Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination

Shenglan Li; Zhendong Mi; Hengyi Zhu; Jingwu Luo; Chun Kit Chan; Geng Yuan; Yanzhi Wang; Pu Zhao

cs.CVOct 2, 2026

Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination

Authors: Shenglan Li, Zhendong Mi, Hengyi Zhu, Jingwu Luo, Chun Kit Chan, Geng Yuan, Yanzhi Wang, Pu Zhao, +1 more

Organizations: Stevens Institute of Technology · Northeastern University · Arizona State University · University of Georgia

Abstract

Vision-language-action (VLA) models increasingly incorporate intermediate reasoning to improve robotic manipulation, yet existing approaches primarily reason about observed states without explicitly anticipating future scene evolution. Extending such reasoning to explicit future rollouts at every inference step, however, introduces substantial computational overhead. We propose IG-VLA, a VLA reasoning framework that enables models to imagine the future and internalize the gist. Our Latent Spatiotemporal Reasoning learns to imagine task-relevant future scene evolution directly in visual representation space, guiding action prediction without costly pixel-level video generation. To further reduce inference overhead, we introduce Scene Gist Memory, which internalizes reasoning-derived scene-behavior associations into a compact Scene Gist Token, preserving the benefits of future reasoning while bypassing explicit future imagination at inference. Extensive experiments on LIBERO, LIBERO-Plus, and VLABench demonstrate the effectiveness and efficiency of IG-VLA. On the LIBERO-Plus Language suite, both the reasoning and gist policies outperform the strongest baseline by nearly 6% in success rate. The gist policy also achieves up to 6.38x speedup over baselines, reducing inference latency from 1081ms to 169.5ms per action chunk on a single NVIDIA A6000 GPU. These results demonstrate that future spatiotemporal reasoning can be effectively internalized for efficient VLA deployment.

Figures & tables

Figure 1: Language-based vs. latent visual reasoning. (a) Language-based reasoning may lose fine-grained spatial information, leading to imprecise actions; (b) Latent visual reasoning preserves evolving scene geometry through visual future reasoning, enabling more precise actions.

Figure 2: Human Automaticity vs. Scene Gist Memory.

Figure 3: Overview of our framework. (a) the overall architecture; (b) the latent spatiotemporal reasoning module learns to predict future visual keyframes; (c) the gist modules internalize reasoning-derived behavioral information into a compact Scene Gist Token.

Methods	Spatial		Object		Goal		Long		Avg.
Methods	SR $\uparrow$	Rank $\downarrow$	SR $\uparrow$	Rank $\downarrow$	SR $\uparrow$	Rank $\downarrow$	SR $\uparrow$	Rank $\downarrow$	SR $\uparrow$	Rank $\downarrow$
Diffusion Policy	78.3	29	92.5	21	68.3	30	50.5	30	72.4	30
Octo	78.9	28	85.7	29	84.6	24	51.1	29	75.1	28
CoT-VLA	87.5	23	91.6	23	87.6	22	69.0	22	81.1	23
WorldVLA (8) (256 $\times$ 256)	85.6	25	89.0	26	82.6	26	59.0	25	79.1	24
WorldVLA (8) (512 $\times$ 512)	87.6	22	96.2	18	83.4	25	60.0	24	81.8	22

Table 1: Comparison on the LIBERO benchmark. * represents that the LLM backbone is frozen during training. Our proposed approach is trained on the LIBERO dataset. All metrics are average success rates (%). The best results are highlighted in bold .

\triangle

denotes IG-VLA with Latent Spatiotemporal Reasoning replaced by Scene Gist Memory.

Methods	Camera	Robot	Language	Light	Background	Noise	Layout	Avg.
$\pi_{0}^{\diamond}$	79.6	21.1	72.5	84.7	86.2	68.3	69.4	67.4
$\pi_{0.5}^{\diamond}$	70.3	41.7	81.1	97.3	94.6	71.8	84.9	75.7
ACoT-VLA ⋄	91.2	62.5	80.3	95.1	91.5	88.3	84.9	84.1
ACoT-VLA	96.6	70.4	79.7	95.1	97.1	95.9	85.0	88.0
IG-VLA (Ours)	95.4	68.9	85.6	97.5	97.2	94.9	85.5	88.5
IG-VLA (Ours) △	95.3	69.0	85.1	97.4	97.5	94.7	84.6	88.3

Table 2: Supervised fine-tuning results on LIBERO-Plus.

\diamond

denotes a pipeline with a frozen LLM backbone, and

\triangle

denotes IG-VLA with Latent Spatiotemporal Reasoning replaced by Scene Gist Memory. Best results are in bold .

