Paper deep dive
ParallelWorld: Test-Time Scaling for Embodied Reasoning
Min Chen, Shengjun Zhang, Yuxin Li, Zhang Zhang, Xin Fei, Chong Xia, Yueqi Duan
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Embodied Reasoning constitutes a fundamental capability of embodied intelligence, serving as the basis for autonomous perception, reasoning, and interaction within physical environments. Recent studies have shifted the paradigm of embodied reasoning from static perception toward dynamic exploration, where agents acquire task-relevant information through interactions with the environment. However, existing active reasoning approaches generally generate exploration trajectories incrementally without long-horizon planning. Even recently emerged test-time scaling frameworks often resort to myopic, single-step lookaheads, which struggle to resolve the delayed feedback inherent in complex, occluded spatial environments. To address this limitation, we propose ParallelWorld, a multi-horizon test-time scaling framework for embodied reasoning. Instead of greedy, single-step trials, ParallelWorld empowers agents to simulate and evaluate multi-step future trajectories in parallel before committing to an action. Specifically, we introduce a verifier-guided tree-search paradigm. Starting from the current state, ParallelWorld branches into multiple parallel trajectories and rolls them out continuously across a multi-step horizon. At each simulation step, a verifier agent evaluates the intermediate state transitions, dynamically pruning unpromising branches and prioritizing paths with the highest information gain. Once the multi-step prospective simulation is complete, the agent synthesizes the long-horizon outcomes to commit to the optimal action sequence. Finally, an answer agent performs reasoning over the selected trajectory to produce the final reasoning. Extensive experiments on ESI-Bench demonstrate that ParallelWorld consistently improves active perception and reasoning performance.
Tags
Links
- Source: https://arxiv.org/abs/2608.22971v1
- Canonical: https://arxiv.org/abs/2608.22971v1
Trouble viewing inline? Open PDF directly →
Full Text
40,827 characters extracted from source content.
Expand or collapse full text
]1Tsinghua University, 2National University of Singapore †footnotetext: ∗* Equal contribution, † Project Leader [ Project Page]https://chen-min-22.github.io/ParallelWorld-page/ ParallelWorld: Test-Time Scaling for Embodied Reasoning Min Chen1,∗ Shengjun Zhang1,∗,† Yuxin Li1 Zhang Zhang1 Xin Fei2 Chong Xia1 Yueqi Duan1 Affiliation: [ Abstract Embodied Reasoning constitutes a fundamental capability of embodied intelligence, serving as the basis for autonomous perception, reasoning, and interaction within physical environments. Recent studies have shifted the paradigm of embodied reasoning from static perception toward dynamic exploration, where agents acquire task-relevant information through interactions with the environment. However, existing active reasoning approaches generally generate exploration trajectories incrementally without long-horizon planning. Even recently emerged test-time scaling frameworks often resort to myopic, single-step lookaheads, which struggle to resolve the delayed feedback inherent in complex, occluded spatial environments. To address this limitation, we propose ParallelWorld, a multi-horizon test-time scaling framework for embodied reasoning. Instead of greedy, single-step trials, ParallelWorld empowers agents to simulate and evaluate multi-step future trajectories in parallel before committing to an action. Specifically, we introduce a verifier-guided tree-search paradigm. Starting from the current state, ParallelWorld branches into multiple parallel trajectories and rolls them out continuously across a multi-step horizon. At each simulation step, a verifier agent evaluates the intermediate state transitions, dynamically pruning unpromising branches and prioritizing paths with the highest information gain. Once the multi-step prospective simulation is complete, the agent synthesizes the long-horizon outcomes to commit to the optimal action sequence. Finally, an answer agent performs reasoning over the selected trajectory to produce the final reasoning. Extensive experiments on ESI-Bench demonstrate that ParallelWorld consistently improves active perception and reasoning performance. 