Generation efficiency is another pillar of interactive WMs for AD. Lots of cool works appearing in this area, including the Parallel Decoding Distillation (PDD) by Arash Vahdat & team #ssad2026

Popular strategies to ensure stability over long rollouts. Some of them are starting to make their way into pre-training #ssad2026

One of the major challenges for enabling true interactivity is finding a way to deal with the many different frequencies of things you want to put in your state.
Every 10Hz we want to step our WM with new actions that in turn update the state #ssad2026

The challenge with interventions in your environment: the state and actions might be correlated #ssad2026

(2) how to enable and integrate interactivity in the state space
(3) how to represent the action space #ssad2026

Key considerations when building a world model for AD:
(1) reconstructing pixels faithfully to run models inside but without over-focusing on high frequencies items in your state representations #ssad2026

A zoo of ways to craft your state representation with focus on semantics, geometry, task relevance/predictability, or rich info from DINO distillation #ssad2026

What should the tokenizer encode in this state. This is where methods can differ: frozen visual autoencoder, frozen representation encoder, joint encoder + predictor and no decoder #ssad2026

The architecture of much of the world models nowadays:
- a tokenizer to compress and produce your state
- a world model module that consumes states and conditions and produce next states #ssad2026

World models were around since a while and made a comeback in late 2010s.
Kudos to Alex Kendall and Wayve for laying out this vision to the world early on and this technical bet on how things will evolve #ssad2026

Quick check on recent (hyper-)popularity of world models #ssad2026

Answering interventions requires world models.
Answering counterfactuals requires controllable world models. #ssad2026

Autonomous Driving needs more than predictive models. World models go from associations from what usually happens to interventions (how?) and then to counterfactuals (why?) #ssad2026

Kicking off day 3 of #ssad2026 with Sofía Dudas talking about World Models for Autonomous Driving

Fireside chat with today’s speakers: @kashyap7x.bsky.social, German Ros, Fatma Güney, hosted by Gabriel Villalonga #ssad2026

Exciting areas outlined by Fatma:
- evolving 3D-aware states
- real world failures with the help of simulation #ssad2026

Taking on better dedicated reward models for driving: crash, near crash, suboptimal driving #ssad2026

Were these reliable? Not quite.
Reliability is a different beast and it can take different forms across tasks: semantic segmentation, 3D reconstruction, HD map prediction #ssad2026

Looking now at LLMs and think-ahead. We can use a large teacher to supervise the future actions of a small student. If you squint a bit you can see an early form of on-policy distillation for driving #ssad2026

Moving on from low to mid-level representations with slots. They work well but progress in other areas, eg SAM2 with implicit tracking, made them eventually look at other problems. That’s research: looking at things that don’t work yet #ssad2026