2025-02-18s

video-SALMONN-01: An open-source audio-visual LLM enhances video understanding with a novel reasoning-intensive dataset and the pDPO method, achieving significant accuracy gains.

SYSGEN: A novel pipeline generates effective system messages for LLMs using open-source models, improving model responses and addressing data scarcity in supervised fine-tuning.

SAFE-SQL boosts Text-to-SQL accuracy by intelligently generating and filtering self-augmented examples for in-context learning, surpassing existing methods in challenging scenarios.

PhysReason benchmark evaluates physics-based reasoning in LLMs, revealing critical limitations and guiding future improvements.

MagicArticulate automates 3D model animation preparation by generating skeletons and skinning weights, overcoming prior manual methods’ limitations, and introducing Articulation-XL, a large-scale benc…

HUMANUP: A novel two-stage reinforcement learning framework enables real-world humanoid robots to autonomously recover from falls on various terrains.

Large language models struggle with mathematical word problems, demonstrating flaws in reasoning despite achieving high accuracy; a new study highlights these persistent gaps in generalization abiliti…

LLMs’ performance on language complexity tasks (LIX & ADD) reveals a strong correlation with general capabilities, suggesting complexity metrics as noisy zero-shot proxies for model evaluation.

AI models learn intuitive physics from self-supervised video pretraining.

HermesFlow seamlessly bridges the understanding-generation gap in MLLMs using a novel Pair-DPO framework and self-play optimization on homologous data, achieving significant performance improvements.

Diffusion-Sharpening enhances diffusion model fine-tuning by optimizing sampling trajectories, achieving faster convergence and high inference efficiency without extra NFEs, leading to improved alignm…

PoPilot, a novel proof-oriented programming LLM, outperforms GPT-40 by 64% under data scarcity by using synthetic data augmentation.

High-quality, task-relevant pretraining data surpasses large-scale pretraining in atomic property prediction, achieving comparable performance at 1/24th the computational cost.

TalkHier, a novel framework for LLM multi-agent systems, uses structured communication and hierarchical refinement to achieve state-of-the-art performance on various tasks, improving collaboration and…

NSA: a novel sparse attention mechanism achieves efficient long-context modeling by combining algorithmic innovations with hardware-aligned optimizations, surpassing full attention models across vario…

LLMs’ knowledge acquisition is unveiled through the lens of evolving knowledge circuits, revealing how new knowledge integration depends on relevance to existing knowledge, exhibiting distinct phases …

Dyve: A novel dynamic process verifier boosts LLM reasoning accuracy by cleverly combining fast, immediate checks with deeper, slower analyses for complex steps, achieving significant performance gain…

Cuckoo: a novel information extraction (IE) model leverages LLM pre-training data, achieving superior performance in few-shot settings by reframing next-token prediction as token extraction.

MIKASA, a new benchmark for memory-intensive reinforcement learning, provides a unified framework for evaluating memory capabilities in diverse scenarios, including complex robotic manipulation tasks.

Fact-checkers need explainable AI: This study reveals how AI tools can better support fact-checkers by providing explanations tailored to their workflows, addressing unmet needs, and improving the eff…