GTPO: Stabilizing Group Relative Policy Optimization via Gradient and Entropy Control

A novel approach called TRIM-KV has been introduced to enhance token retention in memory-bounded key-value caches for large language models (LLMs). This method utilizes a lightweight retention gate to predict the importance of tokens at creation time, allowing for the eviction of less critical tokens when memory limits are reached. The approach aims to address the challenges posed by the quadratic cost of self-attention and the growing size of KV caches during long-horizon inference.

A recent study investigates the impact of explicit world-modeling objectives on the internal representations and performance of Transformers, particularly in the context of a controlled Rubik's Cube task. The research compares standard next-token prediction with two world-modeling strategies, revealing that explicit modeling enhances representation quality and downstream performance after reinforcement learning post-training.

DVPO, or Distributional Value Modeling-based Policy Optimization, has been introduced as a new reinforcement learning framework aimed at enhancing the post-training phase of large language models (LLMs). This framework addresses the challenges posed by noisy supervision and aims to improve both robustness and generalization by utilizing conditional risk theory and token-level value distributions.

AdaptVision has been introduced as a new paradigm in Vision-Language Models (VLMs), focusing on adaptive visual token acquisition to enhance efficiency in visual question answering tasks. By employing a coarse-to-fine approach, the model selectively acquires visual information as needed, addressing the computational overhead associated with traditional methods that rely on fixed-ratio compression.

Kardia-R1 has introduced KardiaBench, a benchmark designed to enhance emotional reasoning in conversational agents by utilizing a dataset of 178,080 QA pairs from 671 real-world profiles, addressing the limitations of existing systems that lack personalized emotional understanding.

Recent research highlights the challenges of pruning reasoning language models (RLMs) like OpenAI's o1 and DeepSeek-R1, which are crucial for multi-step reasoning tasks. The study reveals that traditional pruning methods can severely impair the accuracy and coherence of these models, even at moderate levels of sparsity.

A new adaptive curriculum mechanism called CAPO (Curriculum Advantage Policy Optimization) has been proposed to enhance cross-domain reasoning tasks in reinforcement learning. This mechanism aims to improve reasoning capabilities by utilizing advantage signals, initially focusing on positive samples to establish a solid foundation before incorporating negative signals for better discrimination.

The introduction of SeeNav-Agent marks a significant advancement in Vision-Language Navigation (VLN) by addressing common errors in perception, reasoning, and planning that hinder navigation performance. This framework incorporates a dual-view Visual Prompt technique to enhance spatial understanding and a novel step-level Reinforcement Fine-Tuning method, Step Reward Group Policy Optimization (SRGPO), to improve navigation task rewards.