SpatialThinker-7B: Advanced Spatial Reasoning

• SpatialThinker-7B is a multimodal model that integrates high-resolution visual inputs with autoregressive language modeling using reinforcement learning and scene-graph grounding.
• It employs a novel RL framework with lexicographically gated dense spatial rewards to ensure structured observation, scene analysis, and precise answer generation.
• The model demonstrates robust performance on both synthetic spatial VQA benchmarks and real-world tasks, outperforming prior supervised and RL baselines.

SpatialThinker-7B is a multimodal LLM (MLLM) specifically engineered for advanced spatial reasoning and visual question answering (VQA) using dense spatial rewards and structured scene-graph grounding. Built on the Qwen2.5-VL-7B backbone, it unifies high-resolution visual inputs with autoregressive language modeling, reinforcing 3D-aware reasoning through an online reinforcement learning (RL) framework and a novel data synthesis pipeline. SpatialThinker-7B demonstrates enhanced spatial understanding and generalizes robustly to real-world and abstract VQA tasks, outperforming both supervised and prior RL baselines as well as established proprietary models, all with markedly data-efficient training.

1. Model Architecture and Reasoning Paradigm

SpatialThinker-7B employs the Qwen2.5-VL-7B vision-language backbone, wherein a high-resolution vision encoder (processing RGB images from $512\times512$ up to $2048\times2048$ pixels) pairs with a 7-billion parameter autoregressive transformer language decoder. All parameters—including visual and language components—are updated during RL.

The model operationalizes reasoning as a sequential policy $\pi_\theta$ over token sequences $y = (s_1, \ldots, s_T, a)$, given image $X_\mathrm{img}$ and text $X_\mathrm{text}$: $$\pi_\theta(y|X_\mathrm{img}, X_\mathrm{text}) = \prod_{t=1}^T \pi_\theta(s_t|X_\mathrm{img}, X_\mathrm{text}, s_{<t}) \cdot \pi_\theta(a|X_\mathrm{img}, X_\mathrm{text}, s_{\le T})$$

The reasoning process is externally structured into four stages, each marked by explicit tokens:

• <observe>: Natural language scene descriptions localizing regions of interest
• <scene>: JSON-encoded subgraph $G_q$ representing objects (with IDs, labels, bounding boxes) and spatial relations (triplets) relevant to the query
• >: Chain-of-thought step detailing hypothesis and deduction based on $G_q$ and the visual context

  • <answer>: Discrete multiple-choice selection

  Cross-attention layers fuse vision-derived features with text embeddings; all scene-graph-related content is passed as explicit JSON within the text stream, treated as regular tokens by the transformer encoder.

  2. STVQA-7K Data Synthesis and Quality Control

  SpatialThinker-7B’s training leverages STVQA-7K, a curated spatial VQA dataset generated via a semi-automated pipeline:

  • Base graphs and predicates derive from human-annotated Visual Genome (VG150). The set is extended with 34 new spatial predicates (e.g., near/far, taller/shorter, inside/beneath, facing_away).

  • Claude Sonnet 4 is employed to synthesize multiple-choice spatial VQA pairs across nine categories: spatial relations, depth ordering, distance, size, orientation, containment, reach/interactions, region location, counting, and existence.
  • Each sample features four answer choices (A–D) and a uniform label distribution.
  • Every instance is rated by Claude Sonnet; from approximately 56,000 generated pairs, the top 10,000 by rating are retained.
  • Filtering is further strengthened by a GPT-4o pass@2 consistency check: a question is kept if at least one GPT-4o answer matches the synthetic label.
  • The final dataset has 7,587 QA pairs, split into 6,895 for training and 692 for validation.
  • Scene-graph adaptation for each QA example uses lemmatized keyword extraction to form question-focused subgraphs $G_q$, keeping bounding boxes in absolute pixel format to support full scale-aware supervision through CIoU.

  3. Reinforcement Learning with Lexicographically Gated Dense Spatial Rewards

  The SpatialThinker-7B RL protocol is governed by a lexicographically gated, multi-objective reward composition: $$R_{\rm total} = \mathbf{I}[R_{\rm fmt}=1]\;\Bigl(w_{\rm fmt}\,R_{\rm fmt} + w_{\rm cnt}\,R_{\rm cnt} + w_{\rm acc}\,R_{\rm acc} + \mathbf{I}[R_{\rm acc}=1]\;w_{\rm spa}\,R_{\rm spa}\Bigr)$$ with $w_{\rm fmt}=0.1$, $w_{\rm cnt}=0.2$, $w_{\rm acc}=0.5$, $w_{\rm spa}=0.2$. The indicator function $\mathbf{I}$ enforces that spatial rewards are only considered if prior format and accuracy conditions are met.

  Reward Components:

  • Format ($R_{\rm fmt} \in \{0,1\}$): Ensures ordered presence of <observe>, <scene>, <think>, and <answer> tags, and that scene JSON is parseable.
  • Count ($R_{\rm cnt} \in [0,1]$): Penalizes discrepancies in number of predicted/ground-truth objects or relations using weighted normalized error (object: $A_{\rm obj}=0.7$, relation: $A_{\rm rel}=0.3$).
  • Accuracy ($R_{\rm acc} \in \{0,1\}$): Binary, based on exact match with the correct answer.
  • Spatial ($R_{\rm spa} \in [0,1]$): Applied only if $R_{\rm acc}=1$, matching predicted vs. ground-truth objects using the Hungarian algorithm with a composite cost:

  $$C(o_i, o_j) = A_{\rm spa}(1-\mathrm{IoU}(b_i, b_j)) + A_{\rm sem}(1-\mathrm{sim}(\ell_i, \ell_j))$$

  ($A_{\rm spa}=1.0$, $A_{\rm sem}=2.0$), and scored via average Complete IoU, which incorporates area overlap, center distance, and aspect ratio alignment.

