// Deep-tech · Artificial reasoning
A reasoning engine that learns from almost nothing.
Lindwurm is a novel, neuro-symbolic reasoning system built on a compact set of logical primitives and a self-regulating learning loop — designed to acquire new skills from a handful of examples, not from oceans of data.
// The opportunity
Scaling alone is hitting a wall — the next step is efficiency.
Today's frontier AI improves mostly by getting bigger: more parameters, more data, more compute. Yet on tasks that require genuinely novel reasoning — where memorisation cannot help — even the largest models stay far below human performance. The bottleneck is no longer scale; it is the ability to generalise from very little. Lindwurm attacks exactly this gap, treating sample-efficient, adaptive reasoning as a first-class engineering goal rather than a by-product of size.
A different substrate
Not another wrapper around a frozen large model. A purpose-built engine that keeps learning at run time.
Few-shot by design
Rules are inferred from a few demonstrations, then applied — the way people solve unfamiliar puzzles.
Objectively testable
Built to be measured on independent, public benchmarks so progress is verifiable, not asserted.
// What's different?
Not another model — an engine that develops itself.
Today's familiar systems are fixed at the moment of training: what they know was in the training data. Lindwurm works on a different principle.
The goal is not a bigger model but a generally reasoning system — one that keeps improving its own workings.
// The approach
Logical primitives, wrapped in a self-regulating learning loop.
Lindwurm combines two ingredients that are usually kept apart: a symbolic core and an adaptive, plasticity-driven controller.
1 · A core set of logical primitives
The engine ships with a compact vocabulary of elementary logical and structural operations — its base skill set. Complex behaviour emerges from composing these primitives, not from hard-coding task-specific answers. This keeps the system general: it can be pointed at problems it has never seen.
2 · Homeostatic self-learning
A feedback-driven controller strengthens, weakens, or forgets connections between primitives based on how well a candidate solution fits the evidence. Useful patterns are reinforced; misleading ones are inhibited and fade. This self-regulating (homeostatic) dynamic is what lets Lindwurm improve itself rather than remain frozen.
Because learning happens at inference time from the examples in front of it, Lindwurm does not depend on memorising a giant training corpus. That is the property most current systems lack — and the one that matters for real generalisation.
// How it works
Two examples — and the rule is found.
A simple, ARC-style illustration: from the demonstration pairs the engine derives the rule by itself — reflection across the vertical axis — then applies it to a task it has never seen.
1 · Demonstrations
2 · Inferred rule
- axial reflectionreinforced
- translationinhibited
- colour swapinhibited
3 · New task
Solved after two demonstrations.
Illustration: the rule is not programmed in — the engine derives it from the demonstrations, while the homeostatic loop fades the hypotheses that do not fit.
// Where we are
The engine runs. Now it needs to be stress-tested at full width.
The core reasoning engine is built and operational. The remaining work is a disciplined test-and-tune phase: running Lindwurm across its full breadth, hardening the learning loop, and validating it against external, independent benchmarks so its performance is objective and comparable.
Current position — engine complete, entering full-scale evaluation.
Why compute is the blocker
To run Lindwurm across its full range of behaviour — and to iterate quickly on the learning loop — we need sustained access to high-performance GPU capacity. At small scale the engine works; at full width it needs the hardware to breathe.
Why external test systems matter
Self-reported results are not enough. Independent, public evaluation makes our claims objective and comparable to the state of the art — exactly the credibility a deep-tech reviewer should demand.
Status updates
Public debut of the program: test programme, audit protocol, and milestones published.
// The 16-month plan
From a working engine to an independently benchmarked result.
Bring the engine up to full scale
Secure GPU capacity, port the engine to run across its complete primitive space, and establish the internal measurement harness.
Milestone: the engine runs on the full primitive set; the internal measurement harness logs every run and makes it re-measurable.
Harden the homeostatic controller
Tune reinforcement/inhibition dynamics for stability and sample efficiency; eliminate failure modes; build reproducible evaluation pipelines.
Milestone: a stable, reproducible learning loop — same input, same result; documented sample-efficiency baselines.
Objective, comparable evaluation
Run Lindwurm on external, public reasoning benchmarks (e.g. the ARC-AGI series) to produce third-party-verifiable, comparable performance.
Milestone: a third-party-verifiable result on at least one public reasoning benchmark (the ARC-AGI series).
Package results & prepare to incorporate
Consolidate findings, document the method, and prepare the company formation and follow-on funding (e.g. FFG, EIC) from a position of measured strength.
Milestone: complete methodological documentation; follow-on applications submitted; ready to incorporate.