1 Introduction Understanding and reasoning about the physical world is a fundamental capability of embodied agents [5, 35, 21]. Unlike conventional systems that reason primarily from fixed observations, embodied agents must build a goal-directed understanding of the physical world through interactions with their environment. Recent advances in multimodal foundation models [8, 16] have substantially improved the reasoning capabilities of embodied agents, enabling them to address increasingly complex tasks in simulated and real-world environments. However, achieving reliable embodied understanding remains challenging, as agents often need to make decisions under incomplete observations, ambiguous visual evidence, and dynamically changing environments. To address these challenges, recent studies have shifted embodied understanding from passive perception toward active exploration, where agents are allowed to interact with the environment and gather additional information before making decisions [14, 23, 43, 39, 42, 40]. Rather than relying solely on existing observations, these approaches formulate understanding as an interactive process involving perception, action, and reasoning. By actively selecting informative observations or executing task-related actions, embodied agents can progressively improve their understanding of the surrounding environment. Despite these progresses, existing active understanding approaches typically generate exploration behaviors in a sequential manner, where each action is selected based only on the current observation and previously acquired information. Such a formulation lacks explicit reasoning over alternative future trajectories and prevents agents from evaluating the potential outcomes of different actions before execution. In contrast, humans often consider multiple possible consequences and mentally simulate alternative strategies before taking actions. This raises an important question: can embodied agents similarly leverage future world simulation to evaluate possible exploration paths and make more effective decisions? In this work, we propose ParallelWorld, a multi-horizon test-time scaling framework for active embodied reasoning. Starting from a restorable simulator state, ParallelWorld expands every retained branch with executable camera and task-dependent physical actions, and renders their prospective visual outcomes. A verifier agent evaluates the resulting frontier and prunes uninformative branches using a predefined branch-width schedule, which preserves multiple hypotheses during expansion and converges toward the most promising trajectory at later steps. The selected states are then replayed and expanded in the next iteration. Since physical execution cannot simultaneously follow alternative trajectories or roll back to previous states, an answer agent reasons only over the highest-ranked root-to-leaf route, producing the final prediction when its confidence becomes sufficiently high or the exploration budget is exhausted. In this way, ParallelWorld enables embodied agents to move beyond reactive single-trajectory exploration and make decisions using question-relevant evidence from simulated future worlds. Extensive experiments on embodied understanding benchmarks demonstrate that ParallelWorld significantly improves the capability of embodied agents in active reasoning. Our results validate the effectiveness of future world simulation as a mechanism for enhancing embodied understanding and provide a new perspective toward building more capable autonomous agents. Our main contributions can be summarized as follows: • We propose ParallelWorld, an multi-horizon test-time scaling framework that enables agents to explore potential future trajectories through simulated world interactions before execution. • We introduce a verifier-guided exploration mechanism with a verifier agent and an answer agent, where the former evaluates candidate trajectories and the latter performs reasoning over selected evidence for final decision-making. • Experimental results on embodied understanding benchmarks demonstrate that our method improves active perception and reasoning performance, validating the effectiveness of future trajectory simulation for embodied agents. 2 Related Work 2.1 Spatial Reasoning Spatial reasoning has evolved from geometric estimation and explicit scene representations to MLLM-based spatial question answering. Existing methods can be broadly divided into two categories. Geometry-enhanced approaches introduce depth supervision, 3D annotations, scene graphs, or reconstruction objectives to improve spatial grounding[3, 4, 31, 10]. Reasoning-enhanced approaches instead ground inference in explicit spatial coordinates, intermediate reasoning chains, or visual drawings[20, 32]. More recent studies extend spatial reasoning and evaluation from static observations to remembered spaces, real-world videos, and dynamic 4D scenes[34, 15, 29]. Despite these advances, most existing methods reason over fixed observations and therefore remain limited in their ability to actively acquire task-relevant spatial evidence. 