Soft Spatial Reasoning

Rafi Ibn Sultan; Md. Sajid Alam Chowdhury; Saleh Zare Zade; Chengyin Li; Prashant Khanduri; Marco Brocanelli; Dongxiao Zhu

cs.CVSep 30, 2026

Soft Spatial Reasoning

Authors: Rafi Ibn Sultan, Md. Sajid Alam Chowdhury, Saleh Zare Zade, Chengyin Li, Prashant Khanduri, Marco Brocanelli, Dongxiao Zhu

Organizations: Department of Computer Science, Wayne State University · Department of Radiation Oncology, Henry Ford Health · Department of Electrical and Computer Engineering, The Ohio State University · Institute for AI and Data Science, Wayne State University

Abstract

Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain. This early commitment constitutes premature discretization: an incorrect token selection can propagate errors through subsequent reasoning. We propose Soft Spatial Reasoning, a post-training framework that introduces soft thinking for spatial tasks in LVLMs. At each intermediate reasoning step, the LVLM forms a continuous soft state by mixing token embeddings rather than selecting a single token, allowing multiple candidate continuations to influence the next step. The appropriate degree of softness, however, can vary across reasoning steps: retaining multiple candidates may preserve a useful spatial interpretation, but if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. At the core of Soft Spatial Reasoning is AdaptSoft, a controller that uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step. To train AdaptSoft, we introduce a gradient-alignment learning objective that provides a step-specific learning signal for softness control without intermediate reasoning supervision. Across diverse spatial benchmarks, Soft Spatial Reasoning outperforms hard and fixed-soft CoT baselines using the same backbone, as well as a range of existing LVLMs. The source code is available at https://github.com/rafiibnsultan/Soft_Spatial_Reasoning

Figures & tables

Figure 1: Hard versus adaptive soft thinking for spatial reasoning. (a) Hard thinking commits to one token at each CoT step, allowing early errors to propagate. (b) Our Soft Spatial Reasoning forms continuous intermediate states by mixing token embeddings, carrying information from multiple candidate continuations through the CoT. AdaptSoft adjusts the mixture’s softness at each step based on the current reasoning state and predictive uncertainty.

Figure 2: Overview of Soft Spatial Reasoning . (a) The LVLM policy is post-trained to carry multiple candidate continuations through soft states during reasoning, then generate the final answer as discrete tokens. Rollout rewards yield group-relative advantages for the GRPO update. (b) At each reasoning step, AdaptSoft sets the mixture temperature from the current hidden state and candidate-distribution entropy. The resulting mixture weights combine token embeddings into the next soft state.

Figure 3: Gradient-Alignment Learning. At each soft reasoning step

t

, the temperature

\tau_{i,t}

controls the state

s_{i,t}

, whose contribution

g_{i,t}

to the LVLM output-layer gradient is compared with a reference gradient

G_{\mathrm{ref}}

computed from a disjoint subset of rollouts. The resulting alignment scores

\alpha_{i,t}

are aggregated to form the AdaptSoft loss.

Reference Baselines
Method	Average	Dynamic Reasoning		Spatial Interaction			Complex Logic		Perspective Taking
Method	Average	Mani- pulation	Motion Anal.	Traffic Anal.	Loca- lization	Geospa. Strategy	Pattern Rec.	Geometric Reasoning	Ego Centric	Allo Centric	Hypo- thetical
Random Choice	24.98	24.86	26.30	25.88	23.43	27.27	21.44	24.77	22.55	24.84	25.78
Human Evaluation	92.63	94.62	96.07	91.38	95.11	92.15	89.02	85.90	98.53	94.30	90.26
Proprietary Models
GPT-4.1-mini Open (2025)	48.87	64.32	56.53	59.06	60.19	56.36	29.28	30.19	72.55	39.57	39.28

Table 1: OmniSpatial Jia et al. (2025) results across 10 spatial reasoning task categories. Soft Spatial Reasoning is compared against proprietary models, general open-source LVLMs, soft thinking models, and specialized spatial reasoning models. Proprietary models are included as reference points; the best open-source or specialized result is shown in bold . Average accuracy is weighted by category sample size.

Figure 4: Soft Spatial Reasoning on a randomly selected example from OmniSpatial. AdaptSoft varies softness across the CoT. For readability, greedy decoding is used only to display the soft states as text. Shading shows each span’s mean temperature ( blue : lower; orange : higher). The plot below shows temperature at every reasoning step and labels several words with their temperatures.

Reference Baselines
Method	Average	Question Categories
Method	Average	3D Geom.	Dep. & Occu.	Orientation	Relat. Posit.	Size & Scale	Spati. Navig.
Random Choice	25.00	25.00	25.00	25.00	25.00	25.00	25.00
Human Baseline	87.57	93.70	74.13	91.58	91.51	88.89	87.76
Proprietary Models
GPT-4o-mini Hurst et al. (2024)	46.50	47.06	39.00	47.03	47.17	49.60	49.79

Table 2: SpatiaLab Wasi et al. (2026) results across 6 spatial reasoning task categories in the zero-shot setting. Proprietary models are included as reference points; the best open-source or specialized result is shown in bold . Average accuracy is weighted by category sample size.

