Proposes Skill-It framework using skill dependency graphs to optimize data sampling, improving language model training efficiency.

Mayee F. Chen, Nicholas Roberts, Kush Bhatia et al.

Proposes a goal-conditioned deep RL framework using tactile sensing for object pushing, trained in simulation with sim-to-real transfer, achieving over 90% success on unseen objects.

Max Yang, Yijiong Lin, Alex Church et al.

Proposes C-3PO, a Yule-Simon-based probabilistic position bias model for short-video feeds, improving exposure estimation accuracy.

Olivier Jeunen

QuIP introduces incoherence-based post-training 2-bit quantization for LLMs, with theoretical guarantees and state-of-the-art empirical results.

Jerry Chee, Yaohui Cai, Volodymyr Kuleshov et al.

Proposes LLM-Rec, using four prompting strategies to enrich item descriptions, significantly improving recommendation accuracy.

Hanjia Lyu, Song Jiang, Hanqing Zeng et al.

Proposes RippleEdits benchmark to evaluate ripple effects in knowledge editing, revealing current methods' inability to maintain consistency.

Roi Cohen, Eden Biran, Ori Yoran et al.

WebAgent combines HTML-T5 planning with Flan-U-PaLM code synthesis, raising real-world web-task success by over 50%.

Izzeddin Gur, Hiroki Furuta, Austin Huang et al.

CLIP-KD employs feature alignment and contrastive learning to improve small models, maximizing similarity with teacher features, achieving significant performance gains.

Chuanguang Yang, Zhulin An, Libo Huang et al.

Combines OpenPose and CNN to detect suspicious exam behaviors by analyzing arm angles and duration, achieving 82% accuracy in identifying object exchanges.

Reuben Moyo, Stanley Ndebvu, Michael Zimba et al.

Jina Embeddings excel in MTEB using contrastive training on T5 architecture.

Michael Günther, Louis Milliken, Jonathan Geuter et al.

FLASK introduces a fine-grained, skill-based evaluation framework for LLMs, improving interpretability and reliability.

Seonghyeon Ye, Doyoung Kim, Sungdong Kim et al.

Kernel-based nonlinear meta-learning guarantees faster generalization rates in infinite-dimensional RKHS.

Dimitri Meunier, Zhu Li, Arthur Gretton et al.

This study constructs a comprehensive taxonomy of NLP research, analyzes over 80,000 papers, and identifies key trends, highlighting language models and responsible AI as future priorities.

Tim Schopf, Karim Arabi, Florian Matthes

DNA-Rendering creates a large-scale, high-fidelity human dataset with 1500+ subjects, 67.5M frames, boosting neural human rendering performance.

Wei Cheng, Ruixiang Chen, Wanqi Yin et al.

The Android in the Wild (AITW) dataset offers 715k device interaction examples across multiple Android versions and device types.

Christopher Rawles, Alice Li, Daniel Rodriguez et al.

Hierarchical imitation learning enables robust multi-stage cable routing, improving success rate by 30% over end-to-end methods.

Jianlan Luo, Charles Xu, Xinyang Geng et al.

FlashAttention-2 achieves 2× speedup over FlashAttention by optimizing work partitioning, reaching 50-73% of GPU FLOPs utilization during training GPT-style models.

Tri Dao

nPOLO framework employs multiple scattering to realize low-power optical nonlinearities, enabling multi-layer neural network functionalities solely with linear optics.

Mustafa Yildirim, Niyazi Ulas Dinc, Ilker Oguz et al.

Formulates streaming CTR prediction, analyzes model performance under distribution shifts, improves via parameter tuning and exemplar replay.

Qi-Wei Wang, Hongyu Lu, Yu Chen et al.

InternVid leverages large-scale video-text data and multi-scale captioning to train ViCLIP, achieving state-of-the-art zero-shot action recognition.

Yi Wang, Yinan He, Yizhuo Li et al.