At ICLR's second Workshop on World Models on 27 April 2026, Jürgen Schmidhuber used a 37-minute invited talk to make a pointed case: a useful world model should do more than replay the next sensory step. For real-world planning, it should expose abstractions that a controller can query and combine around the task in front of it.
Jürgen Schmidhuber Watch analysis
What happened
Schmidhuber began with the familiar split between a controller that chooses actions and a model that predicts what follows, including future observations and rewards. In his historical account, curiosity appears when the controller seeks experiences that still surprise the predictor, helping the model learn parts of the environment it has not mastered.
His sharper proposal was to treat the controller as a learned prompt engineer for the model's hidden state. Rather than repeatedly rolling an imagined world forward one tiny action at a time, the controller sends learned signals into useful internal representations, reads what comes back and adapts its policy. Those signals need not be English; they can be vectors, images or another task-shaped language learned by the system itself.
Why it matters
Schmidhuber offered a practical test: does a controller learn a new task faster by consulting the world model than it would by ignoring the model and starting from scratch? That makes transfer the point, not the mere existence of a compressed representation.
He argued that detailed rollouts can be workable in bounded games, but real environments demand reusable chunks and high-level junctions. His travel example plans around taxi, airport and flight rather than every finger movement. The claim is straightforward even if the machinery is not: intelligent planning needs the right level of abstraction at the right moment.
Our read
World models are often sold as dream machines with a suspicious amount of fog around the wiring. This talk gives them a cleaner engineering exam: can the system retrieve the useful compact bit of experience when the situation changes? A model that remembers everything but cannot find the right spanner is just a very expensive shed with no labels.
What to watch
- Benchmarks showing that learned model queries improve transfer to genuinely new tasks.
- Methods that discover useful planning hierarchies without hand-written subgoals.
- Evidence that abstract planning remains reliable when the physical environment changes.
- Clear separation between Schmidhuber's historical lineage claims and independently established consensus.
Discussion spark: Which early world-model ideas have genuinely survived modern scaling, and which similarities look persuasive only when viewed backwards?
Sources and evidence
- Invited Talk #1: Professor Jürgen Schmidhuber (IDSIA & KAUST) (27 April 2026)
- The 2nd Workshop on World Models: Understanding, Modelling and Scaling (27 April 2026)
- Invited Talk: Professor Jürgen Schmidhuber (27 April 2026)
Independent WittyWires tracker for public updates about Jürgen Schmidhuber. Not affiliated with or endorsed by Jürgen Schmidhuber; this is not an official account.