Vision-Language Models (VLMs) remain limited in high-precision numerical prediction for 3D scene understanding. In continuous numerical tasks, existing reinforcement learning methods such as GRPO are sample-inefficient: binary rewards fail to distinguish near-miss predictions from incorrect ones, while group-relative normalization can further weaken informative reward differences when sample groups are uniformly poor. Consequently, near-miss samples contribute little to policy updates despite carrying valuable learning signal. We propose Smooth Numerical Reward Activation (SNRA) and Absolute-Preserving GRPO (AP-GRPO) to address this issue. SNRA converts raw numerical errors into dense rewards via a dynamically parameterized sigmoid, while AP-GRPO modulates group-relative advantages with an absolute reward factor to emphasize informative near-miss samples and suppress uniformly uninformative groups. This improves the effective training contribution of near-miss predictions and substantially enhances data efficiency. On our Numerical3D-50k dataset, AP-GRPO achieves performance comparable to large-scale supervised learning across five benchmarks and four VLM architectures, without architectural modifications.

Vision-Language Models (VLMs) show promise for autonomous driving, yet their struggle with hallucinations, inefficient reasoning, and limited real-world validation hinders accurate perception and robust step-by-step reasoning. To overcome this, we introduce AgentThink, a pioneering unified framework that, for the first time, integrates Chain-of-Thought (CoT) reasoning with dynamic, agent-style tool invocation for autonomous driving tasks. AgentThink’s core innovations include: (i) Structured Data Generation, by establishing an autonomous driving tool library to automatically construct structured, self-verified reasoning data explicitly incorporating tool usage for diverse driving scenarios; (ii) A Two-stage Training Pipeline, employing Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO) to equip VLMs with the capability for autonomous tool invocation; and (iii) Agent-style Tool-Usage Evaluation, introducing a novel multi-tool assessment protocol to rigorously evaluate the model’s tool invocation and utilization. Experiments on the DriveLMM-o1 benchmark demonstrate AgentThink significantly boosts overall reasoning scores by 53.91% and enhances answer accuracy by 33.54%, while markedly improving reasoning quality and consistency. Furthermore, ablation studies and robust zero-shot/few-shot generalization experiments across various benchmarks underscore its powerful capabilities. These findings highlight a promising trajectory for developing trustworthy and tool-aware autonomous driving models.