was significantly better at reasoning. It didn't just guess; it actually understood the biological logic better because it wasn't distracted by trying to force two different languages to fit together.

The Bottom Line

The paper argues that to make AI truly smart at combining different types of data (like biology and language), we shouldn't try to patch them together at the end. Instead, we should build them into a single, unified system from the very beginning. By giving the AI a single dictionary for everything, we remove the "gap" that slows it down, allowing it to reason deeply and accurately.

In short: Don't build a bridge between two islands; build one big island where everything can live together naturally.

Technical Summary: Mind the Gap No More: Achieving Zero-Gap Multimodal Integration via One Tokenizer

Problem Statement

The development of Multimodal Large Language Models (MLLMs) faces a central challenge: effectively integrating heterogeneous inputs (e.g., continuous visual signals or discrete biological sequences) into a cohesive reasoning engine. Current paradigms predominantly rely on modular architectures, where modality-specific encoders extract representations that are subsequently mapped into the LLM's shared latent space via cross-modal projectors.

The authors identify a fundamental bottleneck in this approach: the modality gap. Because modality-specific encoders are optimized independently, they embed data into isolated sub-manifolds within the latent space. This geometric separation forces the LLM to expend significant computational capacity on "geometric reconciliation" (aligning disjoint spaces) rather than engaging in deep cross-modal reasoning. While recent "plug-and-play" approaches have facilitated scaling, they inherently postpone deep interactions and maintain strict representational boundaries prior to forced alignment.

Methodology: One Tokenizer

To address the geometric limitations of modular designs, the paper proposes One Tokenizer, a native architecture that maps all modalities directly into a shared token space using a unified vocabulary.

Theoretical Framework

The authors formalize the modality gap using a statistical framework based on normalized Euclidean distance (inspired by Cohen's d), defined as the distance between expected representations (centroids) of two modalities normalized by their intra-modality variance.

Lemma 1 (Embedding-level gap): The authors prove that modular architectures inherently suffer from a strictly positive initial gap at the embedding layer due to independent initialization in separate subspaces. In contrast, native architectures employing a unified vocabulary drive the expected initial gap to zero as vocabulary size increases, as all tokens share a single underlying distribution.
Theorem 1 (Layer-wise gap preservation): Assuming Lipschitz continuity of LLM layers, the authors demonstrate that the initial geometric state determines deep representation behavior. The initial gap in modular designs is strictly preserved throughout the forward pass, imposing a persistent computational burden. Conversely, the native architecture's zero-gap initialization guarantees an intrinsically integrated representation space across all hidden layers, eliminating the need for post-hoc geometric alignment.

Empirical Testbed: DNA-Text Modeling

To validate these theoretical claims without the confounding noise of tokenizing continuous data (e.g., images), the authors utilize DNA and text as inherently discrete modalities.

One Tokenizer Implementation: The method extends a pre-trained LLM's vocabulary ( $V_{text}$ ) with a discrete DNA $k$ -mer vocabulary ( $V_{DNA}$ ) to create a unified space $V = V_{text} \cup V_{DNA}$ . Both modalities are mapped directly into the shared latent space via a single embedding layer, requiring zero modality-specific encoders or projection parameters.
Baselines: The study compares One Tokenizer against:
1. Modular Baselines: BioReason and ChatNT, which use projection layers or Q-formers to align genomic features with text.
2. SeqCLIP: An enhanced baseline that applies contrastive learning (InfoNCE loss) on DNA-text pairs to explicitly force semantic alignment before generation.

Key Contributions

Theoretical Guarantee: The paper mathematically proves that native architectures with a unified vocabulary achieve and strictly preserve a zero-gap state across all layers, fundamentally eliminating the geometric bottleneck inherent in modular designs.
Controlled Evaluation: The authors establish a noise-free testbed using discrete modalities (DNA and text) to isolate architectural reasoning capacity from tokenization artifacts associated with continuous data.
Empirical Superiority: Extensive evaluations demonstrate that native integration consistently outperforms modular baselines, including those enhanced with complex contrastive alignment mechanisms.

Experimental Results

The authors evaluated the models on two benchmarks: the Nucleotide Transformer (NT) benchmark (classification) and the KEGG-Derived Biological Reasoning Dataset (complex reasoning).

Modality Gap Analysis: Empirical analysis of latent embeddings across layers confirms the theoretical bounds. Modular architectures exhibit a substantial gap at the initial layer that persists through deep layers, forcing the model to spend depth on geometric reconciliation. One Tokenizer, however, maintains a seamless integration with a significantly reduced initial gap and high cross-modal KNN fractions throughout the network.
Classification Performance:
- On the NT benchmark (18 tasks), One Tokenizer achieved the highest average Matthews Correlation Coefficient (MCC) of 0.62, outperforming modular baselines (BioReason: 0.54, ChatNT: 0.55) and the contrastive SeqCLIP (0.56).
- On the KEGG benchmark (37 disease classes), One Tokenizer achieved a state-of-the-art accuracy of 99.31% and an F1-score of 95.24, significantly outperforming modular baselines which showed inconsistencies between accuracy and F1 scores.
Reasoning Performance: Using an LLM-as-a-judge protocol, One Tokenizer achieved a reasoning score of 10.53, significantly narrowing the gap to the ground-truth upper bound (14.03) compared to modular baselines (e.g., SeqCLIP at ~8.55). Case studies revealed that modular models often suffered from domain-inconsistent reasoning or reliance on superficial associations, whereas One Tokenizer facilitated coherent, step-by-step biological deduction.

Significance and Claims

The paper claims that the "modality gap" is not merely an engineering challenge solvable by better alignment modules, but a fundamental structural limitation of modular architectures. By shifting to a unified-vocabulary paradigm, the authors argue that MLLMs can achieve "seamless integration" within the native latent geometry.

The significance of this work lies in demonstrating that for discrete sequences, native integration is a simpler and more powerful route than building complex alignment modules. The results suggest that the geometric unity of a shared token space allows the LLM to extend its pre-trained zero-shot and logical reasoning capabilities directly to biological sequences, rather than treating them as "foreign" entities requiring translation. The authors conclude that a unified tokenization space provides a fundamentally superior framework for deep biological reasoning.

Limitations: The authors note that their study utilizes discrete symbolic modalities to bypass tokenization noise inherent in continuous signals (like vision). A key future direction is investigating whether this seamless integration persists when handling the vector quantization artifacts of continuous modalities. Additionally, the theoretical bounds characterize the gap evolution but do not yet fully explore latent geometry under massive-scale pre-training regimes.