Multi-Objective Reward and Preference Optimization: Theory and Algorithms

A new study introduces a data-efficient fine-tuning strategy for large-scale text-to-video diffusion models, enabling the addition of generative controls over physical camera parameters using sparse, low-quality synthetic data. This approach demonstrates that models fine-tuned on simpler data can outperform those trained on high-fidelity datasets.

A recent study has introduced differential smoothing as a method to mitigate the diversity collapse often observed in large language models (LLMs) during reinforcement learning fine-tuning. This method aims to enhance both the correctness and diversity of model outputs, addressing a critical issue where outputs lack variety and can lead to diminished performance across tasks.

SplatCo has been introduced as a novel structure-view collaborative Gaussian splatting framework designed for high-fidelity rendering of complex outdoor scenes. This framework integrates a cross-structure collaboration module, a cross-view pruning mechanism, and a structure view co-learning module to enhance detail preservation and rendering efficiency in large-scale unbounded scenes.

A recent study explores the automated recognition of instructional activities and discourse from multimodal classroom data, utilizing AI-driven analysis of 164 hours of video and 68 lesson transcripts. This research aims to replace manual annotation methods, which are resource-intensive and difficult to scale, with more efficient AI techniques for actionable feedback to educators.

The introduction of the D³-Predictor presents a significant advancement in dense prediction by addressing the limitations of existing diffusion models, which are hindered by stochastic noise that disrupts fine-grained spatial cues and geometric structure mappings. This new framework reformulates a pretrained diffusion model to eliminate stochasticity, allowing for a more deterministic mapping from images to geometry.

A novel framework for white balance (WB) correction has been proposed, leveraging a perception-inspired Learnable HSI (LHSI) color space. This approach aims to address the limitations of traditional sRGB-based WB editing, which struggles with color constancy in complex lighting conditions due to fixed nonlinear transformations and entangled color channels.

A new study introduces latent-action world models that learn from both action-conditioned and action-free data, addressing the limitations of traditional models that rely heavily on labeled action trajectories. This approach allows for training on large-scale unlabeled trajectories while requiring only a small set of labeled actions.

A new study presents an efficient probabilistic hardware architecture designed for diffusion-like models, addressing the limitations of previous proposals that relied on unscalable hardware and limited modeling techniques. This architecture, based on an all-transistor probabilistic computer, is capable of implementing advanced denoising models at the hardware level, potentially achieving performance parity with GPUs while consuming significantly less energy.