  Policy Optimization:

  Group-Relative Policy Optimization (GRPO) is used: for each query, $N=8$ rollouts are sampled and scored, normalized advantages $A^{(i)}$ are computed, and a PPO-style clipped loss (KL penalty coefficient: $1\times10^{-2}$) is applied.

  4. Scene Graph Representation and Encoding Strategy

  Objects $v_i \in V_q$ are represented by label $\ell_i$ and bounding box $b_i=(x_1, y_1, x_2, y_2)$; relations $e_{ij} \in E_q$ are triplets of the form $(v_i, r_{ij}, v_j)$, with $r_{ij}$ drawn from the extended spatial predicate set.

  • Scene graphs are not modeled by a separate GNN; rather, they are serialized as JSON and processed as part of the language input stream.
  • The text encoder integrates object IDs, label tokens, numerical bbox tokens, and relation names via the standard transformer pipeline.
  • Region-of-interest (RoI) filtering constrains graphs and rewards to question-relevant subgraphs, avoiding reward dilution and overfitting to incidental scene elements.

  5. Training Protocols, Baselines, and Evaluation Benchmarks

  Training is conducted from the base model (Qwen2.5-VL-7B) with no prior supervised fine-tuning on STVQA-7K. Reinforcement learning proceeds for approximately 15 hours on 4×NVIDIA H100 GPUs, using AdamW (lr=$1 \times 10^{-6}$), weight decay ($1 \times 10^{-2}$), BF16 precision, and context length of 16,384 tokens.

  Comparative Baselines:

  • Supervised fine-tuning (SFT) with LoRA (lr=$1 \times 10^{-4}$, 3 epochs)
  • Vanilla GRPO using only format/accuracy rewards (both weights 0.5)
  • Proprietary models: GPT-4o, Claude 3.5 Sonnet
  • Open-source generalist MLLMs: Qwen2.5-VL, LLaVA-NeXT, Cambrian-1, VLAA-Thinker
  • Open-source spatial MLLMs: SpaceLLaVA, SpatialRGPT, RoboPoint, SpaceThinker, SpaceOm, SpatialReasoner, SpatialBot, Visionary-R1, SATORI-R1

  Evaluation Benchmarks:

  • Six spatial: CV-Bench (2D/3D), BLINK (spatial, relative depth), 3DSRBench, MMVP, SpatialBench, SpatialReasonerEval
  • Six real-world/general VQA: MM-Star, VStarBench, RealWorldQA, MME-RealWorld-Lite, RoboSpatial-Home (configuration/compatibility), HallusionBench

  6. Empirical Performance and Ablation Analyses

  SpatialThinker-7B demonstrates substantial improvements on both spatial and general VQA metrics:

  Benchmark	SFT	Vanilla GRPO	GPT-4o	SpatialThinker-7B
  CV-Bench (2D/3D)	70.0%	72.7%	79.4%	78.2%
  3DSRBench	–	–	44.3%	56.4%
  BLINK (avg)	–	–	80.4%	79.3%
  Real-World VQA (avg)	65.8%	65.2%	66.2%	69.7%
  12-Benchmark Mean	64.0%	–	67.8%	71.2%

  • On 3DSRBench, SpatialThinker-7B attains 56.4%, exceeding GPT-4o by +12.1 percentage points.
  • Across six spatial benchmarks, consistently outperforms all open-source spatial MLLMs despite using only 7,000 training samples—orders of magnitude less than competitors trained on millions or with explicit RGB-D input.
  • Real-world VQA metrics indicate robustness and grounding; gains are notable on MM-Star, RoboSpatial-Home, and VStarBench.
  • Dense spatial rewards provide a nearly doubled RL improvement (+7.2% versus sparse RL gain of +4.0%).

  Ablation studies:

  • Naïve addition of count+spatial rewards produces reward hacking (23.7%).
  • Lexicographic gating and RoI filtering restore reward utility (76.3%).
  • Final filtering yields an STVQA-7K validation accuracy of 87.9%.
  • KL regularization proves optimal (KL(0.01): 73.7% on CV-Bench, 3B version), outperforming no-KL or $\chi^2$ constraints.

  7. Insights, Contributions, and Limitations

  SpatialThinker-7B’s principal innovation is the integration of explicit scene-graph grounding and lexicographically gated dense rewards, operationalizing a human-like observe $\rightarrow$ localize $\rightarrow$ think $\rightarrow$ answer pipeline. This structured reward approach notably surpasses sparse RL and SFT while maintaining high data efficiency.

  Further, the approach demonstrates robust generalization to both in-domain (synthetic spatial VQA) and out-of-domain (real-world visual and abstract reasoning) benchmarks, with competitive performance relative to leading proprietary and open-source models.

  Identified limitations:

  • Requires explicit scene graph generation and bounding-box labels.
  • Does not yet support implicit spatial reasoning within latent representations or omitting explicit JSON scene-graphs.
  • Extensions to spatiotemporal vision (e.g., video, navigation) and unified multi-objective policies remain open directions.

  References to architecture diagrams and reward dynamics are provided in Figures 1 and 6–7 of the source publication.