The paper reports a reproducible GPU power benchmark for 18 open‑source LLMs (0.5B–7B parameters) run on a single consumer RTX 4060ti GPU using the Ollama inference engine. Energy metrics such as mean/peak power, total energy per prompt, energy per output token, and throughput were measured, revealing that model architecture and quantization strategy, rather than parameter count alone, drive energy efficiency. The most efficient models were qwen2.5:0.5b and tinyllama:1.1b, while the 7B‑Mistral model consumed up to 8.6× more energy per token, and qwen3.5:0.8b(on) showed unusually high per‑prompt energy due to extended internal reasoning.
By Philipp M. Z\"ahl, Elja Dalipaj, Anika Hennig, Timon Bayer
T-LoopFormer introduces token-level elastic-depth looped transformers that allow each token to decide its own number of loop iterations based on its hidden state, improving token generation accuracy. It also adds a recursion-wise key‑value cache so tokens at different depths only attend to their corresponding cached states, speeding up autoregressive decoding. Experiments demonstrate strong performance on language modeling and zero‑shot reasoning, achieving the lowest decoding latency among comparable models.
By Mingqian Yu, Wenpeng Zhang, Shaobo Cui, Peilin Zhao
The paper introduces OASIS, a method designed to stabilize dual‑normalized attention‑residual architectures by employing explicit null routing and token‑to‑depth null coupling. OASIS mitigates attention sinks and activation outliers, improving low‑bit quantization performance across several language‑model backbones. Empirical results show significant reductions in attention norms and perplexity, with notable gains on long‑context benchmarks.
By Haozheng Luo, Haoran Dai, Jingyuan Huang, Shaoyang Zhang, Xi Chen, Eric Hanchen Jiang, Yijiang Li, Chenghao Qiu, Chenwei Xu, Zhenyu Pan, Haotian Zhang, Binghui Wang, Yan Chen
The paper introduces a geometry-aware hyperbolic residual quantization method for variational autoencoders, addressing inconsistencies that arise when extending residual vector quantization to hyperbolic space. It restores telescoping behavior in the forward pass via Hyperbolic Residual Aggregation and improves gradient flow in the backward pass with a discounted Hyperbolic Straight-Through Estimator. Experiments on hierarchical prediction, recommendation, image tokenization, and neural audio coding demonstrate enhanced stability and structural organization of hyperbolic residual codes, while highlighting a trade‑off between compression efficiency and hierarchical organization.
By Alessio Colombo, Melika Ayoughi
arXiv:2609.30465v1 Announce Type: cross
Abstract: Mixture-of-experts (MoE) models activate few experts per token but store the full expert pool. Expert pruning reduces this storage burden; at a fixed...
By Mingyang Song, Mao Zheng
arXiv:2609.30820v1 Announce Type: cross
Abstract: Looped transformers reuse weights across recurrence steps, making low-bit quantization especially attractive. We identify two distinct failure modes...
By Nux Li
arXiv:2609.30840v1 Announce Type: cross
Abstract: One-step generators enable high-quality visual generation with a single network evaluation, but their post-training is difficult: general implicit ge...
By Austin Wang, Ziheng Cheng, Lexing Ying
arXiv:2609.31315v1 Announce Type: cross
Abstract: Unobserved common causes are pervasive in real-world time series and can induce spurious associations that causal discovery methods mistake for direc...
By Mohammad Fesanghary
arXiv:2609.31603v1 Announce Type: cross
Abstract: Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to i...
By Ali Holmov, Yiran Huang, Kirill Bykov, Zeynep Akata
arXiv:2609.30342v1 Announce Type: new
Abstract: iKFAD is a recently proposed optimiser that replaces adaptive learning rates with adaptive friction in the momentum dynamics, yet performs as well as A...
By Rajit Rajpal, Benedict Leimkuhler
arXiv:2609.30643v1 Announce Type: new
Abstract: We consider the problem of learning the underlying causal directed acyclic graph (DAG) structure corresponding to a structural equation model (SEM) wit...
By Anamitra Chaudhuri, Anirban Bhattacharya, Yang Ni
arXiv:2609.30738v1 Announce Type: new
Abstract: KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score,...