Methods	Guidance	In-dist.		Category		Commonsense		Instruction		Texture		Avg.
Methods	Guidance	IS $\uparrow$	PS $\uparrow$	IS $\uparrow$	PS $\uparrow$	IS $\uparrow$	PS $\uparrow$	IS $\uparrow$	PS $\uparrow$	IS $\uparrow$	PS $\uparrow$	IS $\uparrow$	PS $\uparrow$
$\pi_{0}^{\diamond}$	Linguistics	67.8	62.7	44.0	33.6	54.9	43.0	58.0	38.7	50.6	42.5	55.0	44.1
$\pi_{0.5}^{\diamond}$	Linguistics	75.0	60.8	49.6	35.3	57.5	41.6	57.1	30.3	62.0	47.4	60.2	43.1
ACoT-VLA	Action	79.8	66.1	54.1	38.9	52.3	37.8	56.8	39.6	74.6	54.6	63.5	47.4
IG-VLA (Ours)	Action	80.80	61.64	54.20	32.21	58.20	38.97	65.20	48.49	74.99	49.23	66.68	46.11
IG-VLA (Ours) △	Action	80.48	61.23	54.35	31.75	57.82	39.18	64.76	48.22	74.71	48.88	66.26	45.68

Table 3: Comparison on VLABench. IS and PS denote Intention Score and Progress Score.

\diamond

denotes pipeline with a frozen LLM backbone, and

\triangle

denotes IG-VLA with Latent Spatiotemporal Reasoning replaced by Scene Gist Memory. Best results are in bold .

Table 7

Figure 4: GPU time breakdown(batch size 1, BF16), averaged over three profiled forward passes.

Figure 7: Real-world evaluation on the dual-arm UR3. (a) Pick-and-place-a-screwdriver task execution. (b) Success rate comparison with baselines over 16 trials per method.

Explore similar work

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation
Jun 16, 2026Tianyi Lu, Hui Zhang, Zijie Diao +8Long-Horizon Robotic Manipulation Language-Conditioned Robot Manipulation
Fudan University · Shanghai Innovation Institue · China Unicom +1
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference
Jul 14, 2026Yuzhou Wu, Yuxin Zheng, Muchun Niu +6Efficient VLA Model Inference Efficient VLA Models
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2
MemoryVLA++: Temporal Modeling via Memory and Imagination in Vision-Language-Action Models
Jun 8, 2026Hao Shi, Weiye Li, Bin Xie +6Long-Horizon Robotic Manipulation World Models for Robotics
Department of Automation, BNRist, Tsinghua University, Beijing, China · Department of Computer Science, The University of Hong Kong, Hong Kong, China · Dexmal, Beijing, China +1

Jun 16, 2026, cs.ROJ/K move · Enter open · S save

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation

Tianyi Lu, Hui Zhang, Zijie Diao, Junke Wang+7

Fudan University · Shanghai Innovation Institue · China Unicom+1

Long-Horizon Robotic Manipulation Language-Conditioned Robot Manipulation VLM Reasoning Vision-Language-Action Models Visually Grounded Reasoning Reasoning for VLA Models

Open paper AskPDF arXiv

Most Vision-Language-Action (VLA) models map observations directly to actions without explicit reasoning, limiting their capacity for reasoning-intensive long-horizon tasks. To address this, existing approaches adopt Chain-of-Thought (CoT) reasoning to enable subgoal decomposition and spatial anticipation. However, those methods lack a unified architecture for effective cross-modal reasoning and fail to explicitly include inverse reasoning ability based on the target state. We argue that manipulation planning naturally decomposes into prediction, anticipating the next visual state, and inverse dynamics, inferring the actions to reach it. Bridging both requires a unified autoregressive architecture that interleaves textual and visual reasoning in a single generation process. We propose \textbf{ThinkingVLA}, a generative model that realizes this decomposition within a unified Mixture-of-Transformers architecture. ThinkingVLA consists of a forward CoT that identifies the immediate subgoal and guides the visual forecasting; the predicted image then serves as the target state, grounding an inverse CoT that reasons about spatial relationships and action intent based on the predicted image; and the final action is generated conditioned on this full reasoning context. Extensive experiments on simulation and real-world benchmarks demonstrate that ThinkingVLA consistently outperforms state-of-the-art baselines, with particularly large gains on long-horizon manipulation tasks.

Abstract

Explore similar work

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

MemoryVLA++: Temporal Modeling via Memory and Imagination in Vision-Language-Action Models