// Objectivity
Measured against the hardest public yardstick for reasoning.
We deliberately choose benchmarks that resist memorisation and reward genuine generalisation — most notably the ARC-AGI series, where every task is novel and easy for humans yet hard for AI. Success is defined not by a narrative but by a number that others can reproduce.
Sample efficiency
How few examples are needed to acquire a new skill.
Generalisation
Score on held-out, never-seen tasks — no leakage from development sets.
Compute per task
Result quality relative to the energy and hardware spent.
The test programme — three tiers of validation
Public datasets, run by us
Reproducible, documented runs on public reasoning test sets: ConceptARC, ARC-AGI, RAVEN / I-RAVEN / PGM, Bongard-LOGO, KANDY / Kandinsky patterns.
Prediction-upload leaderboards
Where the organiser holds the hidden solutions: we submit predictions only — the scoring happens independently of us.
Test API for independent audit
Controlled access for independent auditors to probe the engine on their own tasks — not on our demo.
// Responsible research
Aligned with Europe's frameworks — by design.
EU AI Act — in full
We develop with the full scope of the EU AI Act in view: risk management, human oversight, logging, transparency.
Reproducibility
Every published result can be verified by an independent run — same inputs, same outcome.
Public yardsticks
Progress is demonstrated on external, public benchmarks — not asserted in our own words.
Documentation first
The method is auditable end to end; the knowledge lives in the documentation, not in any one head.
// Honest risk assessment
What could go wrong — and how we contain it.
Risk: the learning loop is unstable at scale
Mitigation: phased scale-up with continuous internal metrics; the fine-tuning phase is budgeted specifically for this.
Risk: benchmark scores plateau below target
Mitigation: we develop in a closed feedback loop — a low score is not a failure but an input: it shows exactly what to improve, and the next iteration starts from it. The real risk would be having no test results at all, or results that are not honest.
Risk: compute cost overruns
Mitigation: combine funded GPU time with free research compute (e.g. EuroHPC, research cloud programmes) to extend runway.
Risk: key-person dependency
Mitigation: documentation-first development and knowledge shared across the founding team; a defined plan to add further collaborators as the project incorporates.
// Application directions
Most valuable where data is scarce and the stakes are high.
The range of uses is deliberately broad: Lindwurm is a capability base, not a single application — a natural extension of what Lile works on every day.
Anomaly detection from few precedents
A new abuse pattern is dangerous precisely because there is no data on it yet. Sample-efficient reasoning closes that gap — detection even where no training corpus exists.
DetailsScreening and risk scoring
Sanctions and watchlist screening with explainable decisions: the inferred rule chain can be traced end to end — no black box.
DetailsHard problems
Open-ended engineering and research problems where off-the-shelf tools end — the natural continuation of Lile's hard-problems practice.
Details// FAQ
Frequently asked questions
What happens to the intellectual property?
The method, the architecture, and the code are created within the program and documented as they evolve. Public communication is limited to benchmark results and the principles of the method — implementation details remain protected. Upon incorporation, the IP moves into the new company.
Why grant funding rather than venture capital?
At this stage the goal is measurement, not growth — non-dilutive funding exists precisely for that. After an independently verified result, the next stage, with investors where appropriate, starts from a far stronger position. The order is deliberate: evidence first, capital second.
What if benchmark scores plateau below the target?
Even a partial, independently verified result is valuable: it shows where the approach stands and what to change next. The method and the evaluation infrastructure remain either way — the program measures publicly precisely so that the next step rests on facts.
Why isn't the project open source?
The system's distinctive properties — sample-efficient reasoning that keeps learning at run time — are dual-use: they could also serve to build critical or offensive systems, such as self-guided missiles or drones. Until we have worked out the appropriate ethical and safety mechanisms, the code remains strictly closed. Openness is provided elsewhere: through public benchmarks and independent audit.
How can one collaborate — or become an auditor?
There are two routes: research and industry partners can join through the program's application directions, while for independent auditors the public audit protocol sets the framework — we welcome academic groups, accredited test laboratories, and ML evaluation bodies. In both cases, reach out via the contact form.
// The long view
A research thesis worth testing.
Beyond the immediate benchmarks, Lindwurm explores a deeper hypothesis: that robust, adaptive intelligence rests on a self-maintaining, homeostatic core rather than on static, frozen computation. We treat this as a scientific hypothesis to be earned through measurement — not a claim to be assumed. The funded phase is the first rigorous test of it.
A decisive phase — we are looking for partners and auditors.
Sample-efficient reasoning · a working engine · independent, reproducible benchmarks