Mitigating the Safety Alignment Tax with Null-Space Constrained Policy Optimization

Large Language Models (LLMs) have demonstrated significant capabilities in natural language processing but are often criticized for generating factually incorrect content, known as hallucinations. A recent study investigates the effects of three hallucination-reduction techniques—Chain of Verification, Decoding by Contrasting Layers, and Retrieval-Augmented Generation—on the creativity of LLMs across various models and scales, revealing that these methods can have opposing effects on divergent creativity.

KBQA-R1 has been introduced as a new framework aimed at improving Knowledge Base Question Answering (KBQA) by utilizing Reinforcement Learning to optimize interactions with knowledge bases, addressing limitations of current Large Language Models (LLMs) that often generate inaccurate queries or rely on rigid templates.

Large Language Models (LLMs) have demonstrated significant capabilities in natural language processing; however, they often exhibit overconfidence, leading to discrepancies between predicted confidence and actual correctness. A recent study analyzed nine LLMs across three factual Question-Answering datasets, revealing that the integration of distractor prompts can enhance calibration, resulting in accuracy improvements of up to 460% and reductions in expected calibration error by up to 90%.

The Textual Self-Attention Network (TSAN) has been introduced as a novel approach for optimizing Large Language Models (LLMs) during test-time, allowing for the analysis and synthesis of multiple candidate responses without requiring parameter updates. This method addresses the limitations of previous techniques that focused on revising single responses, thereby enhancing the potential for improved output quality.

Recent research introduces grammar-aligned decoding (GAD), a new approach that aims to improve the output quality of large language models (LLMs) by aligning their sampling with grammar constraints. This method addresses the limitations of grammar-constrained decoding (GCD), which can distort the LLM's output distribution, resulting in grammatical but low-quality outputs.

The introduction of KeyframeFace marks a significant advancement in generating dynamic 3D facial animations from natural language, addressing the limitations of existing datasets that primarily focus on speech-driven animations or unstructured expression sequences. This large-scale multimodal dataset includes 2,100 expressive scripts, monocular videos, and detailed annotations, enabling more nuanced and contextually rich animations.

Test-time scaling (TTS) has been identified as a significant method for enhancing the reasoning capabilities of Large Language Models (LLMs) by allowing for additional computational resources during inference. This study systematically investigates TTS applications in both open-source and closed-source Vision-Language Models (VLMs), revealing varied performance outcomes across different benchmarks.

A recent study has proposed the Plasticity-Ceiling Framework to enhance the utilization of expert trajectories in post-training for large language models (LLMs). This framework aims to optimize the integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), establishing a Sequential SFT-then-RL pipeline as the most effective approach, while providing scaling guidelines for transitioning to RL at specific SFT phases.