Dictionary entry
in-context reinforcement learning
also ICRL or in-context RL
Definition of in-context reinforcement learning
-
: the ability of a sequence model or agent to improve its behavior during interaction by using earlier observations, actions, and rewards in its context, without updating the model's parameters
- The agent demonstrated in-context reinforcement learning when it changed strategy after seeing which earlier actions earned rewards.
- ICRL let one pretrained policy adapt to a new task from its interaction history instead of starting a new gradient-based training run.
Origin & history
The phrase was established by the 2022 paper In-context Reinforcement Learning with Algorithm Distillation. That work trained a causal sequence model on reinforcement-learning histories so it could reproduce policy improvement within its context at inference time, without changing its weights.
Editorial references
Sources & further reading
- In-context Reinforcement Learning with Algorithm Distillation↗ Research Paper · arXiv · 2022-10-25
- A Survey of In-Context Reinforcement Learning↗ Research Survey · arXiv · 2025-02-12
- In-Context Reinforcement Learning under Non-Stationarity: A Survey↗ Research Survey · arXiv · 2026-07
Test yourself
Which of these is the meaning of in-context reinforcement learning?
Cite this entry
"In-context reinforcement learning." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/in-context-reinforcement-learning/. Accessed 15 Oct. 2026.
Dictionary entries near in-context reinforcement learning
- hybrid intelligence
- hyperscaler
- imitation learning
- in-context learning
- in-context reinforcement learning
- industrial AI
- inference
- instruction tuning
- instrumental convergence