By Tianfang Xie, Wei Zhu
KuaFu is a unified behavior‑compression layer that reduces each user behavior item to 2–4 tokens, dramatically shrinking per‑item cache size while preserving fidelity through a four‑stage training process. In production across four profiling tasks, it matches or outperforms uncompressed single‑task models, boosts GPU throughput by 37–350%, and saves 190 GPUs. On public benchmarks it consistently beats prior compressors at the same compression ratio, and on RecBench a 4B KuaFu model outperforms its 8B counterpart by 1.90 points, contributing to a 1.37% lift in overall GMV on Tencent’s advertising and recommendation platform.
By Jiahao Hui, Lin Zhu, Yishen Hu, Jingdong Shu, Zetai Jiang, Xining Ran, Ben Tan, Yeshou Cai, Gong Chen, Haijie Gu, Jie Jiang
ZooWork-ShopRanker is a family of open e‑commerce rerankers (0.6B, 4B, and 8B) that align with human shopping preferences by using large language models as preference oracles to generate training pairs. The flagship 8B model serves as a teacher for the smaller 4B and 0.6B models, which are further refined on judged pairs. A new benchmark, ShopRank‑Bench, contains ~10,000 private‑traffic preference pairs and shows that all ZooWork models outperform the strongest open reranker baseline and their own un‑aligned versions.
By Siqiao Xue, Shuxuan Liu, Ning Hu
The paper introduces Routed Forcing, a method that improves audio‑driven streaming avatar generation by selectively applying different distillation objectives to semantic regions and noise stages. It uses Data‑Forcing Distillation on person regions at high noise levels to restore motion diversity, while retaining Distribution Matching Distillation for mouth and background to keep lip sync and scene stability. Experiments show up to 45% better dynamics and 7–25% higher diversity compared to the previous Self Forcing approach.
By Zihan Su, Siwen Lu, Junhao Zhuang, Zeyue Xue, Haoyang Huang, Guanghao Li, Xiaofeng Tan, Chun Yuan, Nan Duan
The paper introduces HetA-DiT, a heterogeneous attention mechanism for video diffusion models that allocates computation based on token difficulty. A lightweight uncertainty branch predicts denoising difficulty, routing uncertain tokens through dense global attention while applying efficient local attention to reliable tokens. This adaptive routing retains global context where needed, offers a single parameter to balance quality and efficiency, and achieves competitive generation quality while only about 20% of tokens use dense attention.
By Olga Zatsarynna, Denis Korzhenkov, Juergen Gall, Amir Habibian, Mohsen Ghafoorian
Scepsy is a serving system designed to efficiently schedule arbitrary multi‑LLM agentic workflows on GPU clusters. It leverages the observation that each LLM’s share of execution time remains relatively stable across requests, profiling LLMs under various parallelism levels to build an Aggregate LLM Pipeline that predicts throughput and latency. Using this predictor, Scepsy searches for optimal GPU allocations—balancing fractional GPU shares, tensor parallelism, and replica counts—and then heuristically places them on the cluster to reduce fragmentation and honor network topology, achieving up to 2.5× higher throughput and 1.0–3.3× lower latency compared to baseline approaches.
By Otto White, Marcel Wagenl\"ander, Britannio Jarrett, Xijin Zhao, Yanda Tao, Pedro Silvestre, Guo Li, Huanzhou Zhu, Llu\'is Vilanova, Peter Pietzuch
The paper introduces an importance-aware pruning method for diffusion models that is training‑free and focuses on preserving parameters essential to semantically salient image regions. By integrating spatial importance maps derived from conditioning signals or model attention into the pruning objective, the method generates parameter rankings that reflect perceptual relevance rather than uniform reconstruction error. Experiments on the MS‑COCO dataset show that this approach maintains subject fidelity and structural correctness even at high compression ratios, outperforming conventional pruning techniques that cause visible degradation.
By Ba-Thinh Lam, Srijan Das, Hieu Le
Assistive robots increasingly operate in many human-centered environments and perform various human-robot interaction (HRI) tasks, such as object delivery. However, most existing HRI systems rely on R...
Augur is a synthetic decision laboratory that simulates how users will react to product and policy changes before they are released. It constructs a typed knowledge graph from change documents, populates a persona market, runs simulations, and produces an auditable decision memo recommending one of five actions. Using a dataset of 50 real episodes (Gold‑50), the authors evaluate the system’s five‑way release verdicts and find that evaluation design, rather than model capability, largely drives performance differences among frontier and open‑weight models.
By Rahul Khedar, Mayank Malhotra, Avinash Karn