The paper investigates how temporal aggregation and ranking preservation affect aggressive decoding-time KV cache eviction. It shows that exponential‑moving‑average (EMA) aggregation makes many scorer modifications indistinguishable at the eviction‑set level, while certain variants (KeyDiff, key norm, recency, learned scorer) alter rankings and degrade performance. Building on this, the authors propose InertiaKV and its lazy variant, achieving 1.34–1.46× faster decoding, and also explore a score‑free decoding approach that slightly improves quality while eliminating further scoring.
By Bo Zeng, Yu Zhao, Yefeng Liu, Zhihong Lu, Xuanfan Ni, Xintong Wang
Para‑Pipe is a hierarchical mapping framework that combines intra‑ and inter‑stage operator parallelism within a pipelined architecture to optimize deep‑learning inference on heterogeneous System‑on‑Chip (SoC) platforms. By selectively tuning parallelism levels across pipeline stages, it balances throughput and latency while reducing inter‑processor communication overhead. Evaluations on Amlogic and Black Sesame SoCs show Pareto‑optimal configurations, with throughput‑optimized settings achieving up to 11.0 % higher energy efficiency than purely pipelined approaches and 23.3 % over non‑pipelined parallel execution.
The paper investigates how to combine on‑policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) for post‑training reasoning in large language models. It finds that a simple two‑stage approach—first applying OPD, then switching to RL—outperforms single‑signal methods and other joint baselines on logic and math reasoning benchmarks. The authors explain this advantage by showing that OPD broadens the model’s coverage of teacher‑supported solutions while RL sharpens performance within that support, and they provide practical guidance on when to switch based on OPD validation scores.
The paper introduces Hardware‑Aware FP4 FlashAttention‑4, which optimizes attention mechanisms for NVIDIA’s Blackwell 4‑bit floating‑point (FP4) tensor cores. By employing Direct‑P for noncausal inference and a causal path that forwards quantized scores into the backward pass, the method achieves up to 2.13× the bfloat16 forward throughput on an NVIDIA GB200. The causal approach also reconstructs probabilities from saved quantized queries and keys, using 8‑bit floating‑point (FP8) gradients to accelerate a full single‑GPU 8‑billion‑parameter update by up to 1.14×, while distributed training with FP8 probabilities and values shows divergent trajectories compared to tested MXFP4 setups.
We study the ordered Bernoulli-word kernel f(p,n,k)=p^k(1-p)^(n-k) and the geometry generated by its inverse-integer level sets. The binary level 2^(-n) selects p=1/2 as the unique real split-independent anchor.
The study examines how the geometric placement of synthetic data affects discourse-pragmatic function classification. Using 410 annotated instances of the word "look" and synthetic examples generated by Llama 3.1, the authors partitioned the synthetic data by cosine distance in RoBERTa space and compared six training conditions. While all augmentation conditions improved macro‑F and accuracy over a real‑only baseline, the nearest synthetic examples (NEAR) yielded the largest macro‑F gain (0.113) and a distance‑balanced mix achieved the highest accuracy (0.748), though none improved AUC.
SV-WAM is a surround‑view world‑action model that keeps all six camera views for autonomous driving while enabling efficient inference by discarding the video branch during deployment. It uses future‑video prediction as dense training supervision and introduces an action‑centered causal mask to prevent future‑video tokens from influencing action tokens during joint denoising. A differentiable drivable‑area compliance regularizer further improves safety by penalizing vehicle‑footprint corners that approach or cross drivable boundaries. Experiments on NAVSIMv2 and nuScenes show state‑of‑the‑art planning performance with low latency and strong zero‑shot transfer.
The paper investigates how temporal aggregation of token scores during decoding-time KV cache compression affects eviction decisions. It shows that using an exponential‑moving‑average (EMA) aggregation keeps ranking changes minimal for many scorer variants, while others like KeyDiff or recency significantly alter rankings and degrade performance. Building on this, the authors propose InertiaKV and its lazy variant, which achieve 1.34–1.46× faster decoding than full refresh, and also evaluate a score‑free approach that slightly improves quality while eliminating further scoring.
The paper introduces Random Attention, a method that evicts KV cache entries uniformly at random within each attention head while preserving the prompt. It demonstrates that this simple strategy matches or surpasses more complex eviction schemes across four models and six reasoning tasks, achieving 32‑43% higher throughput in vLLM deployments. Experiments reveal that the prompt is the most fragile cache component and that redundancy in the reasoning trace across text and attention heads protects against random eviction, eliminating the need for a selection score.
Spruce is a system that enables secure, private retrieval of large document collections outsourced to untrusted clouds by learning compact binary embeddings that preserve search quality while drastically reducing computation and communication. It replaces expensive corpus-wide embedding scoring with efficient Hamming-distance calculations under a two-server multi-party computation protocol, and introduces a fixed-radius protocol, private cluster pruning, and a one-core dealer to further cut latency and bandwidth usage. Across corpora ranging from 383K to 5.42M documents, Spruce maintains original search quality, achieving up to 6.7× faster full scans and 22.9× speedups with pruning, while retaining over 94% of the original NDCG.
ShallowStream is a framework for streaming video understanding that uses the shallow layers of a multimodal large language model (MLLM) to encode frames and build a lightweight index. During streaming, it maintains an always‑on index via the KV cache of shallow layers, and at query time it scores context frames using shallow‑layer attention and selects diverse evidence for answering. The approach matches the performance of leading streaming methods while cutting per‑frame prefill latency and 10‑second end‑to‑end latency by up to 52.1× and 11.9×, respectively.
