On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs

Reinforcement learning (RL) fine-tuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations--misleading captions or incorrect chain-of-thought (CoT) traces--cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consistency is taken into account across open-source multimodal reasoning models. Entropy-based metrics further show that these perturbations reshape model uncertainty and probability mass on the correct option, exposing model-specific trends in miscalibration. To better understand these vulnerabilities, we further analyze RL fine-tuning dynamics and uncover an accuracy-faithfulness trade-off: fine-tuning raises benchmark accuracy, but can simultaneously erode the reliability of the accompanying CoT and its robustness to contextual shifts. Although adversarial augmentation improves robustness, it does not by itself prevent faithfulness drift. Incorporating a faithfulness-aware reward can restore alignment between answers and reasoning, but when paired with augmentation, training risks collapsing onto shortcut strategies and robustness remains elusive. Together, these findings highlight the limitations of accuracy-only evaluations and motivate training and assessment protocols that jointly emphasize correctness, robustness, and the faithfulness of visually grounded reasoning.

GradientSurf: Gradient-Domain Neural Surface Reconstruction from RGB Video

This paper proposes GradientSurf, a novel algorithm for real time surface reconstruction from monocular RGB video. Inspired by Poisson Surface Reconstruction, the proposed method builds on the tight coupling between surface, volume, and oriented point cloud and solves the reconstruction problem in gradient-domain. Unlike Poisson Surface Reconstruction which finds an offline solution to the Poisson equation by solving a linear system after the scanning process is finished, our method finds online solutions from partial scans with a neural network incrementally where the Poisson layer is designed to supervise both local and global reconstruction. The main challenge that existing methods suffer from when reconstructing from RGB signal is a lack of details in the reconstructed surface. We hypothesize this is due to the spectral bias of neural networks towards learning low frequency geometric features. To address this issue, the reconstruction problem is cast onto gradient domain, where zeroth-order and first-order energies are minimized. The zeroth-order term penalizes location of the surface. The first-order term penalizes the difference between the gradient of reconstructed implicit function and the vector field formulated from oriented point clouds sampled at adaptive local densities. For the task of indoor scene reconstruction, visual and quantitative experimental results show that the proposed method reconstructs surfaces with more details in curved regions and higher fidelity for small objects than previous methods.