Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
arXiv:2608.00311v2 Announce Type: replace Abstract: Long-context inference with large language models (LLMs) is costly: self-attention during prefill scales quadratically with sequence length, and th...
arXiv:2604.19254v2 Announce Type: replace Abstract: Popular low-rank parameter-efficient fine-tuning (PEFT) methods represent adaptation as separate updates to selected backbone weights, without main...
HoliBench is a modular benchmarking and deployment toolkit that jointly measures accuracy, latency, and energy for foundation models across a wide range of devices, from single-board computers to GPU servers. It provides a platform abstraction layer that calibrates cross-device measurements and supports multiple model modalities, inference engines, and quantization levels. Using HoliBench, the authors evaluated 20 models on 7 device types, revealing tradeoffs such as limited latency gains from quantization on low‑precision hardware and diminishing accuracy returns relative to energy consumption, while also showing that single-model profiles can predict multi-model pipeline performance within a few percent.
The paper introduces Neuro‑JEPA, a sparse multimodal foundation model that learns unified representations of brain MRI across T1w, T2w, and FLAIR sequences using a latent predictive objective and a Mixture‑of‑Experts architecture. It was pretrained on over 1.5 million scans from 428,647 studies and systematically evaluates architectural, masking, objective, and sparsity choices for robust multimodal representation learning. Across 47 tasks from three health systems and 12 public datasets, Neuro‑JEPA consistently outperforms a simple CNN baseline, demonstrating its effectiveness for both clinical and research applications.
arXiv:2607.08098v2 Announce Type: replace Abstract: Event cameras are increasingly adopted in embodied perception for their microsecond temporal resolution, high dynamic range, and resilience to moti...
The paper introduces SynthSentry, a model‑agnostic method for detecting synthetic data contamination in language‑model training corpora. It computes a distributional divergence score based on lexical diversity collapse, n‑gram tail truncation, and perplexity variance across reference models, requiring no access to the generating model or synthetic labels. Experiments on English corpora contaminated by small open‑weight generators and an instruction‑tuned model show that SynthSentry ranks contamination severity accurately, maintains low false‑positive rates after calibration, and does not degrade downstream fine‑tuning performance at the tested scale.
AsyncFlow is an asynchronous streaming reinforcement learning framework designed to improve the post‑training phase of large language models. It introduces a distributed data storage and transfer module that enables panoramic data management and fine‑grained scheduling, allowing automated pipeline overlapping and dynamic load balancing. The framework also employs an asynchronous producer‑consumer workflow to reduce computational idleness by deferring parameter updates within staleness thresholds, and it is architecturally decoupled from training and inference engines, providing modular, customizable user interfaces. Experiments show an average throughput improvement of 1.59× over the state‑of‑the‑art baseline.
Semantically Aligned Gradient-Driven Context-Preserving Image Editing (IABEdit) is a model‑agnostic framework that embeds differentiable semantic verification into the training of generative image editors. By using a frozen vision‑language model to extract spatially‑aware descriptors from ground‑truth edits and a trainable aligner to reproduce them from generated outputs, the residual becomes a gradient that teaches the generator both what to edit and where, without adding inference‑time VLM cost. IABEdit is compatible with various backbones (e.g., U‑Net in Stable Diffusion and MMDiT in FLUX) and improves structural fidelity on MagicBrush, achieves state‑of‑the‑art instruction adherence on RealEdit and EMU Edit, and outperforms the proprietary Gemini agent on the D‑LORD surveillance benchmark under heavy occlusion. "whyItMatters":"IABEdit demonstrates that incorporating semantic verification during training can produce more accurate, well‑localized edits and outperform existing methods even in challenging surveillance scenarios, as shown by its superior metrics and human/GPT‑4o evaluations."
The paper investigates on-device language models (ODLMs) for predicting stress in a mobile health context, focusing on privacy-preserving, cloud-independent inference. Using zero‑shot prompting, the authors evaluate ODLMs across multimodal data—objective sensor features and subjective self‑reports—measuring predictive accuracy, latency, and throughput. Results indicate that sensor features slightly outperform self‑reports, and that lightweight sub‑2B models deliver low latency with predictable resource usage, underscoring both the potential and practical limits of ODLMs for mobile mental health.
The paper introduces PACEvolve, a framework that improves self‑evolving agents powered by Large Language Models by addressing their tendency to become trapped in local contexts and repeat flawed hypotheses. It does so through three techniques: Hierarchical Context Management to prune memory, Momentum‑Based Backtracking to escape local minima, and a self‑adaptive Collaborative Evolution policy to balance refinement and knowledge transfer. These methods enable the agents to maintain a global view of search momentum and achieve state‑of‑the‑art results on complex evolutionary benchmarks.
