Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control

arXiv 2026 45.9 method

TLDR

Qantara introduces a joint training objective for JEPA world models enabling multiple inference paradigms from a single checkpoint, achieving SOTA on OGBench-Cube.

Reasoning

The paper presents a novel training method combining Brownian-bridge interpolant and flow matching, enabling latent planning, behavior cloning, and inverse dynamics from one checkpoint. Strengths include SOTA results and multi-paradigm flexibility; weaknesses include complexity and limited scope to latent world models.

Read-first score

Read-first score 45.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 54.

Recency 6%

100

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

Topical relevance 29%

77.1

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

Methodology quality 18%

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=checkpoint

Citation impact 18%

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

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 464.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

video world model

world simulator

generative world model

Deep Analysis

Innovations

Single JEPA checkpoint serving three inference paradigms (latent planning, behaviour cloning, inverse dynamics) without retraining
Joint training objective combining Brownian-bridge interpolant on state axis with noise-to-data flow matching on action axis
Video-inverse composition for inverse dynamics: predict next latent without action conditioning, then extract action
Training mass concentration on edges of (action-time, state-time) noise square to align with inference queries

Methodology

Qantara is an end-to-end JEPA trained with a joint objective that pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. Training concentrates probability mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor. A single checkpoint supports latent planning, behaviour-cloning action sampling, and inverse dynamics via a video-inverse composition that first predicts the next latent without action conditioning and then extracts the action.

Key Results

On the LeWM control suite, Qantara achieves a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. Uniform interior sampling during training reduces Push-T planning from 90.1 to 53.3 SR.

Limitations

Performance is sensitive to training distribution: replacing edge-concentrated sampling with uniform interior sampling drastically reduces planning success rate (90.1 to 53.3 SR on Push-T).

Resources

Paper