AI Firehose

Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows ArXiv link for Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows

Foley-Flow transforms video-to-audio generation by achieving unmatched semantic and rhythmic coherence through innovative masked modeling and dynamic flows. It outshines existing methods, producing synchronized audio that enhances viewer experience. https://arxiv.org/abs/2603.08126

High-Fidelity Pruning for Large Language Models ArXiv link for High-Fidelity Pruning for Large Language Models

HFPrune is a novel pruning method for large language models that utilizes information entropy to assess neuron importance, enhancing predictive accuracy and reducing computational costs, outperforming other techniques for efficient AI in resource-limited settings. https://arxiv.org/abs/2603.08083

More Bang for the Buck: Process Reward Modeling with Entropy-Driven Uncertainty ArXiv link for More Bang for the Buck: Process Reward Modeling with Entropy-Driven Uncertainty

EDU-PRM automates effective reasoning paths, needing only 1.5% of traditional training data while achieving state-of-the-art performance in math reasoning. By using entropy-driven sampling, EDU-PRM enhances accuracy and efficiency, improving complex problem-solving. https://arxiv.org/abs/2503.22233

TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward ArXiv link for TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward

TDM-R1 revolutionizes few-step diffusion models by integrating non-differentiable rewards, improving text-to-image generation. This approach enables faster, efficient high-quality image creation with fewer generating steps. https://arxiv.org/abs/2603.07700

Scale Dependent Data Duplication ArXiv link for Scale Dependent Data Duplication

This study reveals that as language models grow, they struggle with semantic duplicates, treating them like exact copies. This effect could undermine training efficiency and model quality, urging a rethink of data diversity in AI development. https://arxiv.org/abs/2603.06603

Test-Time Meta-Adaptation with Self-Synthesis ArXiv link for Test-Time Meta-Adaptation with Self-Synthesis

This study presents MASS, a meta-learning approach that enables large language models to adapt in real-time by generating synthetic data tailored to problems, boosting performance in math reasoning tasks, allowing AI to enhance its capabilities dynamically. https://arxiv.org/abs/2603.03524

SecAgent: Efficient Mobile GUI Agent with Semantic Context ArXiv link for SecAgent: Efficient Mobile GUI Agent with Semantic Context

SecAgent, an innovative mobile GUI agent, addresses mul- tilingual data scarcity and history inefficiencies, achieving top performance with 3B parameters and a semantic context approach for smartphone automation. https://arxiv.org/abs/2603.08533

A Simple "Motivation" Can Enhance Reinforcement Finetuning of Large Reasoning Models ArXiv link for A Simple "Motivation" Can Enhance Reinforcement Finetuning of Large Reasoning Models

A paper presents MeRF (Motivation-enhanced Reinforcement Finetuning), improving large reasoning models by integrating task motivations into training. This yields noteworthy performance gains in reasoning tasks, highlighting the power of in-context learning in AI. https://arxiv.org/abs/2506.18485

SynPlanResearch-R1: Encouraging Tool Exploration for Deep Research with Synthetic Plans ArXiv link for SynPlanResearch-R1: Encouraging Tool Exploration for Deep Research with Synthetic Plans

SynPlanResearch-R1 enhances research agents by encouraging deeper exploration during tool use, significantly boosting performance in web QA tasks. By synthesizing effective tool-use trajectories, this innovative approach outperforms state-of-the-art methods. https://arxiv.org/abs/2603.07853

Estimating Treatment Effects under Algorithmic Interference: A Structured Neural Networks Approach ArXiv link for Estimating Treatment Effects under Algorithmic Interference: A Structured Neural Networks Approach

Researchers propose an estimator that reduces algorithmic interference in creator-side experiments, showing how standard methods can mislead decisions. This approach identified a promo algorithm's negative impact, while estimators suggested benefits. https://arxiv.org/abs/2406.14380

LD-RPS: Zero-Shot Unified Image Restoration via Latent Diffusion Recurrent Posterior Sampling ArXiv link for LD-RPS: Zero-Shot Unified Image Restoration via Latent Diffusion Recurrent Posterior Sampling

