Proposes Meta Neural Trojan Detection (MNTD) using jumbo learning, achieving 97% AUC in black-box Trojan detection across diverse datasets.

Xiaojun Xu, Qi Wang, Huichen Li et al.

Introduced Talk2Car dataset with 11,959 natural language commands for urban scene target recognition; evaluated state-of-the-art models achieving up to 50.51% IoU.

Thierry Deruyttere, Simon Vandenhende, Dusan Grujicic et al.

Introduced Jericho environment to study language agents in interactive fiction games, achieving a 10.7% score improvement.

Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Côté et al.

DeepProbLog integrates neural networks with probabilistic logic programming, enabling symbolic and subsymbolic inference, program induction, and end-to-end training, advancing neuro-symbolic AI.

Robin Manhaeve, Sebastijan Dumančić, Angelika Kimmig et al.

Proposes Attainable Utility Preservation (AUP) to reduce reward misspecification risks by preserving multi-objective optimization capabilities.

Alexander Matt Turner, Dylan Hadfield-Menell, Prasad Tadepalli

Proposes a compositional formal verification framework for autonomous systems with DNNs, integrating clustering-based safety region analysis and assume-guarantee reasoning.

Corina S. Pasareanu, Divya Gopinath, Huafeng Yu

BabyAI benchmarks grounded compositional learning with MiniGrid, behavioral cloning, and PPO; IL needs 8.4k–409k demonstrations.

Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem Lahlou et al.

CURIOUS combines modular UVFA with intrinsic learning progress to enable self-organized multi-goal reinforcement learning, demonstrating robustness and developmental self-organization.

Cédric Colas, Pierre Fournier, Olivier Sigaud et al.

The paper introduces SPL metric to standardize evaluation of navigation tasks, enhancing research on agents in 3D environments.

Peter Anderson, Angel Chang, Devendra Singh Chaplot et al.

Proposed Low-rank Multimodal Fusion method significantly reduces computational complexity in tasks like sentiment analysis.

Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan et al.

TD3 algorithm employs twin critics with minimum value selection, delayed policy updates, and target smoothing, reducing overestimation bias in continuous control tasks, outperforming DDPG.

Scott Fujimoto, Herke van Hoof, David Meger

Proposes Workflow-Guided Exploration (WGE) with DOMNET for web tasks, achieving 100x sample efficiency and high success rates with minimal demonstrations.

Evan Zheran Liu, Kelvin Guu, Panupong Pasupat et al.

PDDLStream integrates symbolic planners with black-box samplers via adaptive optimistic planning, significantly improving efficiency in high-dimensional robotic tasks.

Caelan Reed Garrett, Tomás Lozano-Pérez, Leslie Pack Kaelbling

SRL survey: AE, E2C, ICM and priors learn compact control states that speed RL.

Timothée Lesort, Natalia Díaz-Rodríguez, Jean-François Goudou et al.

DeepMind Control Suite offers standardized continuous control benchmarks with MuJoCo, supporting multiple algorithms and task types for RL research.

Yuval Tassa, Yotam Doron, Alistair Muldal et al.

AlphaZero uses neural networks and Monte Carlo tree search to surpass human champions in chess, shogi, and Go within 24 hours of self-play training.

David Silver, Thomas Hubert, Julian Schrittwieser et al.

Proposes a compression-based macro discovery framework that extracts, evaluates, and diversifies action sequences to accelerate RL learning in related tasks.

Francisco M. Garcia, Bruno C. da Silva, Philip S. Thomas

Multi-agent self-play training induces emergent complex behaviors surpassing environment complexity.

Trapit Bansal, Jakub Pachocki, Szymon Sidor et al.

Proposes mixed precision training using FP16 storage, FP32 master weights, and loss scaling, achieving accuracy parity with FP32 while halving memory usage.

Paulius Micikevicius, Sharan Narang, Jonah Alben et al.

End-to-end negotiation model with dialogue rollouts significantly improves agreement success and strategic behavior.

Mike Lewis, Denis Yarats, Yann N. Dauphin et al.