2.2 Active Embodied Reasoning Active perception formulates sensing as a decision process in which an agent selects interactions to reduce uncertainty about task-relevant variables[2, 1, 17]. Embodied question answering further instantiates this principle by requiring agents to interact with an environment before producing an answer[9, 21]. Recent embodied agents extend this paradigm through confidence-aware exploration, goal-conditioned perception, multimodal tool use, spatial-belief construction, and retrospective trajectory selection[23, 43, 39, 42, 40]. ESI-Bench evaluates active embodied reasoning across diverse camera-motion and physical-manipulation tasks[14]. However, existing methods generally acquire evidence along a single realized trajectory, selecting each action from the current observation history. ParallelWorld differs from these methods by constructing multiple prospective trajectories and using a verifier to retain informative branches. 2.3 Test-Time Scaling Test-time scaling improves model performance by allocating additional inference computation to repeated sampling, verification, and structured search[28, 7, 19, 37, 33]. In agentic settings, language-model reasoning has been combined with environment feedback and action-level tree search[41], while compute-optimal allocation of test-time resources has also been studied more generally[26]. A complementary line of work uses explicit or learned models of the environment to evaluate prospective outcomes. Learned dynamics support imagination-based control and long-horizon planning[11, 12, 6, 24, 13], whereas classical approaches perform lookahead through POMDP inference, Monte Carlo tree search, or model-predictive control[18, 25, 30]. Recent multimodal systems further apply visual world models to spatial reasoning, exploration, and robotic action refinement[36, 38, 22, 27]. ParallelWorld is complementary to learned world-model approaches: the present implementation isolates verifier-guided search under an exact, restorable simulator transition, while replacing this transition with a learned world model remains an important direction for real-world deployment. 3 Methods 3.1 Overview Figure 1: Overview of ParallelWorld. Given a question, an initial observation, and the corresponding simulator state, ParallelWorld enumerates executable camera and physical actions from each retained world and renders their prospective outcomes. A verifier ranks the candidate worlds and applies a predefined branch-width schedule, retaining multiple branches at odd steps and converging to one branch at even steps. The selected simulator states are replayed and expanded in the next iteration. At each checkpoint, the answer agent reasons over the verifier’s top-1 root-to-leaf route and predicts the answer and confidence, terminating the exploration when the prediction is sufficiently confident or the exploration budget is exhausted. ParallelWorld is an active test-time scaling framework that improves embodied reasoning by imagining multiple possible exploration trajectories before committing to a particular strategy. Instead of generating a single action sequence autoregressively, our method constructs a tree of future worlds. At each exploration step, all retained worlds are expanded, and a verifier agent selects the most informative branches. An answer agent then reasons over the accumulated evidence from the selected branches and determines whether further exploration is necessary. The overall framework is illustrated in Figure 1. Given a question q, an initial environment state s0s_0, and an initial observation o0o_0, the objective is to predict an answer y through active interaction. We define the executable action space as =cam∪task,A=A_cam _task, where camA_cam contains camera translations and rotations, while taskA_task contains task-dependent physical interactions, such as picking, placing, pouring, and stacking. At step t, ParallelWorld maintains a set of retained trajectories ℬt=τt(1),…,τt(Kt),τt(i)=(a1(i),…,at(i)),B_t= \ _t^(1),…, _t^(K_t) \, _t^(i)= (a_1^(i),…,a_t^(i) ), where KtK_t denotes the branch retention width. Each trajectory corresponds to an independently restorable simulated world. 3.2 Prospective World Expansion A simulated branch must preserve both geometric and task-dependent information. We therefore represent its runtime state as ξt=(st,zt,pt,ωt), _t= (s_t,z_t,p_t, _t ), where sts_t is the simulator state, ztz_t is the task-specific interaction state, ptp_t is the camera pose, and ωt _t records the states of relevant objects. The root branch is initialized as ℬ0=(τ0,ξ0),τ0=∅.B_0= \( _0, _0) \, _0= . At exploration step t, each retained branch is expanded with every executable action. For a parent branch i and action a∈a , the prospective state and its corresponding observation are computed as ξt(i,a)=(ξt−1(i),a),ot(i,a)=(ξt(i,a)), _t^(i,a)=T ( _t-1^(i),a ), o_t^(i,a)=O ( _t^(i,a) ), where T denotes the simulator transition and O denotes visual rendering. The complete candidate frontier is t=(τt−1(i)⊕a,ξt(i,a),ot(i,a))|τt−1(i)∈ℬt−1,a∈.C_t= \ ( _t-1^(i) a, _t^(i,a),o_t^(i,a) )\; |\; _t-1^(i) _t-1,\ a \. Consequently, the number of simulated future worlds at step t is |t|=|ℬt−1|⋅||.