LD-RPS presents a technique for image restoration using latent diffusion and recurrent sampling to recover from diverse degradations without needing paired datasets. This unified method, augmented by text guidance, surpasses traditional models in zero-shot settings. https://arxiv.org/abs/2507.00790

From Static Inference to Dynamic Interaction: A Survey of Streaming Large Language Models ArXiv link for From Static Inference to Dynamic Interaction: A Survey of Streaming Large Language Models

This survey explores the streaming paradigm for Large Language Models, shifting static interactions to dynamic exchanges. It categorizes methodologies, paving the way for real-time applications like assistants and robots, aiming for streaming intelligence. https://arxiv.org/abs/2603.04592

LMOD+: A Comprehensive Multimodal Dataset and Benchmark for Developing and Evaluating Multimodal Large Language Models in Ophthalmology ArXiv link for LMOD+: A Comprehensive Multimodal Dataset and Benchmark for Developing and Evaluating Multimodal Large Language Models in Ophthalmology

LMOD+ is a multimodal dataset and benchmark for large language models in ophthalmology, boosting dataset size by 50% for key disease diagnosis and demographic prediction. Results show promise, but performance gaps in MLLMs indicate a need for advancements in AI. https://arxiv.org/abs/2509.25620

Agora: Teaching the Skill of Consensus-Finding with AI Personas Grounded in Human Voice ArXiv link for Agora: Teaching the Skill of Consensus-Finding with AI Personas Grounded in Human Voice

MIT's Agora project uses AI for civic education, enabling users to enhance consensus skills through varied policy perspectives. Initial findings show students reported improved problem-solving skills and created better consensus statements than traditional methods. https://arxiv.org/abs/2603.07339

ImageEdit-R1: Boosting Multi-Agent Image Editing via Reinforcement Learning ArXiv link for ImageEdit-R1: Boosting Multi-Agent Image Editing via Reinforcement Learning

ImageEdit-R1 is a new multi-agent framework utilizing reinforcement learning for advanced image editing. The approach consistently outperforms existing models through collaboration among specialized agents, establishing a new standard for context-aware edits. https://arxiv.org/abs/2603.08059

FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference ArXiv link for FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference

FreeKV enhances key-value retrieval in large language models, achieving 13× speedup while preserving accuracy. By using speculative retrieval and hybrid memory layouts, FreeKV addresses long context processing challenges in LLMs for advanced AI applications. https://arxiv.org/abs/2505.13109

Diffusion-SAFE: Diffusion-Native Human-to-Robot Driving Handover for Shared Autonomy ArXiv link for Diffusion-SAFE: Diffusion-Native Human-to-Robot Driving Handover for Shared Autonomy

Diffusion-SAFE offers a new shared autonomy approach with a 93% handover success rate in driving. It improves safety and control transfer, balancing human intent and automation, revolutionizing interactions with autonomous vehicles. https://arxiv.org/abs/2505.09889

Exploring the Reasoning Depth of Small Language Models in Software Architecture: A Multidimensional Evaluation Framework Towards Software Engineering 2.0 ArXiv link for Exploring the Reasoning Depth of Small Language Models in Software Architecture: A Multidimensional Evaluation Framework Towards Software Engineering 2.0

This study benchmarks SLMs for generating architectural decision records, revealing a reasoning threshold at 3B parameters. Larger models excel in compliance, while mid-sized SLMs benefit from few-shot prompting, providing sustainable AI support for architecture. https://arxiv.org/abs/2603.07091

Can LLM-Simulated Practice and Feedback Upskill Human Counselors? A Randomized Study with 90+ Novice Counselors ArXiv link for Can LLM-Simulated Practice and Feedback Upskill Human Counselors? A Randomized Study with 90+ Novice Counselors

A study shows novice counselors boost their empathetic skills through AI-generated feedback, while practice alone may hinder empathy development, providing a scalable solution to the mental health workforce crisis. https://arxiv.org/abs/2505.02428

Consensus is Not Verification: Why Crowd Wisdom Strategies Fail for LLM Truthfulness ArXiv link for Consensus is Not Verification: Why Crowd Wisdom Strategies Fail for LLM Truthfulness

