Reinforcing Dual-Path Reasoning in Spatial Vision Language Models

2026-06-16

AI Summary

Existing spatial vision-language models exhibit limited performance on tasks requiring multi-step reasoning about depth, distance, and scene relationships, and lack the ability to adaptively select reasoning strategies based on query type. This work proposes SR-REAL, a novel framework that introduces, for the first time, a dual-path reasoning mechanism into spatial VLMs: Language-Only Reasoning (LOR) performs step-by-step linguistic deduction, while Detection-Then-Reasoning (DTR) first extracts 3D geometric cues before conducting quantitative inference. The two paths are jointly trained and optimized via reinforcement learning for adaptive path selection and geometric alignment, leveraging supervised fine-tuning for cold-start initialization, chain-of-thought annotations, and a multi-objective reward combining accuracy, output format correctness, and center detection fidelity. SR-REAL significantly outperforms prior models across multiple spatial reasoning benchmarks, supports both reasoning modalities within a single model, and demonstrates strong cross-dataset generalization without task-specific fine-tuning.

Abstract

Spatial VLMs have made substantial progress in geometric perception, yet complex spatial reasoning requiring multi-step inference over depth, distance, and scene relations remains challenging. Moreover, different spatial queries call for fundamentally different strategies: some are best addressed through purely linguistic, step-by-step deduction, while others require explicit 3D grounding before quantitative inference. We present Dual-Path Spatial Reasoning via Reinforcement Learning for Spatial VLMs (SR-REAL), a unified framework that equips a spatial VLM with two complementary reasoning paths: Language-Only Reasoning (LOR), which performs step-by-step linguistic deduction, and Detect-Then-Reason (DTR), which detects 3D geometric cues (e.g., centers or bounding boxes) via region tokens before explicit geometric inference. SR-REAL begins with a cold-start supervised fine-tuning stage that constructs LOR and DTR chain-of-thought supervision and exposes a region-to-3D interface, followed by RL that optimizes the policy model with accuracy and format rewards; for DTR, a discrete center-based detection reward further refines geometric alignment. Across diverse spatial benchmarks, SR-REAL significantly outperforms spatial VLM baselines: (i) a single RL-trained model supports both reasoning paths, with DTR excelling in region-aware tasks through precise 3D localization and LOR enhancing general spatial reasoning; (ii) jointly training both paths fosters mutual reinforcement; (iii) high-quality, blended cold-start data is crucial for stable RL optimization; and (iv) the model generalizes across datasets and domains without per-task tuning, demonstrating positive transfer between LOR and DTR.

Problem

• spatial reasoning
• visual language models
• 3D grounding
• multi-step inference
• geometric perception

Innovation

• Dual-Path Reasoning
• Spatial Vision Language Models
• Reinforcement Learning
• 3D Grounding
• Chain-of-Thought