Transformers are Sample-Efficient World Models
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:LZZVK5BOrecord.jsonopen to challenge →
read the original abstract
Deep reinforcement learning agents are notoriously sample inefficient, which considerably limits their application to real-world problems. Recently, many model-based methods have been designed to address this issue, with learning in the imagination of a world model being one of the most prominent approaches. However, while virtually unlimited interaction with a simulated environment sounds appealing, the world model has to be accurate over extended periods of time. Motivated by the success of Transformers in sequence modeling tasks, we introduce IRIS, a data-efficient agent that learns in a world model composed of a discrete autoencoder and an autoregressive Transformer. With the equivalent of only two hours of gameplay in the Atari 100k benchmark, IRIS achieves a mean human normalized score of 1.046, and outperforms humans on 10 out of 26 games, setting a new state of the art for methods without lookahead search. To foster future research on Transformers and world models for sample-efficient reinforcement learning, we release our code and models at https://github.com/eloialonso/iris.
This paper has not been read by Pith yet.
Forward citations
Cited by 29 Pith papers
-
Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution
Promptbreeder evolves both task prompts and the mutation prompts that improve them using LLMs, outperforming Chain-of-Thought and Plan-and-Solve on arithmetic and commonsense reasoning benchmarks.
-
MBench: A Comprehensive Benchmark on Memory Capability for Video World Models
MBench is a new benchmark that quantifies long-term memory in video world models via three hierarchical consistency dimensions evaluated on curated real videos.
-
MiraBench: Evaluating Action-Conditioned Reliability in Robotic World Models
MiraBench defines action-conditioned reliability via three levels (physics adherence, action-following fidelity, optimism bias detection) and applies it to 12 model configurations using a 16,000-judgment human corpus,...
-
JEDI: Joint Embedding Diffusion World Model for Online Model-Based Reinforcement Learning
JEDI is the first online end-to-end latent diffusion world model that trains latents from denoising loss rather than reconstruction, achieving competitive Atari100k results with 43% less VRAM and over 3x faster sampli...
-
Render, Don't Decode: Weight-Space World Models with Latent Structural Disentanglement
NOVA represents world states as INR weights for decoder-free rendering, compactness, and unsupervised disentanglement of background, foreground, and motion in video world models.
-
Latent State Design for World Models under Sufficiency Constraints
World models succeed when their latent states are built to meet task-specific sufficiency constraints rather than preserving the maximum amount of information.
-
Agent-Centric Observation Adaptation for Robust Visual Control under Dynamic Perturbations
ACO-MoE employs agent-centric mixture-of-experts to decouple task-relevant features from dynamic visual perturbations in RL, recovering 95.3% of clean performance on the new VDCS benchmark.
-
Agent-Centric Observation Adaptation for Robust Visual Control under Dynamic Perturbations
ACO-MoE recovers 95.3% of clean-input performance in visual control tasks under Markov-switching corruptions by routing restoration experts and anchoring representations to clean foreground masks.
-
Advantage-Guided Diffusion for Model-Based Reinforcement Learning
Advantage-guided diffusion (SAG and EAG) steers sampling in diffusion world models to higher-advantage trajectories, enabling policy improvement and better sample efficiency on MuJoCo tasks.
-
Training Agents Inside of Scalable World Models
Dreamer 4 is the first agent to obtain diamonds in Minecraft from only offline data by reinforcement learning inside a scalable world model that accurately predicts game mechanics.
-
Massive Activations in Large Language Models
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
-
Learning Interactive Real-World Simulators
UniSim learns a universal real-world simulator from orchestrated diverse datasets, enabling zero-shot deployment of policies trained purely in simulation.
-
Mastering Diverse Domains through World Models
DreamerV3 uses world models and robustness techniques to solve over 150 tasks across domains with a single configuration, including Minecraft diamond collection from scratch.
-
Flow Matching in Feature Space for Stochastic World Modeling
FlowWM applies flow matching directly in pretrained feature space with a one-step projection mechanism, improving perception accuracy, mode coverage, and horizon robustness on synthetic and real-world benchmarks.
-
Qwen-AgentWorld: Language World Models for General Agents
Qwen-AgentWorld are language world models that simulate multi-domain agent environments and boost general agent capabilities via decoupled RL simulation and unified foundation model training.
-
Autoregressive Diffusion World Models for Off-Policy Evaluation of LLM Agents
ADWM learns a latent diffusion world model with per-transition independent denoising and policy-conditioned guidance to enable accurate offline evaluation of LLM agent policies.
-
CLAW: Learning Continuous Latent Action World Models via Adversarial Latent Regularization
CLAW is an end-to-end self-supervised method that learns semantically meaningful continuous latent actions and predictive world models from action-free videos to support imitation learning and goal-directed planning.
-
AR Forcing: Towards Long-Horizon Robot Navigation World Model
AR Forcing trains diffusion world models by integrating standard noise prediction loss into an autoregressive loop that uses self-generated predictions as context, reducing train-inference mismatch for improved long-h...
-
Self-supervised Hierarchical Visual Reasoning with World Model
ResDreamer proposes a residual-based hierarchical world model for purely self-supervised visual foresight reasoning that scales with linear communication cost and reports state-of-the-art sample and parameter efficien...
-
Render, Don't Decode: Weight-Space World Models with Latent Structural Disentanglement
NOVA represents scene states as INR weights for analytical rendering without decoders and achieves structural disentanglement of content and dynamics in video world models.
-
OFlow: Injecting Object-Aware Temporal Flow Matching for Robust Robotic Manipulation
OFlow unifies temporal foresight and object-aware reasoning inside a shared latent space via flow matching to improve VLA robustness in robotic manipulation under distribution shifts.
-
Model-Based Reinforcement Learning under Random Observation Delays
A delay-aware model-based RL framework with sequential belief filtering handles random out-of-sequence observations in POMDPs and outperforms MDP baselines while showing robustness to delay shifts.
-
DAWM: Diffusion Action World Models for Offline Reinforcement Learning via Action-Inferred Transitions
DAWM introduces a modular diffusion world model with an inverse dynamics model to produce complete synthetic transitions that improve conservative offline RL algorithms like TD3BC and IQL on D4RL tasks.
-
Physically Interpretable World Models via Weakly Supervised Representation Learning
PIWM aligns latent states in image-based world models with physical variables and constrains their dynamics to known equations via weak distribution supervision, yielding accurate long-horizon predictions and paramete...
-
DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning
DINO-WM builds world models on pre-trained DINOv2 features to enable zero-shot planning from offline data without rewards or demonstrations.
-
Physically Viable World Models: A Case for Query-Conditioned Embodied AI
Embodied AI requires query-conditioned world models that select the simplest physical abstraction sufficient to answer intervention queries.
-
Self-supervised Hierarchical Visual Reasoning with World Model
ResDreamer proposes a residual-reconstruction hierarchical world model for purely self-supervised visual foresight that claims SOTA sample and parameter efficiency in open-world RL.
-
Towards World Models in Biomedical Research
Proposes biomedical world models that learn latent states and intervention-conditioned dynamics to enable simulation of future biological trajectories for discovery in virtual cells, organoids, patients, and surgery.
-
Behavior-Invariant Task Representation Learning with Transformer-based World Models for Offline Meta-Reinforcement Learning
The work introduces behavior-invariant latent task representations via information-theoretic learning in a Transformer world model plus conservative penalties on imagined rollouts to improve generalization in offline meta-RL.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.