Research map / Large language models / Language, safety and interpretability
Context and Diversity Matter: The Emergence of In-Context Learning in World Models
Fan Wang, ZHIYUAN CHEN, YUXUAN ZHONG, Sunjian Zheng, Pengtao Shao, Bo Yu, Shaoshan Liu, Jianan Wang, Ning Ding, Yang Cao, Yu Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase context length from T=1 to T=1000 | L2World (Maze-32K-L) · Maze unseen envs | PSNR: 16.37 -> 24.66 | Improved |
| train on 32K diverse environments instead of 128 (ablation) | L2World · Maze unseen envs, T=10000 | PSNR: 21.52 -> 24.65 | Improved |
| train on 128 environments instead of 32K (ablation) | L2World · Maze seen envs, T=10000 | PSNR: 25.05 -> 26.00 | Improved |
| use long trajectories (10K) instead of short (100) (ablation) | L2World, 32K envs · Maze unseen envs, T=10000 | PSNR: 20.31 -> 24.65 | Improved |
| use L2World instead of Dreamer (LSTM) | L2World (Maze-32K-L) · Maze unseen envs, T=10000 | PSNR: 22.12 -> 24.65 | Improved |
| use L2World instead of NWM (diffusion, 4 frames) | L2World (Maze-32K-L) · Maze unseen envs, T=10000 | PSNR: 17.85 -> 24.65 | Improved |
| pre-train on Maze-32K-L before ProcTHOR-5K (ablation) | L2World · ProcTHOR-5K unseen, T=10000 | PSNR: 19.81 -> 22.04 | Improved |
| increase ProcTHOR data from 5K to 40K (2K-length trajectories) (ablation) | L2World pre-trained on Maze-32K-L · ProcTHOR unseen, T=10000 | PSNR: 22.04 -> 23.94 | Mixed |
| train on 4 environments instead of many (cart-pole) (ablation) | world model · Random cart-pole unseen envs | prediction error | Hurt |
| train on both Scope 1+2 with 8K envs instead of Scope 1 only (ablation) | world model · Random cart-pole | prediction error | Improved |
| early-stop checkpoint instead of over-training (4 envs) (ablation) | world model · Random cart-pole unseen envs | prediction error | Improved |
| shuffle 20% or 50% of context observations (ablation) | L2World (Maze-32K-L vs Maze-128-L) · Maze | PSNR | Hurt |
Closest papers by meaning
- Context is Environment (ICLR 2024 · similarity 0.88)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.87)
- In-Context Learning Strategies Emerge Rationally (NeurIPS 2025 · similarity 0.87)
- Can In-Context Learning Support Intrinsic Curiosity? (NeurIPS 2026 · similarity 0.87)
- Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization (NeurIPS 2025 · similarity 0.86)
- World Models Should Prioritize the Unification of Physical and Social Dynamics (NeurIPS 2025 · similarity 0.86)
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.