|C_t|=|B_t-1|·|A|. Before simulating a candidate, its parent state is restored so that every action is evaluated from the same parent world. This operation is particularly important for physical actions, since object poses, grasp constraints, and task-specific variables may be modified during interaction. For actions that produce informative intermediate states, we additionally retain their intermediate visual frames. Thus, the evidence associated with a candidate c is e(c)=o(c),o1act(c),…,omcact(c),e(c)= \o(c),o^act_1(c),…,o^act_m_c(c) \, where o(c)o(c) is the post-action observation and mcm_c is the number of available intermediate frames. 3.3 Verifier-Guided Exploration Directly retaining the entire frontier would cause the number of trajectories to grow exponentially. ParallelWorld controls this growth using a verifier agent VϕV_φ, which evaluates the potential usefulness of each candidate for answering q. Its input consists of the question, the reference observations, the previous exploration history, and the visual evidence from the current candidate frontier: vc=Vϕ(q,ℐref,ℋt−1,τc,e(c)),c∈t.v_c=V_φ (q,I_ref,H_t-1, _c,e(c) ), c _t. Here, vcv_c denotes the verifier’s implicit assessment of the evidence quality of candidate c. The verifier favors branches that reveal task-relevant objects or relations, resolve visual ambiguity, and provide complementary rather than duplicated observations. The retained branch set is written as ℬt=Kc∈t(vc,Kt),|ℬt|=min(Kt,|t|).B_t=K_c _t (v_c,K_t ), |B_t|= (K_t,|C_t|). In practice, VϕV_φ directly outputs an ordered set of branch identifiers, and the above equation represents this set-valued selection process. The retention width is determined a predefined branch-width strategy: Kt=g(t,q),K_t=g(t,q), where g follows an alternating expansion-and-convergence strategy and assigns different retention widths according to the task category. Odd-numbered steps retain multiple branches to preserve diverse future hypotheses, whereas even-numbered steps set Kt=1K_t=1 to consolidate the search around the most informative trajectory. Larger widths are used during early expansion, followed by a smaller stable width at greater depths. Candidate simulations are initially used only for comparison. After verification, the selected actions are replayed from their parent snapshots: ξ^t(i)=(ξt−1π(i),at(i)),τt(i)∈ℬt, ξ_t^(i)=T ( _t-1^π(i),a_t^(i) ), _t^(i) _t, where π(i)π(i) denotes the parent of branch i. Only these selected states are stored for the next expansion step. This replay operation ensures that the retained images, object states, and simulator states remain mutually consistent. 3.4 Top-1 Route Answering In practical deployment, a robot can execute only one action sequence at a time and cannot simultaneously perform alternative physical interactions or roll back to a previous world state. We therefore distinguish internal prospective search from committed execution: the verifier may retain multiple simulator branches for future exploration, but the answer agent receives evidence from only one root-to-leaf route. This top-1 route corresponds to the single trajectory that would ultimately be executed by the robot and avoids combining observations from mutually incompatible worlds. Although the verifier retains KtK_t branches for internal exploration, only the highest-ranked route is provided to the answer agent. We treat the verifier output as an ordered branch set ℬt=(bt[1],bt[2],…,bt[Kt]),B_t= (b_t^[1],b_t^[2],…,b_t^[K_t] ), where bt[1]b_t^[1] denotes the top-ranked branch at step t. The remaining branches are preserved in the simulator and may be expanded in subsequent steps, but they do not contribute evidence to the current answer prediction. Let (b)P(b) denote the parent of branch b. Starting from the top-ranked branch bt[1]b_t^[1], we recover its complete ancestor chain: bt,j∗=t−j(bt[1]),j∈1,…,t.b_t,j^*=P^\,t-j (b_t^[1] ), j∈\1,…,t\. The top-1 trajectory at checkpoint t is therefore ρt∗=(bt,1∗,bt,2∗,…,bt,t∗),bt,t∗=bt[1]. _t^*= (b_t,1^*,b_t,2^*,…,b_t,t^* ), b_t,t^*=b_t^[1]. By construction, all elements of ρt∗ _t^* belong to the same root-to-leaf path: (bt,j∗)=bt,j−1∗,j=2,…,t.P (b_t,j^* )=b_t,j-1^*, j=2,…,t. The answer evidence contains one post-action observation from every depth along this route: ℰttop1=o(bt,j∗)j=1t.E_t^top1= \o (b_t,j^* ) \_j=1^t. Reference observations ℐrefI_ref are appended when available. Intermediate action frames may be used by the verifier to evaluate candidate outcomes, whereas the answer agent receives only the post-action observations along ρt∗ _t^*. Thus, evidence from other selected branches is excluded: o(b)∉ℰttop1,∀b∈ℬj∖bt,j∗.o(b) _t^top1, ∀ b _j \b_t,j^* \. At each checkpoint, the answer agent predicts (y^t,ct,rt)=Aθ(q,ℐref,ℰttop1), ( y_t,c_t,r_t )=A_θ (q,I_ref,E_t^top1 ), where y^t y_t is the predicted answer, ct∈[0,1]c_t∈[0,1] is the confidence, and rtr_t is the corresponding reasoning. Because the top-ranked leaf may originate from a branch that was not ranked first at an earlier step, ρt∗ _t^* is reconstructed from the current leaf rather than by concatenating the independently top-ranked branch