Learning Multi-Timescale Abstractions for Hierarchical Combinatorial Planning

arXiv 2026 47.8 method

TLDR

A model-based hierarchical RL framework with multi-timescale latent dynamics for sequential stochastic combinatorial optimization.

Reasoning

The paper introduces a novel approach combining latent-space tree-search planning with an SMDP-aware world model, addressing the challenge of variable-duration actions in hierarchical RL. Its strengths include a multi-timescale objective for efficient lookahead and a budget policy for resource allocation; however, the abstract lacks details on real-world validation, and the benchmarks may be synthetic.

Read-first score

Read-first score 47.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

80.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.80683558

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 320.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Multi-timescale objective for latent dynamics to reflect effective temporal scales of abstract actions
SMDP-aware world model for variable-duration decisions in hierarchical planning
Subgoal-conditioned budget policy jointly learned with world model for context-aware resource allocation
Integration of latent-space tree-search planner with hierarchical combinatorial planning

Methodology

The method introduces a model-based hierarchical framework combining a latent-space tree-search planner with an SMDP-aware world model that handles variable-duration decisions. A multi-timescale objective structures latent dynamics so transition magnitudes correspond to the temporal scales of abstract actions, enabling efficient lookahead. A subgoal-conditioned budget policy is learned jointly with the world model to support context-aware resource allocation.

Key Results

The proposed method outperforms strong baselines across challenging Sequential Stochastic Combinatorial Optimization benchmarks.

Resources

Paper