#efficient inference
14 resultsCalibrate Before Reason: Robust Visual Token Reduction against Semantic Drift in VLMs
Jiasheng Li, Zhong Ji, Yan Zhang +1
The paper proposes CaRe, a training‑free method that calibrates compact visual representations before reasoning to keep semantic consistency when reducing visual tokens in large vi…
Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently
Simon Roy, Mark Bong, Giovanni Beltrame
The paper studies how to efficiently combine before-and-after satellite images to match natural‑language change queries, comparing attention, state‑space (Mamba), and compressed fu…
ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding
Mingkang Dong, Muxin Pu, Jie Li +8
ObjectStream introduces a training‑free method that extracts latent objects from frozen Video‑LLM representations and uses them as persistent memory anchors to improve streaming vi…
Penelope: Localized Latent Recurrence for Efficient Structured Reasoning
Yutong Chen, Shouqian Shi, Xinran Liu +5
Penelope introduces a method that adds a localized recurrent computation within a decoder-only Transformer to perform structured reasoning efficiently, using a latent space instead…
GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch
GigaWorld Team, Angen Ye, Angyuan Ma +26
The paper introduces GigaWorld-Policy-0.5, a robot control model that learns from future visual dynamics during training but generates actions only at inference, achieving faster (…
VideoSEMA: a scalable and efficient Mamba-like attention for video understanding
Nhat Thanh Tran, Fanghui Xue andShuai Zhang, Fanghui Xue +5
The paper introduces VideoSEMA, a split space‑time attention model for video classification that combines a scalable Mamba‑like spatial attention block with softmax temporal attent…
Attention-Free and Lightweight Token Reduction for Efficient Vision-Language Models
Xuanyi Hao, Zuoyuan Zhang, Zhibo Wang +4
The paper introduces a plug‑and‑play, attention‑free token reduction module for vision‑language models that selects informative and diverse visual tokens using an entropy‑based imp…
FlashDiff: Efficient Regional Execution and Scheduling for Diffusion Model Serving
Yaqi Qiao, Ping He, Songrun Xie +4
FlashDiff is a system that speeds up diffusion model inference by dynamically selecting which latent regions need further processing and efficiently scheduling those regions across…
SlimPer: Make Personalization Model Slim and Smart
Siqi Wang, Xianjie Chen, Shaofeng Deng +42
SlimPer is a transformer‑based recommendation model that treats personalized ranking as iterative refinement of a compact user‑item knowledge base, achieving linear per‑layer cost…
AVQ-Attention: Adaptive Vector-Quantized Attention
Winfried van den dool, Patrick Forré, Amir Habibian +2
The paper introduces Adaptive Vector-Quantized (AVQ) Attention, which dynamically allocates codebook capacity to the most important regions of the key space, preserving O(MN) compl…
EMAGN: Efficient Multi-Attention Graph Network via Learned Clustering for Scalable Traffic Forecasting
Mingxing Xu, Rakesh Chowdary Machineni, Ke Liu +7
The paper introduces EMAGN, a graph neural network that uses learned clustering to linearize multi-head attention, enabling scalable traffic forecasting with reduced computation an…
Recursive Multi-Agent Systems
Jiaru Zou, Rui Pan, Ruizhong Qiu +8
The paper proposes RecursiveMAS, a framework that treats a multi-agent system as a recursive latent‑space computation, enabling agents to iteratively refine each other's thoughts a…
Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers
Ying Fan, Anej Svete, Kangwook Lee
The paper introduces LOTUS, a looped Transformer architecture that performs multi-step reasoning in latent space, achieving reasoning performance comparable to explicit chain-of-th…
Mobility-Aware Cache Framework for Scalable LLM-Based Human Mobility Simulation
Hua Yan, Heng Tan, Yingxue Zhang +1
The paper introduces MobCache, a mobility-aware caching framework that stores and reuses latent-space embeddings of LLM reasoning steps to enable efficient large-scale human mobili…