Static HTML snapshot of the model record for crawlers and no-JS readers.
| Attribute | Value |
|---|---|
| Model | IRIS |
| Lab / Organization | MSR |
| Category | Model-Based RL |
| Subtype | Autoregressive World Model |
| World Model Type | Discrete token-based dynamics model |
| Primary Domain | Atari games |
| Architecture | VQ-VAE tokenizer + autoregressive Transformer dynamics model |
| Modality | Visual |
| Training Method | Discrete tokenization + autoregressive next-token prediction over (observation, action, reward) sequences |
| Status | active |
| Year | 2023 |
| Performance Index | 65/100 (medium confidence, v1.1) |
Main editorial body preserved directly in static HTML.
IRIS (Imagination with auto-Regression over an Inner Speech) treats world modeling as a sequence modeling problem. It tokenizes visual observations into discrete tokens using a VQ-VAE, then models environment dynamics autoregressively using a Transformer. This bridges the gap between language modeling and world modeling, achieving strong Atari 100K results.
IRIS is a discrete token-based dynamics model developed by Microsoft Research in 2023 for atari games.
Short extractable facts for answer engines and no-JS readers.
| Signal | Value |
|---|---|
| Definition | IRIS is a discrete token-based dynamics model developed by Microsoft Research in 2023 for atari games. |
| Short Description | A world model agent that tokenizes observations into discrete tokens and models environment dynamics autoregressively, like a language model over game frames. |
| Benchmark Rows | 1 |
| FAQ Entries | 1 |
| Related Models | 2 |
| Related Guides | 2 |
| Related Research Topics | 3 |
| Last Updated | 2026-02-25 |
Key capabilities associated with this model.
Representative applications attached to this model record.
Balanced assessment surfaced in static HTML.
Published benchmark evidence attached to this model record.
| Benchmark | Metric | Result | Source |
|---|---|---|---|
| Atari 100K | Mean HNS 1.046 x human | Strong performance | Source |
Primary references preserved in static HTML for citation extraction.
| Reference | Link |
|---|---|
| Micheli et al., 2023. Transformers are Sample-Efficient World Models. ICLR 2023. | Open source |
Side-by-side comparisons already connected to this model.
| Comparison | Matchup | Summary |
|---|---|---|
| World Models vs LLMs | World Models vs Large Language Models | World models and LLMs represent fundamentally different approaches to AI. World models learn causal dynamics of physical environments; LLMs learn statistical patterns over text. Both are essential for the future of AI. |
| DreamerV3 vs DIAMOND | DreamerV3 vs DIAMOND | DreamerV3 and DIAMOND are both model-based RL agents that train policies via imagination, but they use fundamentally different dynamics models: RSSM latent dynamics vs. pixel-space diffusion models. |
| IRIS vs DreamerV3 | IRIS vs DreamerV3 | IRIS and DreamerV3 are both leading model-based RL agents but use fundamentally different world model architectures: autoregressive token prediction vs. RSSM latent dynamics. |
| DreamerV3 vs IRIS | DreamerV3 vs IRIS | 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. |
| IRIS vs DIAMOND | IRIS vs DIAMOND | Two approaches to learning game simulators: IRIS uses discrete tokenization with a GPT-like transformer, while DIAMOND leverages diffusion models for higher visual fidelity. |
Crawler-readable guide links tied to this model.
| Guide | Summary |
|---|---|
| Building World Models: A Practical Guide | A practical guide to implementing world models: from choosing architectures and training setups to debugging dynamics learning and policy optimization. |
| Evaluating World Models: Benchmarks, Metrics and Pitfalls | A practical reference on how to benchmark world models: from sample efficiency on Atari 100K and DMControl to long-horizon prediction quality and downstream policy performance. |
Connected research areas surfaced directly in static HTML.
| Topic | Summary |
|---|---|
| Model-Based Reinforcement Learning | What model-based reinforcement learning is, how world models enable imagination-based planning, and why Dreamer, MuZero, PlaNet, and TD-MPC2 matter. |
| World Models vs LLMs | The key differences between world models and LLMs across objective, architecture, planning, physical reasoning, and embodied AI use cases. |
| World Model Evaluation | How to evaluate world models across rollout quality, benchmark performance, planning utility, and downstream transfer instead of relying on visual plausibility alone. |
FAQ answers rendered directly into static HTML for extractable responses.
IRIS treats world modeling as sequence prediction over discrete tokens, similar to how LLMs predict next tokens in text, but applied to game observations and actions.
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 Tyler D. - Technical editor, methodology and benchmark analysis.
This model page synthesizes primary papers, official model pages, benchmark evidence, and related world-models.io context into a reference resource.
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: 2026-02-25.
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 model and lab sources embedded in static HTML.