LeCun urges shift from pixel generation to world models at ECCV 2026

The European Conference on Computer Vision, ECCV, takes place every two years, and in September 2026 it convened in Malmö, Sweden. Viktor Yurchenko, head of Yandex's CV sensor simulation team, brought back a set of notes. We pulled out the points most useful to people building products around generation, multimodality, and autonomous driving.

The numbers and the main debate: pixels versus representations

The conference received 10,473 submissions and accepted 2,834, with just 163 selected for oral presentations (1.6%). Image and video generation was the largest topic by a substantial margin, followed by vision-language with reasoning and 3D. The generative AI hype has yet to fade.

Yann LeCun's keynote offered an ideological counterweight. He noted that AI is already good at writing code and solving math problems, but struggles to act in the physical world: raw video, audio, and sensor signals remain difficult data to handle, while agents need far too many demonstrations. His prescription:

• Stop predicting the next frame. The future is ambiguous at the pixel level, and that level of detail is usually unnecessary for choosing actions.
• Predict in representation space. This is how JEPA works: an encoder encodes the target, a predictor uses the context to predict its representation, and the error is calculated between representations. Unpredictable details are discarded in the process.
• Plan using model-predictive control. A world model simulates the consequences of actions, and the system searches for a sequence that leads to the goal. RL retains a role as a critic or a way to refine the model.
• Build a hierarchy. Higher levels work with abstract changes over long horizons and set subgoals for lower levels.

LeCun's conclusion was blunt: researchers aiming for human-level AI should shift their focus from LLMs and generative architectures to representation learning and world models. This is one participant's position in an ongoing debate, not an established fact. But for anyone choosing an R&D direction, it is a significant signal.

Research you can put to work

Yandex brought three papers, each with practical applications.

The first shows that autoregressive models can handle image classification. The RandAR-based method averages scores across random orderings of visual tokens. On ImageNet-Val, RandAR-XL achieves 81.3% top-1 accuracy, versus 77.2% for DiT-XL and 82.7% for DINOv2-XL. A speedup of up to 25 times was measured against a DiT classifier on a single A100. Under distribution shift, for example on ImageNet-A, RandAR underperformed DINOv2.

The second, DuET, is relevant to anyone working on image editing. If a unified model supports both generation and editing, conditioning on the source image can be temporarily disabled midway through inference, then restored. On FLUX2-Klein 4B, switching at steps [25, 35) out of 50 raises the GEdit quality score from 7.46 to 8.01. The tradeoff is that SSIM on GIE-Bench drops from 0.815 to 0.749, meaning details in the original image change more noticeably. A variant called Selective DuET keeps preservation within a threshold set by a VLM evaluator, but this is neither a human assessment nor a guarantee for every edit.

The third is a rigorous benchmark of eight training-free methods based on Classifier-Free Guidance, using Stable Diffusion 3.5 Medium and FLUX.2 4B Base. The authors complain that, amid the flood of papers, new methods are often compared using weak benchmarks and outdated models. If you are choosing guidance for production, this is a useful reference.

Autonomous driving: the simulator is the bottleneck

The same themes kept coming up at the workshop on reliable autonomy. ML planners are moving toward end-to-end approaches, the industry already uses RL post-training in simulation, and researchers are experimenting with training from scratch and using self-play. This requires a fast simulator with high-quality sensor simulation, which remains an unsolved challenge: even NVIDIA frequently encounters the sim-to-real gap. Logs alone are not enough for a reliable planner; it needs a world model that generates challenging scenarios.

Academic researchers face a barrier when it comes to testing on real vehicles. The authors of PRIX, who achieved SOTA results on several benchmarks, wrote on their final slide that they do not know how the model will behave on an actual car with limited compute.

Motional is tackling rare situations, such as complex roadworks or an object that has fallen into the middle of the road. The company released a dataset of 20,000 scenes, each 20 seconds long (111 hours): cameras, lidar, HD maps, annotations, and reasoning labels. Gemini 3.1 Pro selected challenging scenes using a 10-point scale, after which humans reviewed them. The reasoning comes in three forms: spatial, decision-related, and counterfactual ("what would have happened if we had accelerated instead of braking"). Fine-tuning a VLM on these questions and answers improved its performance.

What to do with this

• If you work on image editing, test the DuET approach: it requires no training, but keep an eye on detail preservation.
• If you build planners for autonomous systems, invest in sensor simulation and collecting long-tail scenes, as well as architecture.
• If you are choosing a research direction, take the "LLMs versus world models" debate into account: for now, this is LeCun's position, not a consensus.

Keep the source's limitations in mind, too: this is a personal report from a Yandex participant, not an independent review, and much of the text recounts claims without independently verifying them.