Advanced Resilience & Computing Mission
Machines that never stop learning.
ARC-M is a research initiative at TUM building a new architecture for continual learning: finite-sized state models that keep learning from experience after deployment, online and in the loop — built for the data that doesn't exist yet, but is being generated every second.
The transformer ate the internet.
The best models are now trained on essentially all the high-quality text that exists. Scaling needs more data, not just more compute — and the stock of human text is running out this decade. The frontier is squeezing the last few percent out of a saturated architecture.
The next frontier isn't more internet. It's the embodied, operational stream — every robot on a line, every agent doing a task, generating millions of data points a day. Almost none of it is captured: a frozen, context-bounded model throws it away the moment it happens. The bottleneck is architecture, not supply.
Our bet: finite-sized state models
A genuinely new architecture, not a transformer with memory bolted on. The model is the learning system. These are the principles we build around.
Finite state
A bounded internal state, not an ever-growing context window. Cost stays constant as experience accumulates.
Continual learning
The state updates from experience, in the loop and online — the forward pass becomes a learning step.
No catastrophic forgetting
Retaining old skills while acquiring new ones — the unsolved core the architecture is built to address.
Data that compounds
Every robot, every shift, every correction makes the system better. Data compounds instead of resetting to zero.
Embodied stream
Learning from the operational data frozen models can't reach, rather than from a static archive of the past.
Falsifiable first
A bounded-compute benchmark that beats long-context and periodic-fine-tuning baselines. Concrete and fast to test.
Why now
The data wall
Transformer scaling is visibly saturating, and the field is openly debating it. The old paradigm is exhausting its only fuel.
Embodied capability
Robots and agents finally do enough real work to generate the stream. The new data source just came online.
Read the argument in full: The Data Wall and Finite-State Models for Continual Learning.
The architecture every robot and agent ships with once the transformer era ends. We're a research initiative betting on what comes next — and building the benchmark to prove it.