Enhancing large language model navigation via two-stage fine-tuning and sparse-reward reinforcement learning

Large language models lack the spatial intelligence needed to interpret highly abbreviated and semantically implicit inland waterway navigation instructions, causing persistent hallucinations. To address this issue, Navigation-GPT-SI, a two-stage spatial intelligence framework built on a 14-billion-parameter (14B) Qwen2.5 model (Qwen2.5-14B), is proposed. Stage one employs supervised fine-tuning on spatial reasoning data to elicit spatial chain-of-thought reasoning and provide a stable policy initialization for reinforcement learning. Stage two introduces Group Discriminative Policy Optimization (GDPO), which applies key-token weighting to focus optimization on semantically decisive spatial units, thereby mitigating token contribution imbalance. At inference, a lightweight neuro-symbolic safeguard is activated under low self-consistency, leveraging the model’s agent capability to invoke external tools and perform deterministic step-by-step reasoning of encounter geometry and navigation light semantics. Navigation-GPT-SI is evaluated against 16 mainstream LLMs, including ChatGPT, Claude, and DeepSeek, on typical inland waterway encounter scenarios. Experimental results show that Navigation-GPT-SI achieves accuracies of 95.9%, 89.0%, and 99.5% in physical space recognition, spatial logical reasoning, and spatial temporal reasoning, respectively, representing improvements of 48.9, 47.1, and 36.8 percentage points, respectively, over the baseline model, and outperforming all evaluated general-purpose language models in these three dimensions. Downstream transfer to 55 real-world collision avoidance scenarios from the DNV (Det Norske Veritas) schema further shows that the 14B-parameter Navigation-GPT-SI achieves a collision avoidance success rate of 90.9% and an average trajectory length of 5.897 NM with a single inference time of 5.4 s. These results highlight the value of domain-specific adaptation: with only 14B parameters, Navigation-GPT-SI matches or exceeds 72B-parameter general-purpose models and even substantially larger ones in success rate, average trajectory length, and average number of turns, indicating that the learned spatial intelligence can effectively transfer to downstream navigation decision-making.