Efficient Serving of LLM Applications with Probabilistic Demand Modeling

Applications based on Large Language Models (LLMs) contains a series of tasks to address real-world problems with boosted capability, which have dynamic demand volumes on diverse backends. Existing serving systems treat the resource demands of LLM applications as a blackbox, compromising end-to-end efficiency due to improper queuing order and backend warm up latency. We find that the resource demands of LLM applications can be modeled in a general and accurate manner with Probabilistic Demand Graph (PDGraph). We then propose Hermes, which leverages PDGraph for efficient serving of LLM applications. Confronting probabilistic demand description, Hermes applies the Gittins policy to determine the scheduling order that can minimize the average application completion time. It also uses the PDGraph model to help prewarm cold backends at proper moments. Experiments with diverse LLM applications confirm that Hermes can effectively improve the application serving efficiency, reducing the average completion time by over 70% and the P95 completion time by over 80%.

SUSEP-Net: Simulation-Supervised and Contrastive Learning-based Deep Neural Networks for Susceptibility Source Separation

Quantitative susceptibility mapping (QSM) provides a valuable tool for quantifying susceptibility distributions in human brains; however, two types of opposing susceptibility sources (i.e., paramagnetic and diamagnetic), may coexist in a single voxel, and cancel each other out in net QSM images. Susceptibility source separation techniques enable the extraction of sub-voxel information from QSM maps. This study proposes a novel SUSEP-Net for susceptibility source separation by training a dual-branch U-net with a simulation-supervised training strategy. In addition, a contrastive learning framework is included to explicitly impose similarity-based constraints between the branch-specific guidance features in specially-designed encoders and the latent features in the decoders. Comprehensive experiments were carried out on both simulated and in vivo data, including healthy subjects and patients with pathological conditions, to compare SUSEP-Net with three state-of-the-art susceptibility source separation methods (i.e., APART-QSM, \c{hi}-separation, and \c{hi}-sepnet). SUSEP-Net consistently showed improved results compared with the other three methods, with better numerical metrics, improved high-intensity hemorrhage and calcification lesion contrasts, and reduced artifacts in brains with pathological conditions. In addition, experiments on an agarose gel phantom data were conducted to validate the accuracy and the generalization capability of SUSEP-Net.

Efficient Unified Caching for Accelerating Heterogeneous AI Workloads

Modern AI clusters, which host diverse workloads like data pre-processing, training and inference, often store the large-volume data in cloud storage and employ caching frameworks to facilitate remote data access. To avoid code-intrusion complexity and minimize cache space wastage, it is desirable to maintain a unified cache shared by all the workloads. However, existing cache management strategies, designed for specific workloads, struggle to handle the heterogeneous AI workloads in a cluster -- which usually exhibit heterogeneous access patterns and item storage granularities. In this paper, we propose IGTCache, a unified, high-efficacy cache for modern AI clusters. IGTCache leverages a hierarchical access abstraction, AccessStreamTree, to organize the recent data accesses in a tree structure, facilitating access pattern detection at various granularities. Using this abstraction, IGTCache applies hypothesis testing to categorize data access patterns as sequential, random, or skewed. Based on these detected access patterns and granularities, IGTCache tailors optimal cache management strategies including prefetching, eviction, and space allocation accordingly. Experimental results show that IGTCache increases the cache hit ratio by 55.6% over state-of-the-art caching frameworks, reducing the overall job completion time by 52.2%.

Astraea: A GPU-Oriented Token-wise Acceleration Framework for Video Diffusion Transformers

Video diffusion transformers (vDiTs) have made impressive progress in text-to-video generation, but their high computational demands present major challenges for practical deployment. While existing acceleration methods reduce workload at various granularities, they often rely on heuristics, limiting their applicability. We introduce ASTRAEA, an automatic framework that searches for near-optimal configurations for vDiT-based video generation. At its core, ASTRAEA proposes a lightweight token selection mechanism and a memory-efficient, GPU-parallel sparse attention strategy, enabling linear reductions in execution time with minimal impact on generation quality. To determine optimal token reduction for different timesteps, we further design a search framework that leverages a classic evolutionary algorithm to automatically determine the distribution of the token budget effectively. Together, ASTRAEA achieves up to 2.4x inference speedup on a single GPU with great scalability (up to 13.2x speedup on 8 GPUs) while retaining better video quality compared to the state-of-the-art methods (<0.5% loss on the VBench score compared to the baseline vDiT models).

