MILP / CP
Where structure is well understood and constraints dominate.
Research overview
Cognitive optimization:
the research programme.
This page describes what Qbitune is researching and how we intend to evaluate it. It is a statement of a research direction, not a report of results. Where a capability has not been demonstrated, the text says so.
01 / ARCHITECTURE
The Qbitune cognitive
optimization architecture.
Real-world evidence becomes an explicit world model; a reason–abstract–plan engine works over that model; the result is compiled into formal problems, solved on heterogeneous compute, verified, and returned with an explanation. Environment changes re-enter the loop.
02 / RESEARCH HYPOTHESES
What we are trying
to find out.
Each of these is an open question. None is a claim about what Qbitune can already do.
- H1
Automatic formulation
That a system can take an unstructured description of an operational situation and derive a formal decision problem that a solver can accept, without a human writing the model by hand.
- H2
Abstraction that transfers
That world models learned for one planning environment carry useful structure into a related one, rather than having to be rebuilt from nothing.
- H3
Planning under change
That hierarchical plans with explicit contingencies degrade more gracefully under disruption than plans recomputed from scratch.
- H4
Verification as a gate
That formal checks over hard constraints can be applied to machine-generated plans cheaply enough to sit in the loop before execution.
- H5
Routing pays
That selecting among classical, quantum-inspired, hybrid and quantum engines per subproblem beats committing to any single engine.
03 / REASONING · ABSTRACTION · PLANNING
How the three map
onto the architecture.
CapabilityWhere it sitsWhat it produces
ReasoningOver the world model, before and during planningDependencies, conflicts, consequences and the uncertainty attached to each
AbstractionBetween raw evidence and the world modelEntities, goals, constraints and relations compact enough to plan over
PlanningOver the abstracted model, across time horizonsHierarchical plans, contingencies and the subproblems handed to the compiler
04 / FORMAL MODELLING
One model,
several formalisms.
The cognitive layer should not be tied to a single mathematical representation. Which formalism suits a subproblem is itself a decision the system has to make.
QUBO / Ising
Where a problem reduces to binary quadratic form and annealing-style engines apply.
Graph formulations
Where relations between entities carry most of the difficulty.
Custom symbolic
Where domain semantics do not survive translation into a standard form.
05 / TRUSTWORTHINESS
Mechanisms, not
assurances.
Trustworthiness has to be built into the architecture rather than asserted about it. These are the mechanisms the research is designed around.
Explicit constraints
Hard rules stay outside the learned components, where they can be tested.
Plan verification
Candidate plans are checked against those constraints before they are proposed.
Explanation
Which goals, constraints and events drove the decision, in a form a planner can read.
Uncertainty
Represented explicitly, so robust alternatives can be compared rather than assumed.
Provenance
A trace of the data, model state and reasoning behind each plan.
Designed for research into transparent, accountable and human-centred cognitive AI. Qbitune makes no regulatory compliance claim.
06 / ADAPTIVE PLANNING
Plans that survive
contact with reality.
Hierarchy
Decompose long horizons so a local disruption does not invalidate the whole plan.
Contingency
Carry alternatives for the failures that are likely rather than replanning blind.
Continual re-planning
Treat change as the normal case: monitor, revise, verify, repeat.
07 / HYBRID COMPUTE
Independent of
the machine underneath.
Qbitune is not based on the assumption that quantum computing is already superior for every optimization task. The cognitive layer should remain independent from the machine used to search for solutions.
Quantum when it helps. Classical when it wins.
08 / BENCHMARK STRATEGY
If it cannot be measured,
it cannot be trusted.
A proposed evaluation framework. These benchmarks do not exist yet — building them is part of the work.
Reasoning
Logical and constraint validity.
Abstraction
Transfer and model correctness.
Planning
Plan feasibility and objective quality.
Adaptation
Performance after unexpected events.
Trust
Explanation fidelity and robustness.
Compute
Time, energy and computational resources.
We aim to contribute reusable benchmarks, evaluation protocols and interoperable interfaces for trustworthy cognitive planning.
09 / OPEN SCIENCE
FAIR by design,
where it applies.
Research outputs are intended to be Findable, Accessible, Interoperable and Reusable where the nature of the output allows it — with open interfaces for cognitive models, planners and solvers, and reproducible evaluation protocols.
10 / PUBLICATIONS
Scientific
publications.
QUBO-based training for VQAs on quantum annealers
Quantum Machine Intelligence, Springer (2026) · doi:10.1007/s42484-026-00441-1
Reformulates a hard optimization problem — training variational quantum algorithms — into a QUBO representation and solves it through a hybrid quantum workflow with recursive refinement. One concrete example of the capability Qbitune is generalising, not the limit of it.
Read on SpringerLink ↗11 / RESEARCH ROADMAP
From scientific hypothesis
to integrated cognitive system.
A research plan, not a status report. Qbitune claims no current technology readiness level.
- 01
Foundations
Formal representations and optimization research.
- 02
Cognitive prototype
Reasoning, abstraction and planning architecture.
- 03
Trustworthy integration
Verification, explainability and uncertainty.
- 04
Real-world validation
Industrial planning demonstrator.
- Target
Integrated cognitive AI prototype
Validated at TRL4. This is the research target — not current status.
12 / EUROPEAN COLLABORATION
Built to work with
a European research ecosystem.
Qbitune is designed to be composable: to interoperate with other cognitive models, planners and solvers, and to take part in larger multi-system demonstrations rather than standing alone.
Interoperability
Open interfaces for cognitive models, planners and solvers.
Benchmarks
Shared tasks and reproducible evaluation.
Joint pilots
Composable technology for larger multi-system demonstrations.
FAIR research
Outputs designed around FAIR principles where applicable.
Potential synergies include European AI research infrastructures, testing facilities, HPC and quantum initiatives. Qbitune has no formal partnership to announce.
COLLABORATE
Bring us a planning problem that machines still struggle to solve.
We are looking for research collaborators and industrial partners working on complex decision environments with changing goals, constraints and resources.