Research shows aggregating model predictions does not enhance truth in unverifiable domains. Shared errors amplify misconceptions, emphasizing the need for external verification for AI reliability, questioning whether adding compute resolves truth issues. https://arxiv.org/abs/2603.06612

VIVECaption: A Split Approach to Caption Quality Improvement ArXiv link for VIVECaption: A Split Approach to Caption Quality Improvement

Adobe's VIVECaption enhances caption quality with a two-sided strategy, addressing flaws in visual language models. By developing a gold-standard dataset and refining character detection, significant improvements in text-to-image alignment are achieved. https://arxiv.org/abs/2603.07401

Improving reasoning at inference time via uncertainty minimisation ArXiv link for Improving reasoning at inference time via uncertainty minimisation

A novel method from Aarhus University employs uncertainty minimization in large language models, boosting multi-step reasoning without heavy sampling. By emphasizing early reasoning stages, it enhances accuracy and efficiency in AI performance across languages. https://arxiv.org/abs/2603.07159

Improving reasoning at inference time via uncertainty minimisation ArXiv link for Improving reasoning at inference time via uncertainty minimisation

Researchers at Aarhus University introduced a method to enhance reasoning in language models through uncertainty minimization. This approach maximizes self-certainty in reasoning steps, boosting performance while cutting costs and outperforming traditional methods. https://arxiv.org/abs/2603.07159

MoE Lens -- An Expert Is All You Need ArXiv link for MoE Lens -- An Expert Is All You Need

A study on MoE models indicates that a single top-weighted expert can approximate ensemble outputs, suggesting ways to optimize inference and lower costs. This insight allows for expert pruning to boost performance while keeping accuracy intact. https://arxiv.org/abs/2603.05806

Aligning Compound AI Systems via System-level DPO ArXiv link for Aligning Compound AI Systems via System-level DPO

Stanford researchers introduced SysDPO, a new framework for aligning compound AI systems. This method addresses optimization challenges and shows performance gains in AI collaboration, with potential applications in healthcare and education. https://arxiv.org/abs/2502.17721

Learning Next Action Predictors from Human-Computer Interaction ArXiv link for Learning Next Action Predictors from Human-Computer Interaction

Stanford unveil LongNAP, a new AI model that accurately predicts user actions by analyzing their multimodal interaction history. The data collection tool, NAPsack, labels over 360K actions from real user behavior, enhancing proactive AI personalization. https://arxiv.org/abs/2603.05923

TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis ArXiv link for TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis

TumorChain transforms tumor analysis by connecting imaging findings to predictions, enhancing diagnostic accuracy. The dataset, TumorCoT-1.5M, with 1.5 million reasoning instructions and 3D CT scans, sets a new standard for interpretability in oncology decisions. https://arxiv.org/abs/2603.05867

Partial Policy Gradients for RL in LLMs ArXiv link for Partial Policy Gradients for RL in LLMs

A study introduces Partial Policy Gradients for reinforcement learning in large language models, optimizing for subsets of future rewards to enhance persona consistency. This innovative framework outperforms traditional methods and reduces persona drift in dialogue. https://arxiv.org/abs/2603.06138

Match4Annotate: Propagating Sparse Video Annotations via Implicit Neural Feature Matching ArXiv link for Match4Annotate: Propagating Sparse Video Annotations via Implicit Neural Feature Matching

Match4Annotate shifts video annotation using implicit neural representations for efficient label propagation in ultrasound, cutting expert costs. This lightweight framework meets benchmarks for inter-video and intra-video annotation, enabling scalable workflows. https://arxiv.org/abs/2603.06471

Just-In-Time Objectives: A General Approach for Specialized AI Interactions ArXiv link for Just-In-Time Objectives: A General Approach for Specialized AI Interactions

Stanford researchers introduce "just-in-time objectives," inferring user goals in real time to personalize AI interactions. Their tool, Poppins, enhances language models, leading to win rates up to 86% over traditional models, transforming AI collaboration. https://arxiv.org/abs/2510.14591