cs.CVApr 1, 2026
SaveSTAR: Mitigating Cascading Errors in Spatial Reasoning via Turn-point Alignment and Segment-level DPO
Organizations: Guangdong University of Finance and Economics · Chongqing University · Westlake University · The University of Hong Kong
Abstract
Structured spatial navigation is a core benchmark for Large Language Models (LLMs) spatial reasoning. Existing paradigms like Visualization-of-Thought (VoT) are prone to cascading errors in complex topologies. To solve this, we propose STAR, a two-stage framework grounded on topological anchors, and introduce the RedMaze-23K dataset with human-inspired turnpoint annotations. The first stage uses supervised fine-tuning to help models internalize spatial semantics and prune redundant paths. The second adopts Spatial-aware Segment-level Direct Preference Optimization (SDPO) to refine self-correction in long-horizon navigation. Experiments show STAR achieves state-of-the-art performance among open-source models: its 32B variant outperforms DeepSeek-V3 (29.27% vs. 25.00%) and reaches 82.4% of GPT-4's performance.
Figures & tables
Fig. 1 : Strategic Navigation Heuristics: Traditional Sequential Exploration vs. Human-inspired Turn-point Anchoring
Fig. 2 : Framework of STAR. STAR adopts a two-stage paradigm: Stage 1 utilizes Supervised Fine-Tuning (SFT) on RedMaze-23K to ground static spatial understanding, while Stage 2 employs Spatial-aware Direct Preference Optimization (SDPO) to refine dynamic decision-making via segment-level alignment. This progression systematically enhances accuracy and robustness by mitigating reasoning drift at critical topological anchors.
| Model | Variant | RP.CR ↑ | RP.SR ↑ | NS.Acc ↑ | TC.Acc ↑ | RU.Acc ↑ |
|---|---|---|---|---|---|---|
| API Models | ||||||
| GPT-4 | – | 58.72 | 41.37 | 57.10 | 29.29 | 44.27 |
| DeepSeek-V3 | – | 51.38 | 25.00 | 49.96 | 38.35 | 21.75 |
| Open-Source Baselines | ||||||
| Llama-3.1-8B | – | 28.43 | 6.92 | 17.75 | 17.63 | 16.14 |
| DeepSeek-R1-14B | – | 24.88 | 4.15 | 15.20 | 13.42 | 15.58 |
TABLE I : Performance Comparison on RedMaze-23K Benchmark Across Model Scales and Training Stages. Metric abbreviations: RP.CR (Route Planning Completion Rate), RP.SR (Route Planning Success Rate), NS.Acc (Next-Step Accuracy), TC.Acc (Turnpoint Comprehension Accuracy), RU.Acc (Rule Understanding Accuracy). The best open-source results are underlined .
| Setting | TP Info | Hints | RP.CR ↑ | RP.SR ↑ | NS.Acc ↑ |
|---|---|---|---|---|---|
| Baseline | – | – | 41.35 | 25.21 | 30.90 |
| Map-Enhanced | ✓ | – | 44.89 | 28.21 | 31.43 |
| Hint-Enhanced | – | ✓ | 43.28 | 29.13 | 33.54 |
| STAR (Ours) | ✓ | ✓ | 46.42 | 29.27 | 35.00 |
TABLE II: Ablation Study on RedMaze-23K. We investigate the impact of TurnPoint (TP) guidance and reasoning hints. Metric abbreviations: RP.CR (Route Planning Completion Rate), RP.SR (Route Planning Success Rate), and NS.Acc (Next-Step Accuracy). Ours represents the final configuration of the STAR framework.
Fig. 3 : Attention Score Comparison (Base VS Ours). Visualizing the attention scores shows that our method focuses more accurately on critical turn-points compared to the baseline.
Fig. 4 : Confidence Score Comparison (Base VS Ours). Comparison of base model and STAR-enhanced model on our dataset, with their confidence scores in selecting the positive sample(A) over various negative samples (B, C, D, E). STAR-enhanced model consistently achieved higher scores across four pairs of positive and negative samples.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Subgroup ( ) | Base ( ) | Turn Range | Complexity |
|---|---|---|---|
| Subgroup 1 | 1 | Low | |
| Subgroup 2 | 3 | Moderate | |
| Subgroup 3 | 5 | High | |
| Subgroup | Customizable |
TABLE III : Difficulty Calibration Parameters. The baseline parameter ensures structural diversity across subgroups.
Fig. 5 : Qualitative Comparison of Reasoning Paradigms in Structured Spatial Navigation. We illustrate the divergence in decision-making logic between three paradigms: (1) CoT (left) relies on purely textual decomposition but frequently fails to ground spatial relationships in complex junctions, leading to invalid moves (e.g., colliding with obstacles). (2) VoT (middle) introduces step-wise textual analysis to find the shortest path; however, it lacks explicit visual anchoring, making it vulnerable to cascading errors during long-horizon reasoning. (3) Ours (STAR) (right) implements Topological Anchoring , where the model is required to confirm its current position by re-visualizing the traversed path and marking its status with a visual icon at each critical turn-point. This structural feedback loop ensures that the model internalizes the maze’s topological constraints, effectively mitigating reasoning drift and ensuring policy consistency from origin to destination.
| Model | GPT-4 | DeepSeek-V3 | ||||
|---|---|---|---|---|---|---|
| Metric | +CoT | +VoT | +Ours | +CoT | +VoT | +Ours |
| CR | 48.55 | 45.66 | 58.72 | 57.51 | 55.16 | 51.41 |
| SR | 22.43 | 26.82 | 41.38 | 40.88 | 36.42 | 25.04 |
| Acc | 48.43 | 37.16 | 57.10 | 53.51 | 40.43 | 50.02 |
| Model | Qwen-7B-Ours | Qwen-32B-STAR | ||||
| Metric | +CoT | +VoT | +Ours | +CoT | +VoT | +Ours |
TABLE IV : Cross-Prompt Performance on RedMaze-23K (All values in %).
Fig. 6 : Mechanics of Spatial-aware Segment-level DPO (SDPO). This figure illustrates the construction of preference pairs by isolating reasoning divergence. (Left) Positive Session (Chosen): The model maintains strict consistency between its textual instructions and visual state updates, correctly navigating through topological anchors to reach the destination. (Right) Negative Session (Reject): While the initial step is correct (prefix-correct), the reasoning drift occurs at Step 2. The failure is characterized by three critical dimensions: (i) Logical Inconsistency: The instruction ”move left” contradicts the generated visual map; (ii) Constraint Violation: The model attempts to ”step over obstacles,” violating the maze’s physical boundaries; (iii) Structural Corruption: The visualization logic collapses, leading to an erroneous map representation. SDPO targets this specific segment for penalty to prevent error propagation.