Multimodal large language models increasingly use visual chain-of-thought to reason about specific image regions, but existing methods have trade-offs. Textified coordinates create a modality mismatch and fragment meaning, while fixed-granularity patches limit precise region selection and often require major architectural changes. The new NV-CoT framework avoids both problems by letting models generate bounding-box coordinates directly as continuous actions, needing only minimal modification to the model architecture.
NV-CoT supports both supervised fine-tuning and reinforcement learning. For the latter, the authors replace categorical token policies with Gaussian or Laplace policies over coordinates and introduce stochasticity through reparameterized sampling, which makes the framework fully compatible with GRPO-style policy optimization. This design allows the model to treat visual grounding as a continuous control problem rather than a discrete language task.
In evaluations across three benchmarks against 11 representative visual reasoning baselines, NV-CoT reportedly improves localization precision and final answer accuracy while also accelerating training convergence. The paper, accepted at NeurIPS 2026 as an oral presentation (top 0.4%), argues that continuous-action visual reasoning is a more effective paradigm for region-grounded reasoning in MLLMs.