Using 68 probing tasks, the study finds BERT loses linguistic information after fine-tuning, while more readable representations improve NLI accuracy.

Alessio Miaschi, Dominique Brunato, Felice Dell'Orletta et al.

Proposed Graph Cross Network (GXN) with Vertex Infomax Pooling (VIPool) improves graph and vertex classification accuracy.

Maosen Li, Siheng Chen, Ya Zhang et al.

Theoretically proves over-parameterization accelerates gradient descent convergence by shrinking the distance to the global optimum faster, based on a single teacher neuron model.

Jun-Kun Wang, Jacob Abernethy

EECBS integrates explicit estimation search and online learning to accelerate bounded-suboptimal multi-agent pathfinding, outperforming state-of-the-art methods.

Jiaoyang Li, Wheeler Ruml, Sven Koenig

Proposes an integrated task and motion planning (TAMP) algorithm combining discrete task structures with continuous motion parameters, achieving 85% success in complex object manipulation tasks.

Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay et al.

SparTerm directly learns sparse text representations in full vocabulary space using importance prediction and gating, achieving state-of-the-art results on MSMARCO.

Yang Bai, Xiaoguang Li, Gang Wang et al.

kNN-MT enhances translation via nearest neighbor search, improving German-English BLEU by 1.5.

Urvashi Khandelwal, Angela Fan, Dan Jurafsky et al.

QAEval uses question-answer pairs generated from references and pre-trained QA models to evaluate summary content, outperforming ROUGE and BERTScore on benchmarks.

Daniel Deutsch, Tania Bedrax-Weiss, Dan Roth

FSD50K is an open, human-labeled sound event dataset with 51,197 clips across 200 classes, fostering large-scale SER research.

Eduardo Fonseca, Xavier Favory, Jordi Pons et al.

Proposed a grapheme-based labeling scheme, creating a dataset of 411k Bengali handwritten graphemes.

Samiul Alam, Tahsin Reasat, Asif Shahriyar Sushmit et al.

CrowS-Pairs dataset quantifies social biases in MLMs via 1508 sentence pairs across nine bias types, revealing widespread bias.

Nikita Nangia, Clara Vania, Rasika Bhalerao et al.

Performer introduces a linear-time Transformer using FAVOR+ to unbiasedly approximate softmax attention, enabling scalable long-sequence modeling.

Krzysztof Choromanski, Valerii Likhosherstov, David Dohan et al.

This paper introduces RAPS, a conformal prediction method with regularization, significantly reducing prediction set size while guaranteeing coverage.

Anastasios Angelopoulos, Stephen Bates, Jitendra Malik et al.

This paper critically examines flaws in current time series anomaly detection benchmarks and introduces the UCR Anomaly Archive for more reliable evaluation.

Renjie Wu, Eamonn J. Keogh

Proposes TAX4CS framework integrating SHAP, LIME, and PDP to enhance transparency and auditability of credit scoring models.

Michael Bücker, Gero Szepannek, Alicja Gosiewska et al.

Introduces MedQA dataset, combining retrieval and comprehension models, with test accuracy of only 36.7%-70.1%, highlighting challenges in medical open-domain QA.

Di Jin, Eileen Pan, Nassim Oufattole et al.

SPARTA employs sparse Transformer matching with inverted index for efficient open-domain QA, outperforming dense vector methods.

Tiancheng Zhao, Xiaopeng Lu, Kyusong Lee

robosuite v1.5, based on MuJoCo, offers a modular simulation platform with multi-robot models and multi-sensor integration, advancing robot learning research.

Yuke Zhu, Josiah Wong, Ajay Mandlekar et al.

Proposed ClariQ task with deep learning models for generating and ranking clarifying questions in open-domain dialogue.

Mohammad Aliannejadi, Julia Kiseleva, Aleksandr Chuklin et al.

This work adapts three large-scale controllable dialogue architectures to manage about 200 styles, improving style consistency and diversity.

Eric Michael Smith, Diana Gonzalez-Rico, Emily Dinan et al.