Call for Papers

PTA: From Pretrained Representations to Acting Agents — NeurIPS 2026 Workshop

Sequential decision-making is undergoing a transition where agents increasingly rely on pretrained representations learned from large-scale interaction data, videos, and multimodal corpora, yet a central challenge remains: how can pretrained representations be aligned with acting agents and transformed into effective behavior?

The PTA workshop will focus on this emerging interface between pretraining and control. Pretrained models can encode rich semantic, temporal, causal, or structural information. Still, acting agents must transform that knowledge into reliable behavior, including choosing actions over time, adapting to new goals, reasoning under uncertainty, recovering from mistakes, and remaining robust under distribution shift.

We invite submissions that investigate how pretrained knowledge can become actionable for sequential decision-making. Beyond proposing new methods, we particularly encourage work that studies the principles, limitations, evaluation methodologies, and trade-offs underlying different approaches. By bringing together researchers from reinforcement learning, representation learning, robotics, planning, and foundation models, the workshop aims to facilitate discussion toward a deeper understanding of when and why pretrained representations enable effective decision making.

Specific Topics of Interest

Topics include, but are not limited to:

1 What makes a pretrained representation actionable?

Focus on representation properties:

  • Successor representations, predictive state representations, and world-model latents
  • Representation geometry, abstraction, and controllability
  • Connections between representation learning, planning, and generalization
2 How should agents align and use pretrained representations at test time?

Focus on test-time decision making:

  • Test-time alignment of representations with planning objectives, constraints, and feedback
  • Test-time search, test-time optimization, online reasoning, and compute allocation
  • Retrieval, memory, verification, and tool use for sequential decision making
3 How can pretrained knowledge be adapted to new tasks and embodiments?

Focus on transfer and adaptation:

  • Pretraining actionable representations that support zero-shot and few-shot sequential decision making
  • Representation reuse across tasks, environments, and embodiments
  • Multimodal and embodied representations for robotic or interactive agents
4 How do we evaluate actionable representations?

Focus on benchmarks and measurement:

  • Evaluation protocols for robustness, transfer, controllability, and distribution shift
  • Benchmarks for long-horizon control and temporally extended decision making
  • Diagnosing representation failures in planning and acting agents
5 What can go wrong when pretrained agents act?

Focus on failure modes and safety:

  • Misalignment between pretrained abstractions and agent actions, goals, embodiments, or environments
  • Hallucinated plans, unsafe exploration, spurious abstractions, and poor uncertainty estimation
  • Compounding errors in sequential decision making, e.g., under environment, goal, and embodiment shift

Submission Instructions & Timeline

We will use OpenReview to manage submissions and the double-blind review process.

Format

Key Dates