Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators

arXiv 2026 68.8 method

TLDR

Astra couples a VLM policy with a world simulator for agentic spatial reasoning, improving cross-view consistency via RL and view consistency tuning.

Reasoning

Strengths include a novel integration of world simulators with VLMs for spatial reasoning, a two-phase RL curriculum, and view consistency tuning. Weaknesses are the abstract being cut off, limited evaluation to one benchmark (MMSI-Bench), and potential lack of generalization to diverse real-world scenarios.

Read-first score

Read-first score 68.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.

Methodology quality 18%

100

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

Recency 6%

100

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

Citation impact 18%

93.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93444584

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

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 390.

Keyword Scores

world simulator

generative world model

interactive world model

world dynamics prediction

world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Agentic spatial reasoning framework that empowers VLMs with action-conditioned visual imagination by coupling an RL-trained VLM policy (Astra-VL) with a world simulator (Astra-WM).
Astra-WM: a Bagel-based world simulator trained with view consistency tuning to generate novel-view observations from context images and natural-language camera motions with improved pose and content consistency.
World-simulator-in-the-loop two-phase RL curriculum that stabilizes tool-use exploration and teaches the model when to invoke the simulator versus direct answering.
Demonstration that both the world simulator and the agentic policy are necessary for effective spatial reasoning, with quantitative gains on MMSI-Bench and MindCube.

Methodology

The paper proposes Astra, an agentic spatial reasoning framework consisting of Astra-VL (an RL-trained VLM policy) and Astra-WM (a Bagel-based world simulator). Astra-WM generates novel-view observations from context images and natural-language camera motions, and is trained with view consistency tuning to ensure pose and content consistency. The RL stage employs a world-simulator-in-the-loop two-phase curriculum to stabilize exploration and learn when to invoke the simulator. Evaluation is performed on MMSI-Bench and MindCube benchmarks, comparing against baselines like Gemini-3-Flash and Qwen3-VL.

Key Results

Astra-WM improves simulator-augmented Gemini-3-Flash on MMSI-Bench from 45.1 to 49.5. Astra-VL improves the Qwen3-VL backbone from 29.8 to 38.8 on MMSI-Bench and from 36.8 to 42.7 on MindCube.

Limitations

The world simulator (Astra-WM) may still have imperfect view consistency, as view consistency tuning is used to mitigate issues.
The RL curriculum is required to stabilize tool-use exploration, indicating that naive training without the curriculum may be unstable or ineffective.
Generalization beyond the evaluated benchmarks (MMSI-Bench and MindCube) is not demonstrated, and the method's reliance on a learned simulator may limit applicability to domains where novel-view generation is challenging.

Resources

Paper