Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Reinforced Generation of Combinatorial Structures: Applications to Complexity Theory

The Narcissus Hypothesis: Descending to the Rung of Illusion

Adaptive Kernel Design for Bayesian Optimization Is a Piece of CAKE with LLMs

Joint Memory Frequency and Computing Frequency Scaling for Energy-efficient DNN Inference

Fine-Grained Detection of AI-Generated Text Using Sentence-Level Segmentation

Multimodal Medical Image Classification via Synergistic Learning Pre-training

MVCL-DAF++: Enhancing Multimodal Intent Recognition via Prototype-Aware Contrastive Alignment and Coarse-to-Fine Dynamic Attention Fusion

A Chain-of-thought Reasoning Breast Ultrasound Dataset Covering All Histopathology Categories

PruneCD: Contrasting Pruned Self Model to Improve Decoding Factuality

AHA - Predicting What Matters Next: Online Highlight Detection Without Looking Ahead

Highly Imbalanced Regression with Tabular Data in SEP and Other Applications

Gender and Political Bias in Large Language Models: A Demonstration Platform

RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation

Interpretable Nanoporous Materials Design with Symmetry-Aware Networks

DivLogicEval: A Framework for Benchmarking Logical Reasoning Evaluation in Large Language Models

MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment

LLM Agents for Interactive Workflow Provenance: Reference Architecture and Evaluation Methodology

Justice in Judgment: Unveiling (Hidden) Bias in LLM-assisted Peer Reviews

Single-stream Policy Optimization

FragmentGPT: A Unified GPT Model for Fragment Growing, Linking, and Merging in Molecular Design

ToMA: Token Merge with Attention for Diffusion Models

Clip Your Sequences Fairly: Enforcing Length Fairness for Sequence-Level RL

MEGS$^{2}$: Memory-Efficient Gaussian Splatting via Spherical Gaussians and Unified Pruning

FediLoRA: Heterogeneous LoRA for Federated Multimodal Fine-tuning under Missing Modalities

Long-Range Graph Wavelet Networks

Agentic Software Engineering: Foundational Pillars and a Research Roadmap

TinyDef-DETR: A DETR-based Framework for Defect Detection in Transmission Lines from UAV Imagery

PDTrim: Targeted Pruning for Prefill-Decode Disaggregation in Inference

The Transparent Earth: A Multimodal Foundation Model for the Earth's Subsurface

An Information-Flow Perspective on Explainability Requirements: Specification and Verification

SoK: Large Language Model Copyright Auditing via Fingerprinting

Unlearning as Ablation: Toward a Falsifiable Benchmark for Generative Scientific Discovery

Linguistic Neuron Overlap Patterns to Facilitate Cross-lingual Transfer on Low-resource Languages

Retrieval Enhanced Feedback via In-context Neural Error-book

OpenWHO: A Document-Level Parallel Corpus for Health Translation in Low-Resource Languages

Hierarchical Evaluation Function: A Multi-Metric Approach for Optimizing Demand Forecasting Models

Generative Medical Event Models Improve with Scale

Benchmarking Quantum and Classical Sequential Models for Urban Telecommunication Forecasting

Integrating Belief Domains into Probabilistic Logic Programs

Automating Steering for Safe Multimodal Large Language Models

QSpark: Towards Reliable Qiskit Code Generation

Prompting for Performance: Exploring LLMs for Configuring Software

Intra-DP: A High Performance Collaborative Inference System for Mobile Edge Computing

Exploring Image Generation via Mutually Exclusive Probability Spaces and Local Correlation Hypothesis

Athena: Enhancing Multimodal Reasoning with Data-efficient Process Reward Models

Image Segmentation and Classification of E-waste for Training Robots for Waste Segregation

LaMP-Cap: Personalized Figure Caption Generation With Multimodal Figure Profiles

Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism

Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation

BroadGen: A Framework for Generating Effective and Efficient Advertiser Broad Match Keyphrase Recommendations

Large Language Models Implicitly Learn to See and Hear Just By Reading

AvatarShield: Visual Reinforcement Learning for Human-Centric Synthetic Video Detection

WavReward: Spoken Dialogue Models With Generalist Reward Evaluators

Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders

Pandora: A Code-Driven Large Language Model Agent for Unified Reasoning Across Diverse Structured Knowledge

Visual Chronicles: Using Multimodal LLMs to Analyze Massive Collections of Images

LookAhead Tuning: Safer Language Models via Partial Answer Previews

CaKE: Circuit-aware Editing Enables Generalizable Knowledge Learners

A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models

Union of Experts: Adapting Hierarchical Routing to Equivalently Decomposed Transformer

Leveraging Large Models to Evaluate Novel Content: A Case Study on Advertisement Creativity

Promote, Suppress, Iterate: How Language Models Answer One-to-Many Factual Queries

Can LLMs Explain Themselves Counterfactually?

