SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning

arXiv 2026 33.3 method

TLDR

SAGE improves long-horizon planning in latent world models by using subgoal-conditioned action generation, achieving higher success rates on PushT and OGBench Cube.

Reasoning

The paper introduces a novel prior-conditioned planner that decomposes goals into subgoals to guide action generation, effectively addressing the proposal quality issue in long-horizon planning. However, the evaluation is limited to two simulated benchmarks (PushT and OGBench Cube), lacking real-world or diverse task validation, and the method's generalizability remains unclear.

Read-first score

Read-first score 33.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 5%

100

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

Methodology quality 15%

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

Topical relevance 25%

55.7

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 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 279.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Prior-conditioned planner that replaces random proposal initialization with structured guidance via a goal-conditioned generator predicting the next reachable latent subgoal for a specified duration.
Use of subgoals at varying durations as priors to capture semantic information across temporal scales, balancing fine-grained local control with long-horizon progress.
Conditioning action sequence generation on predicted subgoals before evaluation and refinement by a frozen world model.

Methodology

SAGE is a prior-conditioned planner for latent world models. At each planning stage, a goal-conditioned generator predicts reachable latent subgoals of varying durations; these subgoals condition the generation of candidate action sequences. A frozen world model then evaluates and refines these proposals before execution.

Key Results

On PushT and OGBench Cube with target offset 150, SAGE improves long-horizon success rates from 12.7% to 64.7% and from 26.7% to 67.3%, respectively, while preserving strong short-horizon performance.

Resources

Paper