Reference Baseline
Method	Overall	Rotation	Among	Around
Random Choice	32.35	36.36	32.29	30.66
Proprietary Models
Gemini-2.5-Pro Team et al. (2023)	47.05	85.50	25.95	38.40
Claude-4-Sonnet Anthropic (2024)	44.75	48.42	44.21	47.62
Open-weight Models

Table 3: Zero-shot results on MindCube Wang et al. (2026a) across three spatial mental-modeling settings. Proprietary models are included as reference points, and Overall is weighted by setting size.

Variant	Avg.	Ego centric	Allo centric	Hypo thetical
Soft Spatial Reasoning (Full)	42.75	81.37	32.53	41.46
w/o $h_{i,t}$ in AdaptSoft	39.22	77.45	29.52	36.14
w/o predictive uncertainty	40.64	78.43	30.85	38.55
w/o gradient alignment	38.68	76.47	29.26	34.94
w/o gradient centering	41.18	79.41	31.12	39.76

Table 4: OmniSpatial Perspective Taking ablations Jia et al. (2025) . All variants share the backbone and policy GRPO setup, with post-training and evaluation on the corresponding training and test splits. Averages are sample-weighted.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: AdaptSoft Temperatures During Training. The line shows the mean temperature $\tau_{i,t}$ across soft steps in each rollout batch; shading spans their minimum and maximum. The dashed line marks $\tau_{0}=0.5$ .

Figure 6: AdaptSoft Temperature and Predictive Uncertainty. Temperatures and predictive uncertainty are measured across soft reasoning steps during inference on the test set. Color shows the number of steps on a logarithmic scale. The solid line shows median temperature in each entropy bin; dashed lines mark the 10th and 90th percentiles.

Figure 7: AdaptSoft Varies Softness Within CoTs. The histogram shows the 90th–10th percentile temperature spread within each test-set reasoning trace generated during inference. The solid line marks the median within-trace spread; the dashed line marks the corresponding spread across trace-median temperatures.

Hyperparameter	Value	Notes
Model configuration
Backbone	Qwen3-VL-8B-Thinking	–
Language-model adaptation	Full fine-tuning	No LoRA
Numerical precision	BF16	FSDP2 mixed precision
Parameter sharding	FSDP2	Parameter and optimizer offload
Data preparation

Appendix

Table 5: Post-training and evaluation configuration for Soft Spatial Reasoning.

Figure 8: System prompt used to standardize the reasoning format during Soft Spatial Reasoning training.

Explore similar work

SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models
Sep 30, 2026Rafi Ibn Sultan, Xiangyu Zhou, Md. Sajid Alam Chowdhury +4Spatial Reasoning Recent Vision-Language Models
Department of Computer Science, Wayne State University · Department of Radiation Oncology, Henry Ford Health · Department of Electrical and Computer Engineering, The Ohio State University +1
ProSR: Process-Shaped Spatial Reasoning for Reliable Chain-of-Thought in VLMs
May 25, 2026Jiangyang Li, Cong Wan, Changjie Wu +8Spatial Reasoning Reasoning Trajectory
1Xi’an Jiaotong University · 2Amap, Alibaba Group · 4Shenzhen University +1
Reinforcing Dual-Path Reasoning in Spatial Vision Language Models
Jun 16, 2026Yatai Ji, An-Chieh Cheng, Yang Fu +13Spatial Reasoning Indoor Localization
The University of Hong Kong · NVIDIA · University of California, San Diego

Sep 30, 2026, cs.CVJ/K move · Enter open · S save

SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models

Rafi Ibn Sultan, Xiangyu Zhou, Md. Sajid Alam Chowdhury, Chengyin Li+3

Department of Computer Science, Wayne State University · Department of Radiation Oncology, Henry Ford Health · Department of Electrical and Computer Engineering, The Ohio State University+1

Spatial Reasoning Recent Vision-Language Models Visual Perception

Open paper AskPDF arXiv

Large Vision-Language Models (LVLMs) have made remarkable progress across visual perception tasks, yet spatial reasoning remains a persistent weakness, especially for questions that require reasoning over visual space. Recent spatial-reasoning methods incorporate generated grounding, where models predict bounding boxes, masks, or other localization outputs for task-relevant objects as part of their reasoning trace. However, these approaches typically optimize final-answer correctness alone, allowing correct answers to be rewarded even when the model does not reason from confidently localized task-relevant objects. We introduce SpatialCORE (Spatially COnfident REasoning), a post-training framework that turns the model's own confidence in generated grounding into a learning signal for spatial reasoning. Its central idea is to reinforce grounding that is both accurate and confident, encouraging the model to reason from confidently localized task-relevant objects. SpatialCORE realizes this through a self-regulating spatial reward that weights each predicted bounding box's matching quality by its coordinate-token confidence. An answer gate further ties grounding optimization to final-answer correctness. SpatialCORE achieves state-of-the-art results among open-source and specialized spatial reasoning models across diverse benchmarks, and transfers effectively in zero-shot settings to unseen data distributions. The source code is available at https://github.com/rafiibnsultan/SpatialCORE.

Abstract

Explore similar work

SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models

ProSR: Process-Shaped Spatial Reasoning for Reliable Chain-of-Thought in VLMs

Reinforcing Dual-Path Reasoning in Spatial Vision Language Models