Structure-prior Informed Diffusion Model for Graph Source Localization with Limited Data

THFlow: A Temporally Hierarchical Flow Matching Framework for 3D Peptide Design

LightThinker: Thinking Step-by-Step Compression

Purest Quantum State Identification

JL1-CD: A New Benchmark for Remote Sensing Change Detection and a Robust Multi-Teacher Knowledge Distillation Framework

Language Models Can Predict Their Own Behavior

DeepResonance: Enhancing Multimodal Music Understanding via Music-centric Multi-way Instruction Tuning

Without Paired Labeled Data: End-to-End Self-Supervised Learning for Drone-view Geo-Localization

Fine-Tuning is Subgraph Search: A New Lens on Learning Dynamics

EventVL: Understand Event Streams via Multimodal Large Language Model

Injecting Explainability and Lightweight Design into Weakly Supervised Video Anomaly Detection Systems

Biology-Instructions: A Dataset and Benchmark for Multi-Omics Sequence Understanding Capability of Large Language Models

Backdoor Attack with Invisible Triggers Based on Model Architecture Modification

Language Models as Causal Effect Generators

EMMA: End-to-End Multimodal Model for Autonomous Driving

Exploring Model Kinship for Merging Large Language Models

Post-hoc Study of Climate Microtargeting on Social Media Ads with LLMs: Thematic Insights and Fairness Evaluation

Variational Bayes Gaussian Splatting

DOTA: Distributional Test-Time Adaptation of Vision-Language Models

GALLa: Graph Aligned Large Language Models for Improved Source Code Understanding

LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture

Your Turn: At Home Turning Angle Estimation for Parkinson's Disease Severity Assessment

MediSyn: A Generalist Text-Guided Latent Diffusion Model For Diverse Medical Image Synthesis

Socially Pertinent Robots in Gerontological Healthcare

“What is Different Between These Datasets?” A Framework for Explaining Data Distribution Shifts

Is Pre-training Truly Better Than Meta-Learning?

Improving Image Captioning Descriptiveness by Ranking and LLM-based Fusion

Mitigating Strategy-Selection Bias in Reasoning for More Effective Test-Time Scaling

Virtual Arc Consistency for Linear Constraints in Cost Function Networks

A Multimodal Conversational Assistant for the Characterization of Agricultural Plots from Geospatial Open Data

Program Synthesis via Test-Time Transduction

OpenLens AI: Fully Autonomous Research Agent for Health Infomatics

Difficulty-Aware Agent Orchestration in LLM-Powered Workflows

Abduct, Act, Predict: Scaffolding Causal Inference for Automated Failure Attribution in Multi-Agent Systems

TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning

One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning

Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation

Created by

Haebom

Author

Muhammad Adnan, Nithesh Kurella, Akhil Arunkumar, Prashant J. Nair

Outline

Diffusion Transformers (DiT) achieve state-of-the-art results in text-to-image and text-to-video generation and editing, but video generation is computationally expensive due to the large model size and the quadratic cost of spatial-temporal attention across multiple denoising stages. Static caching mitigates this by reusing features at fixed stages, but fails to adapt to the generation dynamics, resulting in a suboptimal trade-off between speed and quality. In this paper, we propose Foresight, an adaptive layer reuse technique that reduces inter-stage computational redundancy across all layers while maintaining baseline performance. Foresight dynamically identifies and reuses inter-stage DiT block outputs across all layers to optimize efficiency by adapting to generation parameters such as resolution and denoising schedule. Implemented on OpenSora, Latte, and CogVideoX, Foresight achieves end-to-end speedups of up to \latencyimprv while maintaining video quality. The source code for Foresight is available at https://github.com/STAR-Laboratory/foresight .

Takeaways, Limitations

•

Takeaways: We present Foresight, a novel adaptive layer reuse technique that effectively reduces the computational cost of diffusion transformer-based video generation models. We demonstrate speedups on various models, including OpenSora, Latte, and CogVideoX. We also increase reproducibility and scalability by making the source code publicly available.

•

Limitations: The \latencyimprv value is not specifically provided, making it difficult to accurately assess the extent of the speedup. Further analysis is needed on generalization performance for various video generation models and its dependence on specific hardware environments. Analysis of changes in memory usage due to the application of Foresight is lacking.

View PDF

Made with Slashpage