Proposes multivariate uncertainty modeling in deep learning using Gaussian loss and Kalman filtering, improving autonomous navigation robustness.

Rebecca L. Russell, Christopher Reale

acados is a modular, high-performance embedded optimal control solver based on BLASFEO, supporting CasADi, Matlab, and Python, enabling real-time applications.

Robin Verschueren, Gianluca Frison, Dimitris Kouzoupis et al.

BART combines bidirectional encoder and autoregressive decoder with diverse noising pretraining, boosting NLP tasks including generation and comprehension.

Mike Lewis, Yinhan Liu, Naman Goyal et al.

Proposes Graph Diffusion Convolution (GDC), combining spatial and spectral methods, significantly improving GNN performance.

Johannes Gasteiger, Stefan Weißenberger, Stephan Günnemann

Proposed a BERT-based weakly-supervised method, significantly improving factual consistency detection accuracy in summaries.

Wojciech Kryściński, Bryan McCann, Caiming Xiong et al.

MOVE employs musically-motivated deep embeddings with triplet loss and hard mining, achieving state-of-the-art accuracy and scalability in music version identification.

Furkan Yesiler, Joan Serrà, Emilia Gómez

Relay Policy Learning combines imitation and reinforcement learning with goal-conditioned hierarchical policies to solve long-horizon robotic tasks, outperforming traditional HRL.

Abhishek Gupta, Vikash Kumar, Corey Lynch et al.

HRL4IN employs hierarchical reinforcement learning with multi-space subgoals and embodiment selection, boosting interactive navigation efficiency and energy savings.

Chengshu Li, Fei Xia, Roberto Martin-Martin et al.

RoboNet dataset combined with visual foresight and inverse models enables cross-robot generalization, surpassing single-robot training with 4-20x less data.

Sudeep Dasari, Frederik Ebert, Stephen Tian et al.

DeepCT leverages BERT to generate context-aware term weights, improving first-stage retrieval accuracy by 27-46% on MS MARCO and TREC-CAR.

Zhuyun Dai, Jamie Callan

Unified T5 model transforms all NLP tasks into text-to-text format, achieving SOTA on benchmarks with up to 11B parameters.

Colin Raffel, Noam Shazeer, Adam Roberts et al.

Introduces WHAMR! dataset combining noise and reverberation; proposes deep learning models achieving up to 12.9dB SI-SDR gains in realistic conditions.

Matthew Maciejewski, Gordon Wichern, Emmett McQuinn et al.

Introduces multi-pitch features and prototypical triplet loss to improve cover detection accuracy, achieving significant gains on large datasets and live scenarios.

Guillaume Doras, Geoffroy Peeters

Proposed federated neuromorphic SNN training method reduces communication load while maintaining high accuracy, validated on MNIST-DVS with 92.3% accuracy.

Nicolas Skatchkovsky, Hyeryung Jang, Osvaldo Simeone

mathlib leverages dependent types and automation to build a unified formal mathematics library, supporting classical structures.

The mathlib Community

This work decouples representation learning and classifier training, using simple sampling and classifier calibration to outperform complex methods on long-tail datasets.

Bingyi Kang, Saining Xie, Marcus Rohrbach et al.

Bayesian Symbolic Regression enhances expression simplicity, results closer to ground truth.

Ying Jin, Weilin Fu, Jian Kang et al.

MLQA creates a large-scale, multi-language aligned extractive QA benchmark across 7 languages, enabling comprehensive cross-lingual transfer evaluation.

Patrick Lewis, Barlas Oğuz, Ruty Rinott et al.

Introduces spike-based speech datasets using physiology-inspired audio-to-spike conversion, demonstrating the importance of spike timing for classification accuracy.

Benjamin Cramer, Yannik Stradmann, Johannes Schemmel et al.

Proposed ADR algorithm with memory-augmented RL enables sim2real transfer for robot solving Rubik’s cube, achieving 85% success.

OpenAI, Ilge Akkaya, Marcin Andrychowicz et al.