from every checkpoint. This guarantees that all observations given to the answer agent correspond to one physically consistent trajectory. We define a deterministic validity indicator to distinguish admissible, conclusive predictions from malformed or unresolved responses. Let (q)Y(q) denote the task-specific answer space for question q, and let ⊥Y_ denote invalid or abstaining responses, such as “unknown,” “uncertain,” or outputs that cannot be parsed into the required answer format. We define Valid(y^t;q)=[y^t∈(q)∖⊥],Valid( y_t;q)=I [ y_t (q) _ ], where [⋅]I[·] is the indicator function. This validity check is implemented as a deterministic parser rather than an additional model call. Exploration terminates when the predicted answer is valid and its confidence exceeds a threshold γ: ct>γ∧Valid(y^t;q)=1.c_t>γ ( y_t;q)=1. Given a maximum exploration horizon L, the stopping checkpoint is T=min(t∈1,…,L:ct>γ∧Valid(y^t;q)=1∪L).T= ( \t∈\1,…,L\:c_t>γ ( y_t;q)=1 \∪\L\ ). If no earlier checkpoint satisfies the stopping condition, then T=LT=L by construction. At the stopping checkpoint, the answer agent produces (y^T,cT,rT)=Aθ(q,ℐref,ℰTtop1),( y_T,c_T,r_T)=A_θ (q,I_ref,E_T^top1 ), and the final prediction is y^=y^T. y= y_T. Overall, ParallelWorld implements verifier-guided tree search over simulated environment states. Prospective expansion explores alternative futures, verifier-based pruning concentrates computation on informative trajectories, and top-1 route answering converts the selected visual evidence into the final answer. 4 Experiments Table 1: Quantitative results on ESI-Bench [14]. We report accuracy across Passive Single-View, Active Exploration, ParallelWorld. All active methods use the same answer-model backbone. Category Subcategory Agent Evaluation Passive Active ParallelWorld View Hallucination 53.32% 66.82% 71.40% Partial Occlusion 36.84% 57.89% 77.89% Perceptual Grounding Material Transparency 69.72% 71.10% 73.85% Rigid Containment 55.00% 60.00% 80.00% Physical Structure Deformable Objects 37.76% 40.82% 42.86% Inclined Plane 59.02% 83.61% 85.25% Physical Dynamics Stacking & Stability 1.12% 45.98% 53.95% Reflection Authoring 56.57% 51.52% 60.61% Spatial Relations 38.94% 36.28% 51.33% Specular Reflection Correspondence 42.05% 47.73% 48.86% Linear Alignment 52.13% 68.09% 70.21% Geometric Configuration 16.55% 17.96% 18.31% Spatial Relations Physical Contact 66.39% 68.07% 71.43% Dimensional Size 39.52% 44.91% 49.10% Metric Comparison 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% Cognitive Mapping 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% Enumerative Perception Illumination Variability 0.00% 6.00% 10.00% Unobserved Change 38.62% 70.95% 89.86% Temporal Understanding Agent Observation 34.59% 48.87% 62.41% Action Sequencing Action Order Inference 37.66% 53.25% 61.04% We evaluate ParallelWorld on ESI-Bench [14] to investigate the following questions: 1. Does prospective world simulation improve embodied reasoning over sequential exploration? 2. How does the number of future worlds retained affect reasoning performance and test-time computation? We first describe the experimental setup in section 4.1 and report quantitative results across all ESI-Bench [14] task categories in section 4.2. We then present a qualitative example illustrating how the verifier identifies informative future worlds in section 4.3. Finally, we conduct ablation study on the verifier retention width of our framework in section 4.4. 4.1 Experimental Setup Benchmark. We conduct experiments on ESI-Bench [14], a comprehensive benchmark for embodied spatial intelligence. The benchmark contains 10 task categories and 29 subcategories. Due to an environment issue that prevents the execution of pouring actions, we exclude the Liquid Volume subcategory and evaluate the remaining 28 subcategories. These tasks require agents to actively acquire information through camera movement, physical interaction, or both before producing an answer. Baselines. We selected two paradigms from ESI-Bench [14] as baselines, including Passive Single-View, which predicts the answer from a single observation at the initial pose, and Active Exploration, which sequentially selects camera or physical actions to gather task-relevant evidence before answering. Implementation details. We use GPT-5.4 as the answer agent and GPT-5.5 as the verifier agent. By default, the maximum exploration depth is set to L=15L=15, and the confidence threshold for adaptive stopping is set to γ=0.8γ=0.8. The verifier is constrained to return candidate route identifiers, while the answer agent outputs the task answer, confidence, and reasoning. 