Introduces BeaverTails dataset with disentangled helpfulness and harmlessness annotations, using PPO-Lagrangian for safety RLHF, improving AI safety alignment with 33k QA and 36k preference pairs.

Jiaming Ji, Mickel Liu, Juntao Dai et al.

RLTF employs online reinforcement learning with multi-granularity unit test feedback to enhance code generation accuracy.

Jiate Liu, Yiqin Zhu, Kaiwen Xiao et al.

SVIT introduces a 4.2M high-quality visual instruction dataset, significantly outperforming state-of-the-art multimodal models.

Bo Zhao, Boya Wu, Muyang He et al.

Polybot aligns observation, action, and internal features across robots via wrist cameras, inverse kinematics, contrastive learning, and multi-head policies, enabling zero-shot transfer and improved success rates.

Jonathan Yang, Dorsa Sadigh, Chelsea Finn

Analyzed generalization gap in visual robotic manipulation using Factor World benchmark.

Annie Xie, Lisa Lee, Ted Xiao et al.

Proposes DE-TGN, a deep ensemble framework with Monte Carlo Dropout for human motion prediction and uncertainty quantification.

Kareem A. Eltouny, Wansong Liu, Sibo Tian et al.

Proposes a semiring-weighted Earley parser with complexity reduced to O(N³|G|), enabling efficient large-scale grammar parsing.

Andreas Opedal, Ran Zmigrod, Tim Vieira et al.

LongNet introduces dilated attention, enabling linear complexity for over 1 billion tokens, supporting scalable long-sequence modeling.

Jiayu Ding, Shuming Ma, Li Dong et al.

Proposes <projektor> framework combining optimal transport and neural scaling laws for performance prediction and data selection, improving accuracy and efficiency.

Feiyang Kang, Hoang Anh Just, Anit Kumar Sahu et al.

Spike-driven Transformer combines SNN and Transformer, achieving 77.1% accuracy on ImageNet-1K.

Man Yao, Jiakui Hu, Zhaokun Zhou et al.

Engression neural network distributional regression enables reliable extrapolation by modeling full conditional distributions, outperforming traditional methods in high-dimensional tasks.

Xinwei Shen, Nicolai Meinshausen

RH20T dataset enables one-shot learning for diverse robotic skills with over 110,000 sequences.

Hao-Shu Fang, Hongjie Fang, Zhenyu Tang et al.

MedCPT employs contrastive pre-training on 2.55B PubMed search logs, achieving state-of-the-art zero-shot biomedical IR performance surpassing larger models.

Qiao Jin, Won Kim, Qingyu Chen et al.

GGS method uses graph smoothing and MCMC to improve protein fitness by 2.5×, leveraging graph regularization and gradient-guided sampling.

Andrew Kirjner, Jason Yim, Raman Samusevich et al.

This paper evaluates two belief probing methods in LLMs, finds poor generalization, and discusses philosophical and empirical issues surrounding model beliefs.

B. A. Levinstein, Daniel A. Herrmann

Hierarchical neural code trees enable controllable CAD generation with high complexity, outperforming SOTA in diversity and realism metrics.

Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne et al.

Proposes Twisted Diffusion Sampler (TDS), combining SMC and twisting to achieve asymptotic exactness in conditional diffusion sampling, enhancing protein design and image generation.

Luhuan Wu, Brian L. Trippe, Christian A. Naesseth et al.

Introduces delay learning in deep SNNs via dilated convolutions with learnable spacings, achieving state-of-the-art accuracy on temporal benchmarks.

Ilyass Hammouamri, Ismail Khalfaoui-Hassani, Timothée Masquelier

Proposes Michelangelo framework utilizing shape-image-text aligned latent space with diffusion models, achieving superior quality and diversity in cross-modal 3D shape generation.

Zibo Zhao, Wen Liu, Xin Chen et al.

Sparse Model Soups (SMS) improves pruning by averaging models, enhancing generalization and OOD performance while maintaining sparsity.

Max Zimmer, Christoph Spiegel, Sebastian Pokutta