By Jitai Hao, Ke Yang, Qiang Huang, Jun Yu
The paper presents a method for improving markerless 3D pose estimation in infants by cross‑model distillation. Using unannotated infant video, a frozen Sapiens 2 pose model provides dense pseudo‑labels that guide fine‑tuning of the SAM 3D Body model. On a held‑out dataset of eleven infants, the fine‑tuned model shows significant gains in 2D keypoint accuracy and 3D joint error compared to the original SAM 3D Body model.
By R. James Cotton, Divya Joshi, Colleen Peyton
The paper introduces Latent Drift, a generative forecasting framework that predicts slow-evolving neurodegenerative disease progression by learning changes in a compressed semantic representation rather than full-resolution anatomy. It addresses two failure modes—identity collapse and continuous interpolation trap—by removing pixel-level identity from the prediction target and applying Finite Scalar Quantization to suppress high-frequency nuisance fluctuations. Experiments on longitudinal 3D brain MRI demonstrate that Latent Drift outperforms diffusion and autoregressive transformer baselines in both generative fidelity and clinically relevant metrics.
By Yuxiang Feng, Juncheng Wang, Chao Xu, Wenlong Hou, Huihan Wang, Yijie Qian, Yang Liu, Baigui Sun, Yong Liu, Shujun Wang
The paper introduces LEADer, a framework that uses local epistemic uncertainty to guide active sampling in diffusion-based image restoration. By adjusting prior strength per pixel and pruning sampling trajectories based on uncertainty traces, LEADer balances detail preservation with artifact suppression and accelerates convergence. The method is plug‑and‑play, theoretically guarantees data consistency and stable convergence, and improves performance across multiple state‑of‑the‑art diffusion models with minimal memory overhead.
By Jiaqi Zhang, Zheng Pang, Rongrong Gao, Qiyuan Zhang, Yang Yang
The paper introduces Repo-To-Skill, a method for converting GitHub repositories into reusable AI skills. By distilling operational knowledge from over 1,000 machine‑learning repositories, the authors build the AREX‑Skill Library with more than 5,000 verified skills across 20 areas. Integrating these skills into a research agent—DisCo—yields significant performance boosts on multiple benchmarks, demonstrating the value of reusable, task‑agnostic knowledge.
By Jianlyu Chen, Yuyang Hu, Hongjin Qian, Jiawei Liu, Wenqing Wei, Xiaolong Chen, Defu Lian, Zhicheng Dou, Chaozhuo Li, Qiwei Ye, Zheng Liu
ProSR is a new method for Synthetic Aperture Radar (SAR) super‑resolution that treats the task as a discrete token prediction problem in a quantized latent space. By mapping SAR signal features to discrete scattering primitives and using a self‑supervised backbone to extract label‑free semantic priors, ProSR preserves the impulsive, physically consistent scattering statistics of SAR images. The approach includes Semantic‑Aligned Detail Encoding and a Prototype‑Map‑Guided Attention mechanism, and it has been validated on a large‑scale 0.25 m resolution benchmark from the Umbra Open Dataset, achieving superior visual quality while maintaining essential scattering characteristics.
By Byoungwoo Kim, Munchurl Kim
The paper introduces a multi‑camera computer vision system that computes Post‑Encroachment Time (PET) in real time to assess intersection safety. Using four synchronized cameras and YOLOv11 segmentation on NVIDIA Jetson AGX Xavier devices, vehicle detections are mapped to a unified bird’s‑eye view and processed with a pixel‑level PET algorithm to generate high‑resolution heatmaps. The system records PET data in an SQL database and demonstrates real‑time throughput (2.68 FPS) with 800 × 800 heatmaps, validating a scalable, decentralized approach for high‑resolution intersection safety evaluation.
By Shounak Ray Chaudhuri, Arash Jahangiri, Christopher Paolini
The paper introduces ReTA, a reinforcement‑learning framework that dynamically decides how to augment electronic health record (EHR) graphs with external knowledge graphs (KGs) on a per‑visit basis. ReTA offers three actions—Soft Import, Hard Import, and Skip—allowing it to enrich node features, graft compact KG subgraphs, or leave the graph unchanged depending on the patient’s evolving state and a budget constraint. Experiments on MIMIC‑III and MIMIC‑IV demonstrate that ReTA consistently outperforms strong baselines, transfers across datasets and KGs, and provides interpretable augmentation patterns while maintaining efficiency.
By Chen Chen, Mohsen Nayebi Kerdabadi, Dongjie Wang, Mei Liu, Zijun Yao
The paper introduces a unified rate–distortion framework for discrete visual tokenization, encompassing vector, product, and scalar quantization. It shows that minimizing distortion, rather than maximizing codebook utilization, is the key objective for reconstruction fidelity and establishes fairness conditions for comparing quantizers. Under these conditions, the study confirms the distortion hierarchy VQ–PQ–SQ and demonstrates that modern VQ methods achieve the lowest distortion.
By Xianghong Fang, Wenlong Mou, Yuan Yuan, Dehan Kong, Tim G. J. Rudner
The paper introduces UnInfo, a test‑time adaptation method for vision‑language models like CLIP that addresses image corruption—a realistic distribution shift caused by sensor conditions. UnInfo leverages uniformity‑aware confidence maximization, information‑aware loss balancing, and knowledge distillation from an EMA teacher to preserve embedding uniformity and improve zero‑shot classification accuracy. Experiments show that UnInfo outperforms existing TTA methods on corrupted image datasets.
By Kazuki Adachi, Shin'ya Yamaguchi, Tomoki Hamagami