Cutting-Edge 2025-2026 Papers & Developments
World Models, VLAs, and End-to-End Driving — What's Happening Right Now
1. Major Conference Papers (Late 2025 — Early 2026)
1.1 CVPR 2025 (June 2025)
| Paper | Key Contribution | Impact |
|---|---|---|
| DiffusionDrive (Highlight) | Truncated diffusion for real-time E2E planning, 10x faster than vanilla diffusion | First real-time diffusion planner |
| SplatAD | Joint camera + LiDAR 3D Gaussian Splatting for AD simulation | First 3DGS with multi-modal sensor sim |
| Data Scaling Laws for E2E AD | Log-linear gains from 16 to 8,192 hours of driving data | Validates scaling for driving |
| DriveTransformer | Unified transformer for all driving tasks | End-to-end architecture simplification |
| SparseDrive | 0.06% collision rate, 7.3 FPS, fully sparse E2E driving | Current SOTA on nuScenes planning |
1.2 NeurIPS 2025 (December 2025)
| Paper | Key Contribution | Impact |
|---|---|---|
| WorldModelBench | First benchmark evaluating video generators AS world models | Standardizes world model evaluation |
| Vista | Generalizable driving world model with multi-level controllability | 55% FID improvement, NeurIPS 2024 but widely adopted in 2025 |
| AutoVLA | Autonomous driving VLA with online RL fine-tuning | Bridge between VLAs and RL |
1.3 ICLR 2025-2026
| Paper | Key Contribution | Impact |
|---|---|---|
| GS-LiDAR (ICLR 2025) | Panoramic 2D Gaussians for LiDAR simulation | LiDAR-native 3DGS |
| AdaWM (ICLR 2025) | Adaptive world model alignment for distribution shift | 2x success rate over DreamerV3 |
| Copilot4D (ICLR 2024, continued impact) | Discrete diffusion on LiDAR tokens | >65% Chamfer distance reduction at 1s; >50% at 3s |
1.4 AAAI 2025-2026
| Paper | Key Contribution | Impact |
|---|---|---|
| WorldRFT | RL fine-tuning of world models with GRPO | 83% collision reduction on nuScenes |
| AD-L-JEPA | First JEPA for driving LiDAR pre-training | 1.9-2.7x GPU-hour reduction vs MAE |
| Drive-OccWorld (AAAI 2025) | Action-conditioned 4D occupancy prediction | 33% improvement over UniAD |
| DrivingGPT | Unified driving language (interleaved image+action tokens) | PDMS 82.4% on NAVSIM navmini with front-camera VAE tokens and no ego status |
1.5 CVPR 2026 / ICML 2026 (Expected)
| Paper/System | What's Known | Status |
|---|---|---|
| WorldLens | Full-spectrum evaluation of driving world models in real-world settings, 24 dimensions across 5 axes | Accepted CVPR 2026 |
| Epona (ICCV 2025) | Autoregressive + diffusion hybrid, minute-long drift-free generation | Open code expected |
| Dreamer V4 | Block-causal transformer world model, diamonds in Minecraft from offline data | May 2025 preprint, expected venue 2026 |
2. Industry Developments (2025-2026)
2.1 NVIDIA (January 2026 — CES)
- Alpamayo-R1-10B released: 10B VLA (8.2B Cosmos-Reason backbone + 2.3B action expert) with Chain-of-Causation reasoning
- Alpamayo 1.5: RL post-training via GRPO, text-guided planning, flexible multi-camera
- AlpaSim: Open-source microservice closed-loop simulator
- Physical AI dataset: 1,727 hours across 25 countries, 700K reasoning traces (v1), 3M traces (v1.5)
- Partners: Lucid, JLR, Uber, Berkeley DeepDrive
- Key detail: Model weights are non-commercial (research/eval only). Designed as teacher for distillation.
