COMBAT: Conditional World Models for Behavioral Agent Training
Anmol Agarwal, Sumer Singh, Pranay Meshram, Saurav Suman, Andrew Lapp, Shahbuland Matiana, Louis Castricato, Spencer Frazier · 2026
Read COMBAT: Conditional World Models for Behavioral Agent Training by Anmol Agarwal, Sumer Singh, Pranay Meshram, Saurav Suman, Andrew Lapp, Shahbuland Matiana, Louis Castricato, Spencer Frazier (2026) 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-2603-00825.txt.
Description
Agarwal and seven colleagues build COMBAT, a real time action controlled world model trained on Tekken 3, simulating a dynamic reactive foe. A 1.2 billion parameter Diffusion Transformer, fed by deep autoencoder latents, uses causal distillation and diffusion forcing for real time inference. Skilled foe behaviour emerges from training solely on single player inputs, with no explicit supervision of the foe's policy. Unlike imitation learning, which needs complete action labels, COMBAT learns quick behaviour from partial data.
Catalog record
| CREATOR | Anmol Agarwal, Sumer Singh, Pranay Meshram, Saurav Suman, Andrew Lapp, Shahbuland Matiana, Louis Castricato, Spencer Frazier |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29155 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.00825