Imagine · Verify · Act

ParallelWorld

Test-Time Scaling for
Agentic Embodied Reasoning

Min Chen^1,*Shengjun Zhang^1,*,†Yuxin Li¹Zhang Zhang¹Xin Fei²Chong Xia¹Yueqi Duan¹

¹ Tsinghua University² National University of Singapore

* Equal contribution † Project Leader

arXiv Code Watch video

Explore alternative futures before acting.
Prospective world imagination for understanding and manipulation.

01 / Overview

Reason about what comes next.

An agent should consider how its actions shape future observations and task progress before committing to them.

Embodied AI has emerged as a key paradigm for autonomous systems operating in complex physical environments. While end-to-end models have established strong local capabilities for perception and control, scaling these capabilities to extended, multi-stage tasks remains difficult. Agentic embodied systems address this challenge by introducing high-level planners that organize perception, reasoning, and executable skills over longer interaction horizons. However, existing agentic embodied frameworks typically generate actions for direct execution without explicitly simulating how alternative action sequences shape future observations and task progress, leading agents to commit to suboptimal interactions before considering their long-term consequences.

To address this limitation, we propose ParallelWorld, an adaptive-horizon test-time scaling framework that grounds agentic embodied reasoning in the prospective simulation of action outcomes. Specifically, we design an agent system to simulate and evaluate alternative trajectories in parallel before committing to an action. A verifier assesses intermediate state transitions with respect to the task objective, dynamically pruning unpromising branches and guiding subsequent expansion. Based on the evaluated long-horizon outcomes, the agent selects a trajectory, executes only its first action, and replans as new observations become available.

Experiments on ESI-Bench and LIBERO-Pro demonstrate improvements in active spatial reasoning and language-conditioned manipulation, respectively, highlighting the effectiveness of prospective planning for both understanding and manipulation tasks.

Comparison of end-to-end systems, conventional agentic systems, and ParallelWorld, which evaluates imagined futures before executing an action. — **From direct execution to prospective planning.** ParallelWorld augments agentic planning with multi-step world imagination and verifier-guided action selection.

02 / Method

Imagine in parallel. Act with foresight.

An adaptive imagination loop meets a closed-loop interaction process.

ParallelWorld framework: an agentic planner proposes candidate actions, a simulator or world model generates prospective outcomes, and a verifier evaluates branches before the first selected action is executed. — The planner adaptively decides how many candidates to propose and whether to continue imagination. The verifier retains promising branches and provides comparative feedback.

Imagine

Propose alternative primitive actions and roll out their prospective observations across multiple steps.

Verify

Evaluate outcomes against the task objective, prune unpromising branches, and guide further expansion.

Act & replan

Execute only the first action of the selected trajectory. Observe the resulting state and begin a new search.

Embodied understanding

Imagine to see more.

Simulator-based imagination restores candidate states to explore views and interactions that reveal task-relevant evidence.

Robotic manipulation

Imagine to act better.

Image-based world imagination predicts the visual outcomes of executable primitives to compare progress toward the goal.

03 / Experiments

One framework. Two embodied settings.

Evaluated on active spatial reasoning and language-conditioned manipulation under perturbations.

+20.00 pp

Partial Occlusion

ESI-Bench · 57.89% → 77.89%

+18.91 pp

Unobserved Change

ESI-Bench · 70.95% → 89.86%

+27.0 pp

Spatial position-swap

LIBERO-Pro · 56.0% → 83.0%

Percentage-point gains over Active Exploration on ESI-Bench and HarnessVLA on LIBERO-Pro.

Active spatial reasoning

ParallelWorld improves over Active Exploration on 26 of 28 evaluated subcategories. Liquid Volume is excluded due to a simulator issue with pouring actions.

