Skip to content
ResearchAXIVELLE Labs

The questions we are working on.

Research at AXIVELLE is organised around seven themes. Together they describe what it would take to build AI systems that reason well, act reliably and work naturally with people.

This page sets out a research agenda. AXIVELLE has not yet published research results; when it does, they will be linked here with their methods and limitations.

01Research themes

Seven themes, studied together.

The themes overlap by design. Progress in reasoning changes what agents can do; efficiency determines what can be deployed; interaction determines whether any of it is useful.

Theme R1

Model development

How should models be built, adapted and combined for dependable use?

We study training and adaptation methods, beginning with small, well-controlled experiments that can be understood fully before they are scaled.

  • Domain adaptation
  • Training data quality
  • Model composition
Theme R2

AI reasoning

How can a model plan, check and correct its own work?

Reasoning research looks at multi-step problem solving: decomposition, verification, and knowing when an answer is not yet good enough.

  • Planning and decomposition
  • Self-verification
  • Calibrated uncertainty
Theme R3

Multimodal intelligence

What does it take to understand the world across more than language?

We explore how text, images, documents and structured data can be represented and reasoned about together.

  • Document and chart understanding
  • Visual reasoning
  • Cross-modal grounding
Theme R4

Agentic systems

How do agents act usefully while remaining predictable and overseeable?

Agents introduce new questions about memory, tool use, coordination and control. We treat oversight as a design requirement, not an afterthought.

  • Tool use and permissions
  • Multi-agent coordination
  • Human oversight mechanisms
Theme R5

AI efficiency

How much capability can be delivered per unit of compute, time and cost?

Efficiency determines who can use AI and where. We study how to deliver quality with less latency and less computation.

  • Inference efficiency
  • Model routing
  • Cost-aware system design
Theme R6

Intelligent automation

Which kinds of work can be automated reliably, and how do we know?

We examine automation end to end: task selection, exception handling, measurement, and the point at which a person should step in.

  • Process reliability
  • Exception handling
  • Outcome measurement
Theme R7

Human–AI interaction

How should people and AI systems work together?

Interfaces shape what AI can do for people. We study interaction patterns that keep people informed, in control and able to trust what they see.

  • Explanation and transparency
  • Control and steerability
  • Collaborative interfaces
02Method

How research is done here.

  1. STEP 01

    Question

    Every project starts from a question that can be answered, written down before any experiment runs.

  2. STEP 02

    Experiment

    Small, controlled experiments first. Scale is added only when the small case is understood.

  3. STEP 03

    Evaluation

    Results are measured against evaluations designed to reflect real use, including failure cases.

  4. STEP 04

    Iteration

    What was learned — including what did not work — shapes the next question and the next system.

03Standards

What we hold our research to.

Research that will end up inside products people depend on has to be careful. These are the standards AXIVELLE Labs works to.

  • Measure before claiming

    A result is described only as far as the evidence supports it.

  • Report the limits

    Every finding is stated with the conditions under which it holds and where it fails.

  • Reproduce it

    Experiments are tracked so that a result can be run again and checked.

  • Research toward use

    We prioritise questions whose answers change how systems are built.

Research tooling

See the Labs Workspace concept.

An interface concept for tracking experiments, evaluations and datasets.