2.2 Wayve
- GAIA-3: 15B parameters, 10x data of GAIA-2, 9 countries
- Wayve AV2.0: Single model navigating 500+ cities, no fine-tuning needed
- US expansion: Adapting to US roads with just 500 hours of incremental data
- Uber partnership: L4 trials in London planned for 2026
- Valuation: $8.6B
2.3 Waymo
- EMMA: End-to-end multimodal model built on Gemini, chain-of-thought reasoning
- Waymo World Model: Built on DeepMind Genie 3, generates camera AND LiDAR data
- Sim Agents: Scenario generation for testing
- 200M+ autonomous miles in production
- Scaling laws paper (June 2025): Validated power-law improvements for motion planning
2.4 Tesla
- FSD V13: 48 neural nets, fully end-to-end from V12+
- Temporal-Voxel Transformers: Latest architecture evolution
- 3B+ FSD miles: Largest real-world E2E driving dataset
- Dojo 2: Restarted, targeting 100K H100-equivalent scale in 2026
2.5 comma.ai
- openpilot 0.11 (March 2026): First driving model fully trained in learned simulation
- DiT world model: 2B parameters (500M/1B variants also available)
- CVPR 2025 paper: On-policy learning from world model outperforms reprojective sim (52.49% vs 48.10% distance engagement)
- 325+ car models, 100M+ miles, MIT licensed
2.6 Google DeepMind
- Genie 2/3: Interactive world models generating photorealistic environments at 24fps/720p
- Scaling to driving: Waymo World Model uses Genie 3 backbone
- V-JEPA 2 (Meta, not DeepMind): 1B params, ~15x faster planning than video-generation baselines (16s vs ~4min)
2.7 Physical Intelligence
- pi0.5: Updated foundation model for robotics
- Flow matching action head: More stable than diffusion for action generation
- Cross-embodiment: Transfers across robot types — driving is the next target
3. Emerging Paradigms
3.1 World Models as Simulators (The "Learned Sim" Paradigm)
The biggest paradigm shift: world models ARE the simulator. Instead of building a physics engine, you train a model that generates realistic sensor data given actions.
| System | Approach | Status |
|---|---|---|
| comma.ai openpilot | Train driving policy entirely in DiT world model | Production (March 2026) |
| Waymo World Model | Genie 3 generates camera+LiDAR for testing | Research+ |
| NVIDIA Cosmos | Foundation model for world simulation | Available |
| Wayve GAIA | Generative world model as simulator | Internal use |
Why this matters for airside: You don't need to build a CARLA-like airport simulator. Train a world model on your bags, then train your planner inside the world model. comma.ai proved this works.
3.2 VLAs Replacing Modular Stacks
The VLA paradigm (single model: sensor input → language reasoning → action output) is replacing modular perception-prediction-planning pipelines.
| Stage | 2024 | 2025-2026 |
|---|---|---|
| Perception | Separate detector + tracker | VLA backbone handles it |
| Prediction | Separate motion forecaster | World model predicts implicitly |
| Planning | Separate trajectory optimizer | VLA directly outputs trajectories |
| Explanation | None | Chain-of-Causation reasoning |
3.3 RL Post-Training for Driving
Following the ChatGPT playbook: pre-train → SFT → RLHF. For driving: pre-train world model → supervised fine-tune on driving data → RL post-train in world model imagination.
| System | RL Method | Result |
|---|---|---|
| Alpamayo 1.5 | GRPO (3 reward signals) | AlpaSim 0.73→0.81, minADE 1.22→1.11m |
| WorldRFT | GRPO in latent world model | 83% collision reduction |
| SafeDreamer | Lagrangian-constrained RL | 94.3% cost reduction |
| Think2Drive | MBRL in latent space | Expert-level CARLA in 3 days |
3.4 Tokenize Everything
The field is converging on: tokenize all modalities → predict next token → scale.
