amortized planning
Imagine you have to solve a new puzzle every time you enter a room. Amortized planning means building a reusable strategy or model beforehand, so that each new puzzle takes much less effort to solve. Instead of starting from scratch, you learn to quickly generate or select a good plan, spreading the initial learning cost across many planning problems.
The papers show this idea works across diverse domains. Coding agents learn to write planners that outperform hand-engineered ones. Learned world models or trajectory priors let agents simulate outcomes or generate action sequences without per-query optimization. Even multi-agent systems and navigation tasks benefit from amortized design, achieving higher accuracy and faster planning by shifting reasoning to a precomputation or learning phase.
- Papers
- 9
- Released code
- 2
- First seen
- June 2026
- Latest
- Oct 2026
7 papers in the last two months, against 1 in the two before.
Start here
Three papers to read first, in the order they came out- 1Three-stage training instills world model planning into LLM agents
A three-stage training paradigm enables LLM agents to internalize a world model for future-aware planning and outperforms baselines on search and math reasoning.
June 2026 - 2Reusable latent trajectories improve world-model control while cutting planning time
LeFlow replaces iterative action-space planning with amortized latent trajectory generation, cutting planning time by an order of magnitude while improving success rates.
Aug 2026 - 3Coding agents synthesize robot planners that generalize to unseen instances
Coding agents outperform hand-engineered planners on generalized task and motion planning problems by synthesizing reusable programs.
Sept 2026
How fast it is moving
Often studied with
agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.3 togethermulti-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.3 togetherlatent world modelA learned internal, compressed representation of environment dynamics, enabling future-aware planning, generation, and reasoning directly in latent space.3 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Top Universitycs.RO
Coding agents synthesize robot planners that generalize to unseen instances
Coding agents outperform hand-engineered planners on generalized task and motion planning problems by synthesizing reusable programs.
Fondazione Bruno Kessler, Carnegie Mellon University · Sept 2026
- Top Universitycs.CVcode
Reusable latent trajectories improve world-model control while cutting planning time
LeFlow replaces iterative action-space planning with amortized latent trajectory generation, cutting planning time by an order of magnitude while improving success rates.
University of Washington · Aug 2026
- Chinese Techcs.AI
Three-stage training instills world model planning into LLM agents
A three-stage training paradigm enables LLM agents to internalize a world model for future-aware planning and outperforms baselines on search and math reasoning.
Fudan University, Shanghai Innovation Institute · June 2026
- Top Universitycs.LG
Generative models as proposal distributions improve long-horizon decision making
Generative Trajectory Planning uses a generative model as a proposal distribution for action sequences enabling efficient long-horizon planning.
Harvard University · Oct 2026
- Top Universitycs.LG
LLM agents learn to organize document collections before questions arrive
Training a builder agent with reinforcement learning to reorganize documents into a filesystem workspace improves downstream question answering and transfers to unseen domains.
Stanford University, Carnegie Mellon University · Oct 2026
- Academiccs.LG
System that builds agent teams per query outperforms all baselines
A learned architect predicts harness utility, enabling per-query multi-agent system construction without executing candidates during search.
Arizona State University · Oct 2026
- Top Universitycs.RO
Learned control coordinates frozen diffusion models into multi-agent generators
Coordinated multi-agent diffusion steering (CMDS) learns an amortised control to coordinate frozen pretrained diffusion models using an assembly-level reward.
UCL, TUM · Oct 2026
- Academiccs.AI
Agents can budget external checks without sacrificing task quality
BudgetPM allocates limited external observations across stored agent intentions to maximize task quality under a hard budget.
Xi’an Jiaotong University · Oct 2026
- Research Labcs.CVcode
Dual-horizon foresight improves UAV navigation from language instructions
ForeFly predicts both proximal and route-critical futures to improve aerial navigation from language instructions.
Institute of Automation, Chinese Academy of Sciences, University of Chinese Academy of Sciences · Sept 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.