We invite submissions on a broad range of topics related to learning in big worlds. These include:
- Methods that learn to construct an agent state from the history of an agent's observations and actions. Some examples include algorithms for efficient gradient-based recurrent learning, sparse attention architectures, architectures for learning predictive representations, and algorithms for feature search.
- Methods for exploration that operate on the agent state instead of the environment's state.
- Proposals and benchmarks for evaluating algorithms in big worlds. For example, benchmarks that limit the agent's computation or memory.
- Methods for continual learning to deal with unforeseeable situations. Examples include algorithms that prevent catastrophic loss of plasticity and catastrophic forgetting, and computationally efficient algorithms that can learn using only the computation available at deployment.
- Methods for learning models of the agent state and planning with approximate models. Some examples include abstract models of the agent's representation and planning algorithms that can effectively use inaccurate models.
- Methods for discovering temporal abstractions and using temporal abstractions for exploration and planning. For example, this includes option discovery from experience and methods that can plan with option models.
- Meta-learning methods that learn some aspect of the learning algorithm. For example, these include methods that learn to assign credit by modulating activations or step-size parameters, and methods that adapt the agent's hyperparameters online from experience.
- Theory that makes assumptions about the agent's capabilities but not about the environment's complexity. For example, a convergence proof of a linear learning method that works for arbitrary features.
Strong submissions to our workshop should focus on what learning algorithms can control, without relying on simplifying assumptions about the environment. A submission is not a good fit if it assumes that the environment is governed by a small set of simple causal mechanisms, that all states can be enumerated, or that the environment is fully observable or deterministic. Likewise, a theoretical result is not relevant to learning in big worlds if it assumes the agent can represent the optimal policy, the true value function, or a perfect model of its environment.
A submission may, however, study a single aspect of learning in big worlds while making simplifying assumptions about others. For example, a paper on planning with inaccurate models may use a fully observable environment in its experiments, so long as the model used for planning is itself inaccurate. We encourage authors to use their judgment to decide whether their work engages with at least one aspect of learning in big worlds.
Submission Instructions
Papers should be anonymized, submitted in PDF format and use the RLC style file (available here). Paper length should be 4 to 8 pages, excluding references. A cover page is not needed.
You may submit recently published work (published after September 2024). Submissions are not archival. All accepted papers will have a poster presentation at the workshop, and one paper will be selected for a 15 minute oral presentation.
The review process will be double-blind. Please submit your paper through OpenReview. If you have questions, then please email [email protected].
Important Dates
- Paper submission deadline
- 15 May 2026 (AoE)
- Decision notification and reviewer feedback
- 29 May 2026
- Final version deadline for accepted papers
- 15 July 2026 (AoE)