LAMA-AI 2026: First International Online Forum on Language, Mathematics, and Artificial Intelligence Online November 21-22, 2026

------------------------------------------------------------------------------------------------------------------------------

Just one thing: don't follow the existing research. Think for yourself. That's very important.

Susumu Kitagawa, Nobel Prize in Chemistry 2025

------------------------------------------------------------------------------------------------------------------------------

What is “a house” or “(a) language”? And how do humans and machines perceive, construct, process, and evaluate them? A hint.

First Call for Papers

PDF of the call here.

Following the international online workshop series Linguistics Meets ChatGPT: From Prompt to Theory (organized by MANOVA AI / Gauss:AI Global), we are happy to announce the First International Online Forum on Language, Mathematics, and Artificial Intelligence (LAMA-AI).

We start from the following premises:

• Language is at the core of Large Language Models (LLMs).
• LLMs are at the heart of AI systems that integrate language, image, and audio processing.
• AI models are implemented via programming languages and processed by computers in binary code.

These factors make linguistics and mathematics instrumental for AI. It should be noted that both words and numerals are tokenized before entering an LLM; in both cases, tokens are pieces of form detached from traditional semantics. For example, the tokenizer may segment the year “2026” into the tokens “202” and “6”; similarly, the word metalanguage may be segmented as “metal” and “anguage.” As one might imagine, tokenization complicates the identification of structural units traditionally recognized in both mathematics and linguistics — such as “26” in 2026 or “language” in metalanguage. (Examples and basic information on tokenization are available at https://platform.openai.com/tokenizer; the examples above were checked for GPT-5.x & O1/3.) Nevertheless, LLMs produce fluent language and solve complex mathematical problems. Although LLMs may make mistakes, their performance is sufficiently convincing that many perceive them as capable of reasoning.

Reactions to LLMs within linguistics and mathematics have differed.

• Mathematicians have been largely cooperative; leading figures, such as Terence Tao, have even spoken of a fundamental reassessment of intelligence, and their involvement has helped LLMs achieve success in solving International Mathematical Olympiad problems.
• Reactions within linguistics have often been more cautious, not to say skeptical. Many scholars have emphasized genuine limitations of LLMs, including data transparency, environmental impact, and high error rates specifically in the context of traditional linguistic analyses. These concerns are legitimate. However, much of the linguistic discussion still relies on familiar categories such as words, morphemes, syntax (trees), dependencies, meaning, and grammar, without sufficiently considering the model-specific units and operations through which LLMs generate language. Not to mention that LLMs do not “see” language, strictly speaking, but operate only with numerical representations of tokens.

An additional source of confusion is terminology: in linguistics, “token” is commonly understood as an occurrence of a linguistic unit, often a word, whereas in LLMs tokens are computational pieces of form produced by a tokenizer. Tokens need not correspond to words, morphemes, or other units of linguistic analysis, and they are not associated with stable meanings in the traditional linguistic sense. Meaning-like effects in LLMs arise from the combinability of units in large sequences and from their distributional relations in high-dimensional vector space. This is related to Distributional Semantics, but it is not the same. These mismatches seem to obscure the fact that LLMs generate fluent language without internally representing language in the way linguistic theory usually assumes.

Aim of the Forum

The aim of this conference is fourfold:

1. To shed light on LLMs’ internal organization — tokenization, embeddings, vectors, transformer architecture, and high-dimensional vector spaces — and their implications for our respective fields of inquiry.
2. To bring together linguists, math ematicians, cognitive scientists, and other researchers interested in AI in a stimulating atmosphere for the exchange of ideas on applying LLMs to scientific work.
3. To identify limitations of LLMs and propose qualified, field-specific solutions that can be implemented computationally to advance current AI models.
4. To foster a genuine, unfragmented dialogue by running LAMA-AI as a single, sequential session with no parallel tracks.

Topics

We welcome presentations on the following and related topics:

1. LLMs’ internal organization versus classical assumptions in linguistics and mathematics
2. Data collection and resource creation: Do scientists need to curate data, including creation of corpora, if LLMs are fed massive datasets? Can an LLM serve as a reliable repository for scientific data?
3. Methodological advantages and challenges in using LLMs for scientific research
4. Output-based evaluation: How reliable are benchmarks?
5. Language processing in humans and machines
6. Tokenized data as a source for scientific investigation
7. Alternative unit groupings in linguistics, mathematics, and LLM tokenization: bracketing paradoxes (e.g. un+happier vs. unhappy+er), and place-value regrouping (e.g. 1800 as “1 thousand + 8 hundreds” vs. “18 hundreds”); and LLM tokenization examples discussed in the CFP
8. Big LLMs versus baby LLMs: Do we need large amounts of data?
9. Improvement scenarios for LLM efficiency
10. AI and the definition/meaning of basic units within specific scientific fields
11. AI-driven solutions to field-specific scientific tasks
12. AGI and the cross-modal processing of language, image, and audio (given that text corresponds to written language, images to meaning, and audio to spoken language, to some extent)

Invited Speakers

TBA

Submission Guidlines

For 20-minute talks (followed by 10 minutes of discussion), please submit abstracts of up to 500 words, excluding examples and references, via EasyChair by August 15, 2026. Abstract submission opens on June 15, 2026. Submission link: https://easychair.org/conferences/?conf=lamaai2026

Important Dates

1. Notifications of acceptance: September 7, 2026.
2. Early-bird registration: September 8–30, 2026.
3. Regular registration: October 1–November 19, 2026.
4. Program schedule: early October 2026.
5. Conference dates: November 21–22, 2026 (online).

Certificates

Certificates will be issued for both presenting and non-presenting participants.

Proceedings

We plan to publish the forum proceedings. Details TBA.

Important Links

LAMA-AI Forum: https://gaussaiglobal.com/lama-ai-forum

MANOVA AI: https://www.manova-ai.com

Gauss:AI Global: https://gaussaiglobal.com

LinkedIn: https://www.linkedin.com/company/manova-ai

Facebook: https://www.facebook.com/GaussAIGlobal

All relevant information and updates will be posted on these sites.

Organizing committee

Andela Miladinovic, University of Belgrade, Serbia

Despoina Bikou, Aristotle University of Thessaloniki, Greece

Philippa Adolf, University of Vienna, Austria

Stela Manova, MANOVA AI / Gauss:AI Global, Vienna, Austria (chair)

Yolanda Xavier, NOVA University Lisbon, Portugal