4.2 Main Results Table 1 summarizes the quantitative results across the 28 evaluated ESI-Bench subcategories. ParallelWorld consistently outperforms the conventional Active Exploration baseline under the same answer-model backbone. The improvements are particularly pronounced for Temporal Understanding, where the accuracy on Unobserved Change increases from 70.95% to 89.86%, and for Metric Comparison, where Spatial Distance improves from 58.55% to 67.11%. Although Passive Single-View performs best on Connectivity due to the task setting, ParallelWorld still improves over Active Exploration from 55.00% to 60.00%. Overall, these results demonstrate the effectiveness of prospective simulation for active embodied reasoning. 4.3 Qualitative Analysis Figure 2: Qualitative analysis of verifier-guided exploration. We present a illustrative qualitative comparison on the exploration process of Active Exploration and ParallelWorld on an ESI-Bench [14] task. Gray views indicate unselected candidate branches, whereas colored views indicate branches retained by the verifier. As shown in Figure 2, the sequential active exploration baseline follows a single trajectory whose observations remain ambiguous and eventually predicts an incorrect answer. In contrast, ParallelWorld evaluates multiple candidate actions in simulated future worlds, as illustrated by the stacked views at each step. The verifier selects trajectories that progressively expose the relative positions of the lamp and projector. The selected evidence clearly indicates that the gap of the two objects, enabling the answer agent to produce the correct prediction. This example demonstrates how prospective simulation and verifier-guided selection improve evidence acquisition over single-trajectory exploration. 4.4 Ablation Studies Table 2: Ablation study on verifier branch width. We report accuracy (%), average runtime per question (s), and the mean number of exploration steps under fixed branch width K and the predefined K schedule settings. All configurations use the same answer model, maximum exploration depth, and stopping threshold. Subcategory K=2K=2 K=3K=3 K=4K=4 K Schedule Acc. Time Steps Acc. Time Steps Acc. Time Steps Acc. Time Steps Rigid Containment 55.00 4246.8 5.65 40.00 4091.8 5.15 57.89 3964.3 4.95 80.00 1542.6 4.20 Partial Occlusion 56.84 192.4 1.15 49.47 168.6 1.08 55.79 178.0 1.08 63.16 158.3 1.09 Regional Boundary 66.15 285.0 2.20 65.15 486.2 2.56 56.41 315.5 1.97 61.25 290.5 3.72 Counting with Occlusion 23.33 230.8 1.63 23.33 456.7 3.33 23.33 553.5 3.33 30.00 325.2 3.80 Average 55.52 590.8 1.99 50.72 668.7 2.27 51.87 613.5 2.04 59.56 350.6 2.66 As illustrated in Table 2, we compare fixed K=2,3,4K=2,3,4 search with our predefine K schedule. In our schedule setting, the verifier retains a wider set of branches during expansion steps and progressively converges to one branche at later steps, with the schedule adjusted according to the task category. This design balances trajectory diversity ,search concentration and time consumption. The schedule strategy achieves the best average accuracy of 59.56%59.56\%, compared with 55.52%55.52\%, 50.72%50.72\%, and 51.87%51.87\% for K=2,3,4K=2,3,4, respectively. It also has the lowest average runtime (350.6350.6 s per question), despite using a slightly larger average number of exploration steps. The results also show that increasing the fixed branch width does not consistently improve accuracy, since retaining additional branches may introduce redundant or less informative candidates. Our schedule setting does not outperform every fixed setting on every task; for example, K=2K=2 and K=3K=3 perform better on Regional Boundary. Nevertheless, it provides the strongest overall accuracy-efficiency trade-off, achieving the best mean performance across the evaluated subcategories. 5 Conclusion We presented ParallelWorld, an active test-time scaling framework for embodied reasoning. ParallelWorld constructs multiple future worlds through simulated interactions, uses a verifier agent to retain informative exploration trajectories, and employs an answer agent to reason over the selected evidence. Experiments on ESI-Bench show consistent improvements over sequential Active Exploration across the evaluated task categories, with particularly clear gains in temporal understanding and metric comparison. These results demonstrate that prospective world simulation provides an effective mechanism for improving active embodied reasoning. Limitations and Future Work. ParallelWorld incurs additional test-time computation because each retained world must be expanded over the executable action space. This cost can become substantial for tasks with many physical actions or long exploration horizons. Moreover, the quality of exploration depends on both simulator fidelity and verifier reliability, while retaining only selected trajectories may discard complementary evidence from alternative branches. Future work could investigate learned branch-pruning and value estimation, uncertainty-aware verifier models, and hierarchical action proposal strategies to improve the efficiency and robustness of prospective exploration. Extending ParallelWorld to real-world robots and dynamic environments is another important direction. References [1] J. Aloimonos, I. Weiss, and A. Bandyopadhyay (1988) Active vision. International Journal of Computer Vision 1 (4), p. 333–356. External Links: Document Cited by: §2.2. [2] R. Bajcsy (1988) Active perception. Proceedings of the IEEE 76 (8), p. 966–1005. External Links: Document Cited by: §2.2. [3] B. Chen, Z. Xu, S. Kirmani, B. Ichter, D. Sadigh, L. Guibas, and F. Xia (2024) SpatialVLM: endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 14455–14465. External Links: Link Cited by: §2.1. [4] A. Cheng, H. Yin, Y. Fu, Q. Guo, R. Yang, J. Kautz, X. Wang, and S. Liu (2024) SpatialRGPT: grounded spatial reasoning in vision-language models. In Advances in Neural Information Processing Systems, Vol. 37, p. 135062–135093. External Links: Link Cited by: §2.1. [5] Z. Cheng, R. Li, J. Hu, Y. Tu, S. Dai, S. Hu, Y. Shi, L. Shi, and M. Sun (2026) Embodiedeval: evaluate multimodal llms as embodied agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11420–11432. Cited by: §1. [6] K. Chua, R. Calandra, R. McAllister, and S. Levine (2018) Deep reinforcement learning in a handful of trials using probabilistic dynamics models. In Advances in Neural Information Processing Systems, Vol. 31, p. 4754–4765. External Links: Link Cited by: §2.3. [7] K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman (2021) Training verifiers to solve math word problems. External Links: 2110.14168, Link Cited by: §2.3. [8] G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen, et al. (2025) Gemini 2.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Cited by: §1. [9] A. Das, S. Datta, G. Gkioxari, S. Lee, D. Parikh, and D. Batra (2018) Embodied question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 1–10. External Links: Link Cited by: §2.2. [10] Z. Fan, J. Zhang, R. Li, J. Zhang, R. Chen, H. Hu, K. Wang, P. Wang, H. Qu, S. Zhou, et al. (2026) VLM-3R: vision-language models augmented with instruction-aligned 3D reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 31054–31065. Cited by: §2.1. [11] D. Ha and J. Schmidhuber (2018) World models. External Links: 1803.10122, Link Cited by: §2.3. [12] D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi (2020) Dream to control: learning behaviors by latent imagination. In International Conference on Learning Representations, External Links: Link Cited by: §2.3. [13] D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap (2025) Mastering diverse control tasks through world models. Nature 640, p. 647–653. External Links: Document, Link Cited by: §2.3. [14] Y. Hong, J. Liu, H. Yin, M. Li, L. Guibas, F. Li, J. Wu, and Y. Choi (2026) ESI-Bench: towards embodied spatial intelligence that closes the perception-action loop. External Links: 2605.18746, Link Cited by: §1, §2.2, Figure 2, Figure 2, §4.1, §4.1, Table 1, Table 1, §4, §4. [15] Y. Huang, K. Wen, R. Gao, D. Liu, Y. Lou, J. Wu, J. Xu, J. Zhang, Z. Yang, Y. Lin, et al. (2026) Thinking in dynamics: how multimodal large language models perceive, track, and reason dynamics in the physical 4D world. arXiv preprint arXiv:2603.12746. Cited by: §2.1. [16] A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al. (2024) Gpt-4o system card. arXiv preprint arXiv:2410.21276. Cited by: §1. [17] S. Isler, R. Sabzevari, J. Delmerico, and D. Scaramuzza (2016) An information gain formulation for active volumetric 3d reconstruction. In 2016 IEEE International Conference on Robotics and Automation (ICRA), p. 3477–3484. External Links: Document Cited by: §2.2. [18] L. P. Kaelbling, M. L. Littman, and A. R. Cassandra (1998) Planning and acting in partially observable stochastic domains. Artificial Intelligence 101 (1–2), p. 99–134. External Links: Document Cited by: §2.3. [19] H. Lightman, V. Kosaraju, Y. Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe (2024) Let’s verify step by step. In International Conference on Learning Representations, External Links: Link Cited by: §2.3. [20] Y. Liu, D. Chi, S. Wu, Z. Zhang, Y. Hu, L. Zhang, Y. Zhang, S. Wu, T. Cao, G. Huang, et al. (2025) SpatialCoT: advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074. External Links: Link Cited by: §2.1. [21] A. Majumdar, A. Ajay, X. Zhang, P. Putta, S. Yenamandra, M. Henaff, S. Silwal, P. McVay, O. Maksymets, S. Arnaud, K. Yadav, Q. Li, B. Newman, M. Sharma, V. Berges, S. Zhang, P. Agrawal, Y. Bisk, D. Batra, M. Kalakrishnan, F. Meier, C. Paxton, S. Sax, and A. Rajeswaran (2024) OpenEQA: embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 16488–16498. External Links: Document, Link Cited by: §1, §2.2. [22] C. Qian, E. C. Acikgoz, B. Li, X. Chen, Y. Zhang, B. He, Q. Luo, G. Tur, D. Hakkani-Tür, Y. Li, and H. Ji (2026) Current agents fail to leverage world model as tool for foresight. