Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
arXiv:2609.26173v1 Announce Type: new Abstract: Many post-training quantization (PTQ) methods use layer-wise reconstruction, second-order proxy objectives, or activation-aware transformations to redu...
arXiv:2609.26199v1 Announce Type: new Abstract: A large graph is often available only in part: a crawl stopped by its budget, a panel, a partial dump. When the sampled fraction $s$ is known by design...
arXiv:2609.26216v1 Announce Type: new Abstract: A correct teacher solution becomes useful supervision when the receiving student can continue its reasoning. We measure this compatibility with prefix...
arXiv:2609.26402v1 Announce Type: new Abstract: The discovery of novel inorganic materials drives technological breakthroughs in critical fields such as computing and energy storage. Generative AI ha...
arXiv:2609.26219v1 Announce Type: cross Abstract: Long-running LLM agent workflows often revise interior context spans while retaining long suffixes. Although suffix tokens remain unchanged, altered...
arXiv:2609.25008v1 Announce Type: new Abstract: I pretrained a language model end-to-end in Rust - alone, with no team, no PyTorch, and no Python in the training path - for $164 in rented GPU time. I...
arXiv:2609.25054v1 Announce Type: new Abstract: For a long-horizon LLM agent, the memory question is not what was once recorded but what \emph{currently holds}. Most designs answer it only indirectly...
arXiv:2609.25056v1 Announce Type: new Abstract: A feedback-based word deduction framework based on the Jotto problem is proposed, and the problem space is represented as a weighted graph where all va...
arXiv:2609.26638v1 Announce Type: new Abstract: Autoregressive OCR vision-language models accurately convert document images into text and structured markup, but require one sequential decoding step...
arXiv:2503.15242v3 Announce Type: replace Abstract: We introduce BigO(Bench), a novel coding benchmark designed to evaluate the capabilities of generative language models in understanding and generat...
arXiv:2609.25891v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, as critical evidence may o...
arXiv:2605.30263v2 Announce Type: replace Abstract: Recent video diffusion foundation models have achieved remarkable progress in high-quality video generation, yet turning them into real-time intera...
S$^4$R is a method for compressing the Key-Value cache in large language models by building low‑rank subspaces from selectively sampled tokens and performing attention over a sparsely reconstructed KV representation. It initializes key/value bases using a representative prompt subset, reducing reliance on external calibration data while avoiding the high compute cost of full‑prompt decomposition. Experiments on LongBench and RULER with Llama and Qwen models demonstrate up to 5× KV compression with near‑full‑cache accuracy, blending the efficiency of fixed compression with the adaptability of prompt‑dependent approaches.
KwaiMind is a commercial image editing system that combines general editing capabilities with e-commerce specialization. It uses an agent-based data engine with 1.8 million editing pairs and a multimodal diffusion transformer trained through pre‑training, fine‑tuning, preference optimization, and online reinforcement learning. The system is guided by a vision‑language judge and specialized rewards for click‑through rate, text rendering, and product consistency, and it achieves top scores on ImgEdit, GEdit, REDEdit, and the new Ecom‑Bench, while improving predicted and actual CTR in offline and online experiments.
Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historic...
Transformer language models process sequences token by token in an autoregressive manner, making growing contexts increasingly expensive. Yet many adjacent token spans are highly predictable or freque...
PRQuant introduces a training‑free, low‑overhead method for low‑bit quantization of linear layers by permuting input channels that cause the largest quantization error into contiguous tail blocks and precomputing residual weight sub‑tensors. The approach eliminates the need for online gathering during inference, converting scattered residual compensation into a regular tail‑augmented GEMM and thereby reducing latency. Experiments show that PRQuant lowers down‑projection reconstruction error and outperforms standard MXFP4 and other post‑training quantization baselines on five downstream benchmarks, improving accuracy by up to 1.24 points on Qwen3‑4B‑Instruct‑2507.
Moonworks Lunara is a text‑to‑image model that defines Artistic Intelligence as exploration‑driven world realization, preserving semantic, artistic, and compositional structure. It uses a Diffusion Mixture Transformer architecture and a training algorithm that iteratively refines the data distribution with informative samples and human‑created art. Benchmarks show Lunara ranks first in aesthetic quality and second in emotional resonance against seven other image‑generation models, while maintaining a sub‑10B parameter size and sub‑10‑second inference latency.
MTMed3D is a multi-task Transformer-based model that jointly performs 3D detection, segmentation, and classification in medical imaging. It uses a shared Transformer encoder to produce multi-scale features, with separate CNN decoders for each task. Evaluated on BraTS 2018 and 2019, it achieves strong results, especially in detection, while reducing computational cost and inference time compared to single-task models.
StepKV introduces a step-aware approach to compressing the key-value cache used during large language model inference, treating reasoning steps as primary units of retention rather than individual tokens. By linking cache entries to the steps that generated them and estimating each step’s utility from trajectory signals, StepKV assigns a combined token‑ and step‑level score to guide pruning. Experiments on multi‑hop question answering and long‑horizon web reasoning show that StepKV maintains accuracy even under tight cache budgets, outperforming token‑level baselines that suffer sharp performance drops.