Main comparison summary preserved directly in static HTML.
Two model-based RL agents using fundamentally different world model architectures: DreamerV3's RSSM with actor-critic vs. IRIS's autoregressive Transformer with VQ-VAE tokens.
Primary editorial conclusion preserved for non-JS crawlers and readers.
DreamerV3 is the stronger performer (2x human vs 1x human on Atari) and far more general across domains. IRIS is conceptually important because it bridges world modeling and language modeling, suggesting future Transformer-scale world models could benefit from LLM advances. DreamerV3 is the practical choice today; IRIS points toward a promising architectural future.
Extractable difference list generated from the comparison table.
Static decision guidance for no-JS readers.
Choose DreamerV3 when its capabilities best match your research or deployment requirements.
Choose IRIS when its capabilities best match your research or deployment requirements.
DreamerV3 is the stronger performer (2x human vs 1x human on Atari) and far more general across domains. IRIS is conceptually important because it bridges world modeling and language modeling, suggesting future Transformer-scale world models could benefit from LLM advances. DreamerV3 is the practical choice today; IRIS points toward a promising architectural future.
| Dimension | DreamerV3 | IRIS |
|---|---|---|
| World Model | RSSM (recurrent state-space) | Autoregressive Transformer |
| Representation | Discrete categorical latents | VQ-VAE discrete tokens |
| Policy Learning | Actor-critic in imagination | Autoregressive next-token prediction |
| Atari 100K | 2.01x human (SOTA) | 1.046x human |
| Domain Breadth | Atari, DMControl, Minecraft, etc. | Atari (primary) |
| Connection To | State-space model tradition | Language modeling (GPT-style) |
| Scalability Path | RSSM refinements | Transformer scaling laws |
High-level scoring context for the models referenced in this comparison.
| Model | Category | Index v1.1 | Confidence |
|---|---|---|---|
| DreamerV3 | Model-Based RL | 88/100 | high |
| IRIS | Model-Based RL | 65/100 | medium |
| DIAMOND | Model-Based RL | 64/100 | medium |
| DreamerV2 | Model-Based RL | 72/100 | high |
FAQ answers rendered directly into static HTML for extractable responses.
IRIS proves that autoregressive next-token prediction (the same approach powering GPT) can work for world models. This opens the door to scaling world models the same way LLMs scale.
Short extractable summary preserved directly in static HTML.
Editorial provenance and refresh policy preserved directly in static HTML.
Published by world-models.io editorial board.
Lead editor Bernard Grenat.
This comparison page publishes a direct answer, explicit trade-offs, and source-backed evidence that can be validated against primary materials.
Each editorial page is assembled from primary sources, normalized into extractable summaries, checked for factual drift, and reviewed before publication or major refreshes. Last reviewed: 2026-06-21.
Pages are refreshed when a new paper, benchmark, release, architecture update, or stronger primary source materially changes the answer a reader or AI system should retrieve.
Each page links back to relevant primary sources and keeps a stable canonical URL so readers can verify claims, trace context, and reference the most up-to-date version. See the editorial policy.
Primary papers and official sources for the models discussed on this comparison page.