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), San Diego, California, United States, p. 13686–13723. External Links: Document, Link Cited by: §2.3. [23] A. Z. Ren, J. Clark, A. Dixit, M. Itkina, A. Majumdar, and D. Sadigh (2024) Explore until confident: efficient exploration for embodied question answering. External Links: 2403.15941, Link Cited by: §1, §2.2. [24] J. Schrittwieser, I. Antonoglou, T. Hubert, K. Simonyan, L. Sifre, S. Schmitt, A. Guez, E. Lockhart, D. Hassabis, T. Graepel, T. Lillicrap, and D. Silver (2020) Mastering atari, go, chess and shogi by planning with a learned model. Nature 588 (7839), p. 604–609. External Links: Document Cited by: §2.3. [25] D. Silver and J. Veness (2010) Monte-carlo planning in large POMDPs. In Advances in Neural Information Processing Systems, Vol. 23, p. 2164–2172. External Links: Link Cited by: §2.3. [26] C. V. Snell, J. Lee, K. Xu, and A. Kumar (2025) Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In International Conference on Learning Representations, External Links: Link Cited by: §2.3. [27] Z. Sun, Y. Sun, H. Huang, and A. Knoll (2026) ω-EVA: Envision, Verify, and Act with Latent Interactive World Models. External Links: 2606.09457, Link Cited by: §2.3. [28] X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou (2023) Self-consistency improves chain-of-thought reasoning in language models. In International Conference on Learning Representations, External Links: Link Cited by: §2.3. [29] K. Wen, R. Chen, H. Zheng, Y. Lin, P. Pan, C. Li, W. Cong, J. Zhang, J. Lu, C. Lin, et al. (2026) DynamicVerse: a physically-aware multimodal framework for 4d world modeling. In Advances in Neural Information Processing Systems, Vol. 38, p. 108604–108634. Cited by: §2.1. [30] G. Williams, N. Wagener, B. Goldfain, P. Drews, J. M. Rehg, B. Boots, and E. A. Theodorou (2017) Information theoretic MPC for model-based reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), p. 1714–1721. External Links: Document Cited by: §2.3. [31] D. Wu, F. Liu, Y. Hung, and Y. Duan (2025) Spatial-MLLM: boosting MLLM capabilities in visual-based spatial intelligence. In Advances in Neural Information Processing Systems, Vol. 38, p. 13569–13597. Cited by: §2.1. [32] J. Wu, J. Guan, K. Feng, Q. Liu, S. Wu, L. Wang, W. Wu, and T. Tan (2025) Reinforcing spatial reasoning in vision-language models with interwoven thinking and visual drawing. In Advances in Neural Information Processing Systems, Vol. 38, p. 143297–143330. Cited by: §2.1. [33] Y. Xie, K. Kawaguchi, Y. Zhao, X. Zhao, M. Kan, J. He, and Q. Xie (2023) Self-evaluation guided beam search for reasoning. In Advances in Neural Information Processing Systems, Vol. 36, p. 41618–41650. External Links: Link Cited by: §2.3. [34] J. Yang, S. Yang, A. W. Gupta, R. Han, L. Fei-Fei, and S. Xie (2025) Thinking in space: how multimodal large language models see, remember, and recall spaces. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 10632–10643. Cited by: §2.1. [35] R. Yang, H. Chen, J. Zhang, M. Zhao, C. Qian, K. Wang, Q. Wang, T. V. Koripella, M. Movahedi, M. Li, et al. (2025) Embodiedbench: comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. arXiv preprint arXiv:2502.09560. Cited by: §1. [36] Y. Yang, J. Liu, Z. Zhang, S. Zhou, R. Tan, J. Yang, Y. Du, and C. Gan (2025) MindJourney: test-time scaling with world models for spatial reasoning. In Advances in Neural Information Processing Systems, Vol. 38. External Links: 2507.12508, Link Cited by: §2.3. [37] S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y. Cao, and K. R. Narasimhan (2023) Tree of thoughts: deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, Vol. 36, p. 11809–11822. External Links: Link Cited by: §2.3. [38] S. Yu, Y. Zhang, Z. Wang, J. Yoon, H. Yao, M. Ding, and M. Bansal (2026) When and how much to imagine: adaptive test-time scaling with world models for visual spatial reasoning. External Links: 2602.08236, Link Cited by: §2.3. [39] G. Zhang, M. Ding, J. Wu, R. Liao, and V. Tresp (2025) ReEXplore: improving MLLMs for embodied exploration with contextualized retrospective experience replay. External Links: 2511.19033, Link Cited by: §1, §2.2. [40] P. Zhang, Z. Huang, Y. Wang, J. Zhang, L. Xue, Z. Wang, Q. Wang, K. Chandrasegaran, R. Zhang, Y. Choi, et al. (2026) Theory of space: can foundation models construct spatial beliefs through active exploration?. In The Fourteenth International Conference on Learning Representations, Cited by: §1, §2.2. [41] A. Zhou, K. Yan, M. Shlapentokh-Rothman, H. Wang, and Y. Wang (2024) Language agent tree search unifies reasoning, acting, and planning in language models. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 62138–62160. External Links: Link Cited by: §2.3. [42] W. Zhou, X. Xiong, Y. Peng, M. Tao, C. Zhao, H. Dong, M. Tang, and J. Wang (2025) PhysVLM-AVR: active visual reasoning for multimodal large language models in physical environments. External Links: 2510.21111, Link Cited by: §1, §2.2. [43] M. Zhu, H. Zhong, C. Zhao, Z. Du, Z. Huang, M. Liu, H. Chen, C. Zou, J. Chen, M. Yang, and C. Shen (2025) Active-O3: empowering multimodal large language models with active perception via GRPO. External Links: 2505.21457, Link Cited by: §1, §2.2.