PrismAudio: Decomposed Chain-of-Thoughts and Multi-dimensional Rewards for Video-to-Audio Generation

A new algorithm, Deep Gaussian Process Proximal Policy Optimization (GPPO), has been introduced to enhance uncertainty estimation in Reinforcement Learning (RL), particularly in control tasks requiring a balance between safe exploration and efficient learning. GPPO utilizes Deep Gaussian Processes to approximate both policy and value functions, maintaining competitive performance with existing methods while offering calibrated uncertainty estimates.

The introduction of Perceptual-Evidence Anchored Reinforced Learning (PEARL) marks a significant advancement in multimodal reasoning, addressing the limitations of traditional Reinforcement Learning with Verifiable Rewards (RLVR) in Vision-Language Models (VLMs). PEARL enhances reasoning by anchoring it to verified visual evidence, thus mitigating issues like visual hallucinations and reward hacking.

A recent study explores the enhancement of Reinforcement Learning (RL) in 3D visuospatial tasks through a human-informed curriculum design, aiming to improve the technology's effectiveness in complex problem domains. The research highlights the challenges faced by state-of-the-art RL methods, such as PPO and imitation learning, in mastering these tasks.

A new study has introduced a framework for deterministic inference across varying tensor parallel sizes, addressing the issue of training-inference mismatch in large language models (LLMs). This mismatch arises from non-deterministic behaviors in existing LLM serving frameworks, particularly in reinforcement learning settings where different configurations can yield inconsistent outputs.

Recent research has critically evaluated the effectiveness of Reinforcement Learning with Verifiable Rewards (RLVR) in enhancing the reasoning capabilities of large language models (LLMs). The study found that while RLVR-trained models perform better than their base counterparts on certain tasks, they do not exhibit fundamentally new reasoning patterns, particularly at larger evaluation metrics like pass@k.

A new framework called Efficient LLM-Aware (ELLA) has been proposed to enhance heterogeneous graph learning, addressing the challenges posed by complex relation semantics and the limitations of existing models. This framework leverages the reasoning capabilities of Large Language Models (LLMs) to improve the understanding of diverse node and relation types in real-world networks.

Researchers have introduced L2V-CoT, a novel training-free approach that facilitates the transfer of Chain-of-Thought (CoT) reasoning from large language models (LLMs) to Vision-Language Models (VLMs) using Linear Artificial Tomography (LAT). This method addresses the challenges VLMs face in multi-step reasoning tasks due to limited multimodal reasoning data.

A recent study introduces a novel method for eliciting Chain-of-Thought (CoT) reasoning in base large language models (LLMs) through gradient-based representation optimization. This approach addresses the limitations of existing hidden state manipulation techniques, which often lead to degraded text quality and distribution shifts. By reformulating the challenge as an optimization problem, the method aims to guide hidden states towards reasoning-oriented trajectories while preserving linguistic integrity.