The paper investigates how the time cost of 802.11 contention affects federated learning. Using ns-3 simulations, the authors measure frame-delivery ratios and saturation throughput across various client densities and offered loads, then employ a FedAvg trainer that uses these ratios to estimate communication time. Across 720 runs with diverse datasets, partitions, densities, loads, and seeds, all models reached target accuracy within the round budget, with communication time-to-target increasing significantly as client density rose. The study also compares uniform and persistent heterogeneous participation, finding no statistically significant accuracy gap, though confidence intervals are wide. The results are specific to the evaluated configurations and do not generalize to all convergence or fairness scenarios.
The paper investigates how block‑diffusion language models can use a constant‑size cache to enable efficient parallel decoding. By employing sequence mixers that summarize completed blocks into a reusable state and a block‑causal training objective, the authors pretrain three 3B block‑diffusion denoisers (attention, Mamba, and hybrid) on 300 B tokens. The resulting state‑space cache remains O(1) in memory and latency regardless of context length, yielding significant speed‑up and memory savings compared to traditional attention‑based caches, especially at very long sequences.
The paper introduces a self‑supervised neural network that unifies single‑frame Fresnel coherent diffraction imaging (CDI) and overlapped ptychography. By using a fixed, pre‑estimated probe and optimizing with a Poisson negative log‑likelihood objective, the method reconstructs object patches from either a single diffraction frame or multiple overlapping measurements, achieving high SSIM scores and a ten‑fold improvement in photon‑dose efficiency. Demonstrations on synthetic patterns and real datasets from APS and LCLS show robust, high‑throughput reconstructions, with a 36× speedup over iterative solvers for a 10,304‑frame workload.
arXiv:2609.06729v2 Announce Type: replace Abstract: Accurate 3D brain tumour segmentation from multi-modal Magnetic Resonance Imaging (MRI) is essential for clinical diagnosis and treatment planning....
Federated learning enables collaborative model development without centralizing patient data, but annotation reliability at participating institutions cannot always be assumed. In ophthalmic imaging,...
The paper introduces a tail‑risk‑aware scheduling strategy for agentic LLM workflows that decouples readiness from immediate release of model turns. By jointly selecting which ready turn to release and controlling the amount of unfinished work kept in the queue, the method uses a mean‑CVaR objective to adapt to evolving tail risk and online turn‑work estimates. Experiments on real software‑engineering task traces show comparable performance to eager release under light load and a significant reduction in the 95th‑percentile workflow flow time, achieving up to a 3.5× speedup under contention.
The paper introduces SOLID, a framework that enables operations research language models to self-improve without relying on verified answers or external evaluators. SOLID uses solver-generated artifacts from the model’s own rollouts to create pseudo-references, clustering objectives and applying group-relative advantages for dense self-supervision. Experiments on multiple OR benchmarks show that SOLID enhances solution accuracy for both general-purpose and OR-tuned models compared to outcome-only training.
The paper introduces the Power Flexibility Index (PFI) to measure how large language model (LLM) training performance changes when GPU power is reduced. Using 131 training runs on H200 and H100 GPUs, the study finds that LLM jobs have significant but variable power elasticity and identifies telemetry signals that can predict PFI during runtime. The authors demonstrate that allocating power based on PFI maximizes overall token throughput, recovering about 1.5k tokens/s per job under a 30% power reduction, which represents 63% of the gap between equal-weight and perfect-information allocations.
X-AuT is a progressive framework for compressing the audio encoder of speech large language models. It selects layer combinations via short behavioral probes and restores performance through representation alignment, cross‑scale distillation, scheduled student‑policy supervision, and LoRA finetuning, while keeping the language‑model backbone frozen. On ten Chinese–English benchmarks, reducing Qwen3‑ASR‑0.6B’s encoder from 18 to 16 layers lowers macro‑average error from 5.61% to 5.27%, and a 14‑layer model achieves 5.75% error with 20.7% fewer parameters.
The paper introduces SurgicalRoomAgent, a voice‑interactive multi‑agent system for smart operating rooms that leverages large language models to understand natural language, control devices, record intraoperative events, and generate surgical reports. Its layered architecture includes a voice interaction pipeline (wake, ASR, turn detection, agent reasoning, TTS) and an agent core (skill registry, task planner, device manager). Three key technologies—KV Cache prefix warming, streaming partial JSON parsing with early parallel task execution, and progressive skill prompt disclosure—reduce latency and maximize context efficiency, enabling real‑time operation within a 16,384‑token limit.