About Our Research

Understanding our mission, our collaborative network, and the organizations supporting our vision.

Research Direction

"We aim to build intelligent and versatile systems that address high-impact societal needs across language boundaries, multimodal contexts, and specialized domains."

Omni Language AI Research is a dedicated research collective operating at the intersection of deep learning, linguistics, and multimodal intelligence. By blending rigorous academic methodology with modern engineering, we create systems that don't just process text but comprehend context, reason over complex modalities, and adapt dynamically to new domains.

Our work runs along three directions. They are not separate programmes: they sit on one shared methodological base — data and corpora, continual learning, memory, reinforcement learning, and evaluation — and they feed one another, so a result in one direction usually changes what is possible in the others.

How our three research directions connect A shared foundation of data and corpora, continual learning, memory systems, reinforcement learning, and evaluation supports three directions: natural language processing, multimodal learning, and health intelligence. The directions exchange methods with one another and together produce equitable language access, trustworthy clinical decision support, and reproducible evaluation. SOCIETAL OUTCOMES Equitable language access · Trustworthy clinical decision support · Reproducible evaluation Natural Language Processing Multilingual NLP Continual learning Memory systems RL & test-time scaling Multimodal Learning Vision–language models Grounded reasoning Multimodal steering Benchmarking & evaluation Health Intelligence RL for care pathways Mental health support Multimodal pathology Clinical safety reasoning & representations multimodal patient data SHARED FOUNDATIONS Data & corpora Continual learning Memory systems Reinforcement learning Evaluation & benchmarks

Scroll the diagram horizontally to see it in full.

One base, three directions, shared outcomes. Methods developed for one direction — multilingual adaptation, grounded reasoning, safe policy learning — are reused in the others; the horizontal arrows mark where that exchange is most active.
Natural Language Processing Multimodal Learning Health Intelligence

Natural Language Processing

From many languages to adapted models to test-time reasoning Many languages are mapped into a shared representation space; a base language model is adapted by multilingual continual pre-training and reinforcement learning while retaining earlier knowledge in memory; at inference, extra compute explores and verifies candidate reasoning paths before an output is selected. MULTIPLE LANGUAGES eng zho fin swa ara + many more shared space CONTINUAL ADAPTATION Base language model + Multilingual pre-training + RL alignment retained memory TEST-TIME REASONING query path A path B path C verified output explore · verify · select

Scroll the diagram horizontally to see it in full.

Languages are mapped into one representation space, the model is adapted in stages without discarding what it already knows, and extra compute at inference buys verified reasoning rather than a single guess.

Illustration for natural language processing Large Language Models (LLMs) have demonstrated impressive capabilities, but scaling them globally requires addressing representation, alignment, and learning efficiency. We focus on bridging these challenges through:

  • Massively Multilingual NLP: Scaling and adapting systems to support a few hundred languages, ensuring equitable access to state-of-the-art AI for underrepresented communities.
  • Continual Learning: Investigating training paradigms and architectures that enable models to continuously learn new languages and tasks over time without catastrophic forgetting.
  • Memory Systems: Developing memory-augmented architectures and retention mechanisms to enable models to effectively store and retrieve long-term context.
  • Reinforcement Learning: Applying reinforcement learning techniques to align models across diverse languages and optimize translation and reasoning performance.
  • Test-Time Scaling: Investigating how allocating additional compute during inference (e.g., via search and verification traces) affects translation quality and reasoning performance across diverse languages.
Explore Multilingual NLP Related publications

Multimodal Learning

Vision and language encoders, grounded reasoning, and evaluation Images, audio, and text are encoded into a joint embedding space; a reasoning chain locates, compares, and infers over specific evidence regions to produce a grounded answer; the result is scored along evaluation axes including factuality, robustness and safety, and visual grounding. VISION · AUDIO · LANGUAGE image audio text joint space GROUNDED REASONING evidence regions 1 · locate 2 · compare 3 · infer grounded answer EVALUATION EVALUATION AXES factuality robustness & safety visual grounding benchmarks · metrics · reporting

Scroll the diagram horizontally to see it in full.

Vision, audio, and language meet in a joint space; reasoning stays tied to the evidence it points at; and every claim is checked against the same evaluation axes.

Illustration for multimodal learning Real-world applications require models to perceive, reason, and act across different data modalities. Our research in multimodal AI concentrates on:

  • Vision-Language Models: Developing unified architectures that seamlessly integrate visual and textual information to understand complex scenes, documents, and diagrams.
  • Multimodal Reasoning: Enabling systems to perform multi-step reasoning, cross-modal retrieval, and logical deduction using joint visual and textual contexts.
  • Evaluation and Benchmarking: Establishing rigorous benchmarks and metrics to assess the safety, robustness, and factual consistency of multimodal models in real-world environments.
Related publications

Health Intelligence

Multimodal patient records, safety-bounded treatment policies, and mental health dialogue Histopathology slides, clinical notes, electronic health records, physiological time series, and genomics are combined into a joint space; a reinforcement learning loop maps patient state to a policy, an action, and an observed outcome, all inside explicit clinical safety constraints; conversational systems detect, support, and monitor affect across sessions. PATIENT RECORD histopathology clinical notes EHR time series genomics joint space TREATMENT POLICY state policy action outcome bounded by clinical safety constraints learned from historical care pathways MENTAL HEALTH DIALOGUE detect · support · monitor affect trajectory across sessions

Scroll the diagram horizontally to see it in full.

Evidence from several modalities is fused per patient, policies are learned only inside explicit safety constraints, and language interfaces track well-being over time rather than at a single point.

Illustration for health intelligence AI has the potential to transform healthcare, but clinical environments demand extreme robustness, explainability, and multimodal integration. Our research focuses on:

  • Reinforcement Learning for Care: Designing RL algorithms that learn optimal, personalized treatment paths from historical patient data while respecting medical safety boundaries.
  • Mental Health and Well-being: Leveraging natural language interfaces and domain-specific LLMs to detect emotional changes and support mental health monitoring and psychotherapy.
  • Multimodal Pathology Models: Integrating gigapixel histopathology images with text and genomics to build diagnostic and prognostic models.
FINe-Health Foundry AI for Health AI for Mental Health NLP for Health (Archived)

The work behind these directions

Journal articles, conference proceedings, and preprints, searchable by title, author, venue, and year.

Browse publications

Collaborators

We work closely with leading research groups and academic institutions globally.

ELLIS Institute Finland

ELLIS Institute Finland
TurkuNLP

TurkuNLP
Aalto University

Aalto University
University of Helsinki

University of Helsinki
Helsinki-NLP

Helsinki-NLP
LMU Munich

LMU Munich
University of Manchester

University of Manchester

Funders & Support

We are grateful to the following institutions and funding bodies for supporting our research agenda and enabling scientific progress.

ELLIS Institute Finland

Academic funding and fostering machine learning networks.

University of Turku

Institutional support and research infrastructure.

Foundation PS

Supporting AI research excellence and competitiveness.

EU Horizon Europe / MSCA

Supporting doctoral/postdoctoral training and researcher mobility via the HAIF program.

Microsoft AI for Good Lab

Research collaboration and compute support under the LINGUA program.

Business Finland

Funding research, innovation, and technology partnerships.