Pts3D-LLM: Studying the Impact of Token Structure for 3D Scene Understanding With Large Language Models

Effectively representing 3D scenes for Multimodal Large Language Models (MLLMs) is crucial yet challenging. Existing approaches commonly only rely on 2D image features and use varied tokenization approaches. This work presents a rigorous study of 3D token structures, systematically comparing video-based and point-based representations while maintaining consistent model backbones and parameters. We propose a novel approach that enriches visual tokens by incorporating 3D point cloud features from a Sonata pretrained Point Transformer V3 encoder. Our experiments demonstrate that merging explicit 3D features significantly boosts performance. Furthermore, we show that point-based token structures can rival video-based ones when the points are cleverly sampled and ordered. Our best models from both structures achieve state-of-the-art results on multiple 3D understanding benchmarks. We emphasize our analysis of token structures as a key contribution, alongside transparent reporting of results averaged over multiple seeds, a practice we believe is vital for robust progress in the field.

Adversarial Preference Learning for Robust LLM Alignment

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

Cross-Modal Characterization of Thin Film MoS$_2$ Using Generative Models

The growth and characterization of materials using empirical optimization typically requires a significant amount of expert time, experience, and resources. Several complementary characterization methods are routinely performed to determine the quality and properties of a grown sample. Machine learning (ML) can support the conventional approaches by using historical data to guide and provide speed and efficiency to the growth and characterization of materials. Specifically, ML can provide quantitative information from characterization data that is typically obtained from a different modality. In this study, we have investigated the feasibility of projecting the quantitative metric from microscopy measurements, such as atomic force microscopy (AFM), using data obtained from spectroscopy measurements, like Raman spectroscopy. Generative models were also trained to generate the full and specific features of the Raman and photoluminescence spectra from each other and the AFM images of the thin film MoS$_2$. The results are promising and have provided a foundational guide for the use of ML for the cross-modal characterization of materials for their accelerated, efficient, and cost-effective discovery.

DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers

Visual Prompt Tuning (VPT) has become a promising solution for Parameter-Efficient Fine-Tuning (PEFT) approach for Vision Transformer (ViT) models by partially fine-tuning learnable tokens while keeping most model parameters frozen. Recent research has explored modifying the connection structures of the prompts. However, the fundamental correlation and distribution between the prompts and image tokens remain unexplored. In this paper, we leverage metric learning techniques to investigate how the distribution of prompts affects fine-tuning performance. Specifically, we propose a novel framework, Distribution Aware Visual Prompt Tuning (DA-VPT), to guide the distributions of the prompts by learning the distance metric from their class-related semantic data. Our method demonstrates that the prompts can serve as an effective bridge to share semantic information between image patches and the class token. We extensively evaluated our approach on popular benchmarks in both recognition and segmentation tasks. The results demonstrate that our approach enables more effective and efficient fine-tuning of ViT models by leveraging semantic information to guide the learning of the prompts, leading to improved performance on various downstream vision tasks.

DarkDiff: Advancing Low-Light Raw Enhancement by Retasking Diffusion Models for Camera ISP

High-quality photography in extreme low-light conditions is challenging but impactful for digital cameras. With advanced computing hardware, traditional camera image signal processor (ISP) algorithms are gradually being replaced by efficient deep networks that enhance noisy raw images more intelligently. However, existing regression-based models often minimize pixel errors and result in oversmoothing of low-light photos or deep shadows. Recent work has attempted to address this limitation by training a diffusion model from scratch, yet those models still struggle to recover sharp image details and accurate colors. We introduce a novel framework to enhance low-light raw images by retasking pre-trained generative diffusion models with the camera ISP. Extensive experiments demonstrate that our method outperforms the state-of-the-art in perceptual quality across three challenging low-light raw image benchmarks.

SoloSpeech: Enhancing Intelligibility and Quality in Target Speech Extraction through a Cascaded Generative Pipeline

Target Speech Extraction (TSE) aims to isolate a target speaker's voice from a mixture of multiple speakers by leveraging speaker-specific cues, typically provided as auxiliary audio (a.k.a. cue audio). Although recent advancements in TSE have primarily employed discriminative models that offer high perceptual quality, these models often introduce unwanted artifacts, reduce naturalness, and are sensitive to discrepancies between training and testing environments. On the other hand, generative models for TSE lag in perceptual quality and intelligibility. To address these challenges, we present SoloSpeech, a novel cascaded generative pipeline that integrates compression, extraction, reconstruction, and correction processes. SoloSpeech features a speaker-embedding-free target extractor that utilizes conditional information from the cue audio's latent space, aligning it with the mixture audio's latent space to prevent mismatches. Evaluated on the widely-used Libri2Mix dataset, SoloSpeech achieves the new state-of-the-art intelligibility and quality in target speech extraction and speech separation tasks while demonstrating exceptional generalization on out-of-domain data and real-world scenarios.