Images → VQ-VAE/FSQ tokens → Transformer predicts next image tokens
LiDAR → Pillar/voxel tokens → Transformer predicts next LiDAR tokens
Actions → Discretized bins → Transformer predicts next action tokens
Language → BPE tokens → Transformer predicts next language tokens
Occupancy → VQ-VAE tokens → Transformer predicts next occupancy tokens
ALL IN ONE MODEL: DrivingGPT, GAIA-1, Alpamayo4. Current SOTA Leaderboards (March 2026)
4.1 nuScenes Planning
| Method | Collision Rate | L2 Error | Year | Note |
|---|---|---|---|---|
| SparseDrive | 0.06% | 1.55m | 2024 | Prior planning baseline |
| DiffusionDrive | 0.16% @3s / 0.08% avg | 0.90m @3s / 0.57m avg | 2025 | Official nuScenes log in project README |
| VADv2 | 0.12% | 1.62m | 2024 | Prior planning baseline |
| UniAD | 0.31% | 1.65m | 2023 | Prior planning baseline |
4.2 NAVSIM
| Method | Metric | Benchmark / split | Score | Year | Note |
|---|---|---|---|---|---|
| Latent-WAM | EPDMS | NAVSIM v2 | 89.3% | 2026 | Preprint; not directly comparable to PDMS |
| DiffusionDrive | PDMS | NAVSIM | 88.1% | 2025 | Official project README reports 45 FPS |
| DrivingGPT | PDMS | NAVSIM mini | 82.4% | 2025 | Mini split; do not compare as full NAVSIM |
| GTRS-E | EPDMS | NAVSIM v2 / Navhard | 49.4% | 2025 | NAVSIM v2 challenge metric |
PDMS and EPDMS are different NAVSIM metrics, and the reported splits above are not interchangeable. Treat the table as a source-backed tracker rather than a single leaderboard.
4.3 nuScenes Occupancy Prediction
| Rank | Method | mIoU | Year |
|---|---|---|---|
| 1 | GaussianFormer-2 | 44.2% | 2025 |
| 2 | FB-OCC | 43.5% | 2024 |
| 3 | FlashOcc | 42.1% | 2024 |
| 4 | SurroundOcc | 40.7% | 2023 |
4.4 World Model Video Generation (nuScenes)
| Rank | Method | FID | FVD | Year |
|---|---|---|---|---|
| 1 | DriveDreamer-2 | 11.2 | 55.7 | 2025 |
| 2 | DrivingGPT | 12.78 | 142.6 | 2025 |
| 3 | Drive-WM | 15.8 | 122.7 | 2024 |
5. What's Ahead of the Curve
5.1 Predictions for 2026-2027
World model simulators go mainstream — comma.ai proved it; expect Wayve, NVIDIA, and others to ship products trained primarily in learned simulators.
VLA distillation pipelines mature — Alpamayo is designed as a teacher. Expect standardized distillation → edge deployment pipelines for 100M-500M parameter student models on Orin/Thor.
Multi-modal world models — Models that jointly predict camera + LiDAR + radar futures (Waymo World Model already does camera + LiDAR). This enables full sensor simulation.
Airside as a proving ground — Low speed, structured environment, high economic value = ideal for early world model deployment. Expect 2-3 more companies to enter airside AV with learned approaches.
Regulatory frameworks catch up — ISO/PAS 8800 (Dec 2024) is the first standard addressing AI safety lifecycle. EASA AI Roadmap 2.0 targeting 2028 certification frameworks. FAA likely follows.
Open-source world models reach production quality — OpenDWM, Cosmos, OccWorld ecosystem maturing rapidly. The gap between research and deployable code is shrinking.
5.2 Research Gaps Worth Pursuing
| Gap | Why It Matters | Who Should Work On It |
|---|---|---|
| Airside driving dataset | Zero public datasets exist | You — publish a subset, become the benchmark |
| LiDAR-native world models | Most work is camera-centric | Copilot4D direction, but needs more work |
| Multi-agent world models for GSE | No airside coordination models | Combine with Moonware-style turnaround prediction |
| Safety certification for world models | No proven methodology exists | Build AMLAS + UL 4600 safety case, publish it |
| Jet blast as learned hazard | Only lookup tables today | Learn residuals from CFD + real data |