Research

Studying how language models should be built

DEIA Solutions runs an independent research program alongside its consulting practice. The through-line: language models reason better when structure is explicit — in the language they learn from, in the order they learn it, and in the representations they act on.

Basque-First LLM Pretraining

Most large language models are pretrained English-first. English conveys grammatical roles largely by word order, which lets a model get far on positional inference without ever building explicit structural representations. Basque is different: it is morphologically rich and explicitly case-marked — grammatical roles are written directly into the words themselves.

Our research investigates whether pretraining on Basque first produces models with stronger predicate-argument structure and morphological reasoning — models that know who did what to whom because the training language forced them to encode it, not infer it.

PRISM-IR

PRISM — Process Representation, Intent Simulation & Manifestation. PRISM-IR is its intermediate representation: an open, domain-agnostic format for describing any process, whether simulated, executed, or hybrid.

It is the contract between human intent and machine execution: humans write plain English, LLMs compile it to PRISM-IR, runtimes execute it, and any LLM can reconstruct readable English from it on demand. The specification covers all 43 van der Aalst workflow patterns and makes intention a first-class construct — every process node carries an explicit verb-plus-object intention rather than leaving intent implicit in prose.

Training-Order Experiments

A companion line of work asks how the order of training data shapes what a model learns — including experiments with training language models in reverse order. If curriculum matters as much as corpus, the sequencing of pretraining data is a lever the field has barely begun to use.

Why this matters for our consulting work

The same principle that drives the research — make structure explicit, don't leave intent implicit — drives how we build customer-service agents. Agents built on explicit process representations are easier to audit, easier to improve, and fail more gracefully than prompt-only builds. Research and practice feed each other.