Research map / Label-efficient and robust learning / Self-supervised and contrastive
Simple, Good, Fast: Self-Supervised World Models Free of Baggage
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SGF, a simple reconstruction-free, feedforward world model that learns VICReg-style self-supervised representations with frame/action stacking and augmentations, evaluated on Atari 100k with much shorter training time.
Problem
World models for sample-efficient RL typically rely on RNNs/transformers, discrete representations, and pixel reconstructions, which add complexity and computational cost; the paper asks which components are essential.
Approach
Learn world-model representations with a VICReg-based self-supervised objective (temporal consistency via an action-conditioned predictor on augmented successive observations plus variance/covariance regularization), using a feedforward dynamics model over stacked frames and actions, without reconstructions, discretization, or sequence models.
Key results
- SGF achieves a normalized mean of 0.884 and median of 0.152 on Atari 100k, below EfficientZero (1.943 mean) and DreamerV3 (1.12 mean) but competitive on several games such as Boxing (83.4) and Krull (7813.7).
- SGF trains in about 1.5 hours on a single A100 (approx. 3 hours on V100 equivalent), roughly four times faster than DreamerV3 (12 hours).
- Ablations show removing augmentations, action stacking, frame stacking, or temporal consistency degrades performance on five games; e.g., removing augmentations drops Boxing from 83.4 to 1.6 and removing temporal consistency drops Boxing to 0.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image augmentations (ablation) | SGF · Boxing | mean score: 1.6 -> 83.4 | Helped |
| add action stacking (ablation) | SGF · Breakout | mean score: 22.9 -> 50.7 | Helped |
| add frame stacking (ablation) | SGF · Breakout | mean score: 8.2 -> 50.7 | Helped |
| add temporal consistency (ablation) | SGF · Pong | mean score: -20.6 -> 12.6 | Helped |
| use dimension-contrastive rather than sample-contrastive learning (ablation) | SGF · Breakout | mean score: 25.1 -> 50.7 | Helped |
| increase projector hidden dimension from 2048 to 8192 (ablation) | SGF · Breakout | mean score: 42.2 -> 23.7 | Hurt |
| increase transition hidden dimension from 1024 to 4096 (ablation) | SGF · Breakout | mean score: 42.2 -> 27.1 | Hurt |
| increase actor and critic hidden dimensions from 512 to 2048 (ablation) | SGF · Boxing | mean score: 83.9 -> 77.9 | Hurt |
| increase imagination horizon from 10 to 15 steps (ablation) | SGF · Boxing | mean score: 83.9 -> 57.9 | Hurt |
| increase training frequency to two batches per environment step (ablation) | SGF · Boxing | mean score: 83.9 -> 64.1 | Hurt |
| increase frame and action stack size from 4 to 8 (ablation) | SGF · Breakout | mean score: 42.2 -> 21.8 | Hurt |
| increase frame and action stack size from 4 to 12 (ablation) | SGF · Breakout | mean score: 42.2 -> 14.4 | Hurt |
Surprising findings
- Removing frame stacking barely hurts Kung Fu Master, attributed to enemies facing their direction of movement so velocity is visible in a single frame.
- Making VICReg sample-contrastive (transposing the embedding matrix) significantly worsens Breakout but less so on other games.
Limitations
- Limited to deterministic MDPs; stochastic transitions would require stochastic predictors.
- Feedforward dynamics model restricts it to environments with mainly short-term dependencies, likely why it lacks state-of-the-art performance.
- VICReg requires image observations and suitable augmentations.
Experimental setup
Datasets: Atari 100k
Benchmarks: Atari 100k
Metrics: mean score median score human-normalized mean human-normalized median training runtime
Baselines: SPR EfficientZero IRIS DreamerV3 TWM SimPLe
Builds on: VICReg BYOL DreamerV3 SPR IRIS
Closest papers by meaning
- SPARTAN: A Sparse Transformer World Model Attending to What Matters (NeurIPS 2025 · similarity 0.86)
- WorldModelBench: Judging Video Generation Models As World Models (NeurIPS 2025 · similarity 0.85)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.85)
- stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation (NeurIPS 2026 · similarity 0.84)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.84)
- Trimming the Long-Tail of Visual World Modeling Evaluation (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.