LayerSkip combines layer dropout and early exit with self-speculative decoding, achieving up to 2.16× speedup in large language model inference.

Mostafa Elhoushi, Akshat Shrivastava, Diana Liskovich et al.

Data-to-paper platform autonomously generates verifiable research with 80-90% accuracy.

Tal Ifargan, Lukas Hafner, Maor Kern et al.

GraphRAG leverages hierarchical knowledge graphs and community detection to enable large-scale global understanding, outperforming traditional RAG by 25% on complex datasets.

Darren Edge, Ha Trinh, Newman Cheng et al.

Proposes a supervised uncertainty estimation framework for LLMs using hidden layer activations and external features, outperforming unsupervised baselines.

Linyu Liu, Yu Pan, Xiaocheng Li et al.

ViViDex leverages human videos, combining trajectory-guided reinforcement learning and visual distillation, to enable dexterous multi-finger robot manipulation with minimal demonstration data.

Zerui Chen, Shizhe Chen, Etienne Arlaud et al.

Proposed CoST uses contrastive quantization to improve semantic tokens, boosting Recall@5 by 43%.

Jieming Zhu, Mengqun Jin, Qijiong Liu et al.

Proposes 'Align Your Steps' framework to optimize diffusion model sampling schedules, significantly improving quality in few-step generation.

Amirmojtaba Sabour, Sanja Fidler, Karsten Kreis

SnapKV compresses KV cache by analyzing attention patterns from observation windows, enabling 3.6x faster decoding and handling 380K tokens without fine-tuning.

Yuhong Li, Yingbing Huang, Bowen Yang et al.

Retinexformer combines Transformer-based multi-scale feature fusion and illumination-guided modules, achieving PSNR of 25.30dB on ultra-high-resolution low-light images.

Xiaoning Liu, Zongwei Wu, Ao Li et al.

phi-3-mini is a 3.8B parameter model trained on 3.3T tokens, rivaling GPT-3.5, deployable on smartphones.

Marah Abdin, Jyoti Aneja, Hany Awadalla et al.

Study on LLaMA3 quantization reveals performance drop at low bit-widths using methods like GPTQ and AWQ.

Wei Huang, Xingyu Zheng, Xudong Ma et al.

Introduces instruction hierarchy training, significantly boosting model robustness against prompt injections and jailbreaks.

Eric Wallace, Kai Xiao, Reimar Leike et al.

MoVA adapts a mixture of vision experts to multimodal contexts, enhancing performance through a coarse-to-fine mechanism.

Zhuofan Zong, Bingqi Ma, Dazhong Shen et al.

PhysDreamer leverages video generation priors to estimate material fields, enabling realistic interactive 3D object dynamics without real-world material data.

Tianyuan Zhang, Hong-Xing Yu, Rundi Wu et al.

This paper analyzes the content bias in Fréchet Video Distance (FVD), revealing its overemphasis on frame quality over motion continuity, rooted in feature extractor biases.

Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar et al.

Blink benchmark evaluates 14 core visual perception tasks using multimodal models like GPT-4V and Gemini, achieving only 51.26% accuracy, far below human performance of 95.70%.

Xingyu Fu, Yushi Hu, Bangzheng Li et al.

PROSE-PDE integrates multi-operator learning with symbolic encoding, enabling strong generalization and extrapolation for PDE solutions.

Jingmin Sun, Yuxuan Liu, Zecheng Zhang et al.

The paper introduces explicit 3D-feature viewpoint control for customized diffusion generation, but the supplied text reports no numerical metrics.

Nupur Kumari, Grace Su, Richard Zhang et al.

Proposes SOBER, a kernel quadrature-based batch Bayesian optimization framework with probabilistic lifting, supporting discrete, non-Euclidean spaces, and adaptive batch sizes.

Masaki Adachi, Satoshi Hayakawa, Martin Jørgensen et al.

Proposed a multisense consistency framework based on Fregean sense to evaluate GPT-3.5’s semantic stability across five languages, revealing significant inconsistencies.

Xenia Ohmer, Elia Bruni, Dieuwke Hupkes