ESI-Bench · accuracy (%) · 28 evaluated subcategories
Subcategory	Passive	Active	ParallelWorld
View Hallucination	53.32	66.82	71.40
Partial Occlusion	36.84	57.89	77.89
Material Transparency	69.72	71.10	73.85
Rigid Containment	55.00	60.00	50.00
Deformable Objects	37.76	40.82	42.86
Inclined Plane	59.02	83.61	85.25
Stacking & Stability	1.12	45.98	53.95
Reflection Authoring	56.57	51.52	60.61
Specular Reflection: Spatial Relations	38.94	36.28	51.33
Correspondence	42.05	47.73	48.86
Linear Alignment	52.13	68.09	70.21
Geometric Configuration	16.55	17.96	16.90
Physical Contact	66.39	68.07	73.95
Dimensional Size	39.52	44.91	49.10
Spatial Distance	61.18	58.55	67.11
Connectivity	66.67	55.00	60.00
Traversable Passage	68.33	65.00	75.00
Regional Boundary	45.00	57.50	63.75
Long-Term Navigation	16.67	16.67	18.33
Counting with Occlusion	23.33	26.67	30.00
Spatial Segmentation	20.00	20.00	23.33
Merged Observation	0.00	1.67	6.67
Category Ambiguity	0.00	8.33	13.33
Structural Enclosure	2.50	5.00	7.50
Illumination Variability	0.00	6.00	10.00
Unobserved Change	38.62	70.95	89.86
Agent Observation	34.59	48.87	62.41
Action Order Inference	37.66	53.25	61.04

Accuracy (%), from Table 1 of the paper. All active methods use the same answer model. Bold marks the best result in each row.

Manipulation under perturbations

ParallelWorld improves over HarnessVLA in all eight settings. Each setting contains 100 episodes: 10 tasks × 10 seeds, for 800 episodes in total.

LIBERO-Pro · success rate (%) · 800 episodes
Method	Spatial		Object		Goal		LIBERO-10
Method	T	S	T	S	T	S	T	S
OpenVLA	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
π0	0.0	0.0	0.0	2.0	0.0	0.0	0.0	0.0
π0.5	1.0	20.0	1.0	17.0	2.0	38.0	1.0	8.0
MolmoAct	0.0	0.0	0.0	6.0	0.0	0.0	6.0	0.0
NORA	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
X-VLA	0.0	0.0	8.0	2.0	9.0	1.0	10.0	0.0
AtomVLA	1.0	16.0	0.0	10.0	11.0	2.0	9.0	1.0
HarnessVLA	79.0	56.0	79.0	85.0	62.0	57.0	42.0	29.0
ParallelWorld (Ours)	88.0	83.0	94.0	88.0	65.0	66.0	50.0	30.0

Success rate (%), from Table 2. T: instruction-redirection. S: position-swap. HarnessVLA is reproduced with Codex as the planner; frozen VLA results are taken from the HarnessVLA paper.

04 / Qualitative analysis

Where foresight helps—and where it falls short.

Prospective evaluation helps select informative views and favorable manipulation outcomes. Precise geometry and contact dynamics remain challenging.

Four qualitative examples: successful prospective exploration and manipulation, alongside limitations in precise geometric reasoning and physical prediction. — **Successes and limitations.** Top: imagined alternatives reveal useful evidence and guide manipulation. Bottom: sufficient visual evidence does not guarantee metric accuracy, and visually plausible outcomes do not guarantee physical success.

05 / Ablation

Adapt the search to the decision.

Adaptive search width allocates more candidates when alternatives are ambiguous and reduces redundant exploration when the next step is clear.

Spatial-S subset · 3 tasks × 10 seeds
Method	Success rate (%) ↑	Total time (s) ↓
Direct Execution	23.3	417.4
Fixed Search Width (K = 2)	10.0	2930.0
Fixed Search Width (K = 4)	36.7	4889.5
Adaptive Search Width (Ours)	73.3	2402.2

Table 3 of the paper. Total time is the average per episode. Adaptive search outperforms the tested fixed widths, while adding computation relative to direct execution.

06 / Citation

BibTeX

@misc{chen2026parallelworld,
  title  = {ParallelWorld: Test-Time Scaling for Agentic Embodied Reasoning},
  author = {Min Chen and Shengjun Zhang and Yuxin Li and Zhang Zhang
            and Xin Fei and Chong Xia and Yueqi Duan},
  year   = {2026},
  eprint = {2608.22971},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url    = {https://arxiv.org/abs/2608.22971}
}