Soft Spatial Reasoning
Summary
The paper proposes Soft Spatial Reasoning, a post-training framework that introduces adaptive "soft thinking" for spatial tasks in Large Vision-Language Models, where intermediate reasoning steps mix token embeddings instead of committing to a single discrete token. An AdaptSoft controller dynamically adjusts softness based on hidden state and predictive uncertainty, outperforming hard and fixed-soft CoT baselines on spatial benchmarks.
View Cached Full Text
Cached at: 10/01/26, 04:23 PM
Paper page - Soft Spatial Reasoning
Source: https://huggingface.co/papers/2609.38717
Abstract
LargeVision-LanguageModels(LVLMs)commonlyperformspatialreasoningthroughchain-of-thought(CoT),encodingintermediatereasoningasautoregressivesequencesofdiscretelanguagetokens.Suchhardthinkingrequirescommittingtoasingletokenateachstep,evenwhenthecorrectspatialinterpretationremainsuncertain.Thisearlycommitmentconstitutesprematurediscretization:anincorrecttokenselectioncanpropagateerrorsthroughsubsequentreasoning.WeproposeSoftSpatialReasoning,apost-trainingframeworkthatintroducessoftthinkingforspatialtasksinLVLMs.Ateachintermediatereasoningstep,theLVLMformsacontinuoussoftstatebymixingtokenembeddingsratherthanselectingasingletoken,allowingmultiplecandidatecontinuationstoinfluencethenextstep.Theappropriatedegreeofsoftness,however,canvaryacrossreasoningsteps:retainingmultiplecandidatesmaypreserveausefulspatialinterpretation,butifthosecandidatesimplyconflictingspatialrelations,mixingthemmayinterferewithsubsequentreasoning.AtthecoreofSoftSpatialReasoningisAdaptSoft,acontrollerthatusesthecurrenthiddenstateandpredictiveuncertaintytoadaptthedegreeofsoftnessateachreasoningstep.TotrainAdaptSoft,weintroduceagradient-alignmentlearningobjectivethatprovidesastep-specificlearningsignalforsoftnesscontrolwithoutintermediatereasoningsupervision.Acrossdiversespatialbenchmarks,SoftSpatialReasoningoutperformshardandfixed-softCoTbaselinesusingthesamebackbone,aswellasarangeofexistingLVLMs.Thesourcecodeisavailableathttps://github.com/rafiibnsultan/Soft_Spatial_Reasoning
View arXiv pageView PDFGitHub0Add to collection
Get this paper in your agent:
hf papers read 2609\.38717
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.38717 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.38717 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.38717 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Reinforcing Dual-Path Reasoning in Spatial Vision Language Models
This paper introduces SR-REAL, a unified framework for spatial vision-language models that combines linguistic deduction and 3D geometric reasoning via reinforcement learning, enabling robust multi-step spatial reasoning across diverse tasks.
SpatialCORE: Confidence-Aware Grounded Spatial Reasoning in Large Vision--Language Models
SpatialCORE is a post-training framework that uses a model's confidence in its generated bounding-box grounding as a learning signal for spatial reasoning in large vision-language models, achieving state-of-the-art results among open-source models on spatial reasoning benchmarks.
SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
The paper introduces SpatialASM, a framework that trains large vision-language models to actively plan task-relevant viewpoints and mentally imagine missing spatial evidence via novel-view synthesis scaffolding, improving spatial reasoning across four backbones and seven benchmarks without extra views at inference.
SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning
The paper proposes SVoT, a reinforcement learning framework that generates interleaved, verifiable intermediate states and visualizations for multi-hop spatial reasoning in MLLMs, achieving significant accuracy gains on new benchmarks involving multi-object interactions and numerical reasoning.
Retrieve, Integrate, and Synthesize: Spatial-Semantic Grounded Latent Visual Reasoning
This paper introduces RIS, a framework for spatial-semantic grounded latent visual reasoning in Multimodal Large Language Models to overcome information bottlenecks. It proposes anchoring latent tokens to spatial and semantic evidence, showing improvements on benchmarks like V* and HRBench.