STAR: Mitigating Cascading Errors in Spatial Reasoning via Turn-point Alignment and Segment-level DPO

TL;DR

STAR is a two-stage framework that improves spatial reasoning in LLMs by reducing cascading errors through turn-point alignment and segment-level optimization, demonstrating state-of-the-art results.

Contribution

It introduces a novel two-stage approach with topological anchors and a new dataset, RedMaze-23K, to enhance spatial reasoning and self-correction in LLMs.

Findings

STAR's 32B model outperforms DeepSeek-V3 with 29.27% accuracy.

STAR reaches 82.4% of GPT-4's performance on spatial navigation tasks.

The RedMaze-23K dataset provides valuable turnpoint annotations for training.

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.