Reinforcing Dual-Path Reasoning in Spatial Vision Language Models

## The Core Insight

Spatial vision-language models (VLMs) struggle with complex geometric reasoning—tasks that require understanding depth, distance, and 3D relationships in scenes. The authors observe that not all spatial questions benefit from the same strategy: some are solved better by step-by-step linguistic reasoning (talking through the logic), while others need explicit detection of 3D features (like object centers or bounding boxes) before doing math. Rather than forcing a single path, SR-REAL trains a single model to flexibly use both, achieving significantly better performance across spatial benchmarks without task-specific tuning.

## How It Works

The method unfolds in two stages.

**Cold-start supervised fine-tuning** prepares the ground. The authors construct two types of chain-of-thought supervision. Language-Only Reasoning (LOR) teaches the model to reason purely through text—breaking down a spatial question into logical steps without detecting 3D geometry. Detect-Then-Reason (DTR) teaches the opposite: first use "region tokens" (markers tied to image regions) to detect 3D geometric cues like object centers or bounding boxes, then perform explicit quantitative inference on those detected features. This stage also exposes a "region-to-3D interface"—a learned mechanism that translates image regions into 3D geometric information.

**Reinforcement learning (RL)** then optimizes both paths jointly. The reward signal includes accuracy (did the model answer correctly?) and format rewards (did it output in the expected form?). For DTR specifically, a discrete center-based detection reward pushes the model to align predicted 3D locations with ground truth, sharpening geometric precision. The RL training allows the model to learn when to route to LOR versus DTR based on the input question.

The key innovation is that a single trained policy model learns to dispatch queries to the right reasoning path—the model itself decides whether linguistic deduction or 3D detection-first reasoning is more appropriate.

## Why It Matters

Spatial reasoning is notoriously difficult for vision-language models. Real-world questions about scenes—"Is the chair closer to the table or the window?" or "How far apart are the two objects?"—require either sound logical inference or precise geometric grounding, and these are quite different skills. Most prior approaches force a single strategy, leaving performance on the table.

The results show this dual-path design works: a single model trained with SR-REAL outperforms spatial VLM baselines on diverse benchmarks. DTR excels at region-aware tasks through precise 3D localization, while LOR boosts general spatial reasoning. The authors also find that training both paths together creates mutual reinforcement—each helps the other improve. Critically, the trained model generalizes across datasets and domains without per-task fine-tuning, suggesting the learned routing strategy captures something generalizable about when each mode of reasoning is useful.

The practical takeaway: rather than designing a monolithic reasoner or separate specialist models, letting a single model learn to choose its reasoning strategy—and training it with RL to refine that choice—unlocks stronger performance on a range of spatial understanding tasks.