Contextual Latent World Models for Offline Meta Reinforcement Learning

Mohammadreza Nakhaei, Aidan Scannell, Kevin Luck, Joni Pajarinen · 2026

Read Contextual Latent World Models for Offline Meta Reinforcement Learning by Mohammadreza Nakhaei, Aidan Scannell, Kevin Luck, Joni Pajarinen (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-2603-02935.txt.

Description

Mohammadreza Nakhaei, Aidan Scannell, Kevin Luck, and Joni Pajarinen teach agents to recognize which task they face by how its world behaves. Their 2026 paper addresses offline meta-reinforcement learning, where policies must adapt to new tasks using only fixed past data. Context methods guess the task from recent transitions, but learning good guesses without labels is hard.

Their fix conditions a latent world model, an internal simulator, on the inferred task summary, and trains both parts together. That forces the summary to capture how the task's dynamics differ, not just surface clues. The learned summaries were richer, and adaptation to unseen tasks improved clearly. Gains held across three standard benchmark suites for control and manipulation.

Catalog record

CREATORMohammadreza Nakhaei, Aidan Scannell, Kevin Luck, Joni Pajarinen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97733
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.02935

Browse the full library · About the Tower of Babel