Uncovering Untapped Potential in Sample-Efficient World Model Agents
Lior Cohen, Kaixin Wang, Bingyi Kang, Uri Gadot, Shie Mannor · 2025
Read Uncovering Untapped Potential in Sample-Efficient World Model Agents by Lior Cohen, Kaixin Wang, Bingyi Kang, Uri Gadot, Shie Mannor (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11537.txt.
Description
Lior Cohen and co-authors stack known improvements into one learning agent in this 2025 paper. The head of the text titles the agent Simulus. World model agents learn an inner copy of their game world, then practice inside it. That makes them sample-efficient but complex to build well. Past upgrades were scattered across many papers.
Simulus combines four of them: flexible input tokens, curiosity driven by uncertainty, smarter replay of past dreams, and a new way to predict rewards. On three test suites it set new marks for agents that plan-free learn this way. Each part helped alone, and together they helped more. The code and trained weights are public.
Catalog record
| CREATOR | Lior Cohen, Kaixin Wang, Bingyi Kang, Uri Gadot, Shie Mannor |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 82 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-11537