FuncGenFoil is a new generative model that designs and edits high-fidelity airfoils as continuous functions, outperforming prior methods in accuracy and diversity while enabling flexible, resolution-free shape generation.

A method SceneDesigner and a new dataset ObjectPose9D enable accurate and flexible multi-object 9D pose manipulation.

This paper presents ZeroTIR, revealing agent‑level RL scaling laws that tie training steps, code‑call frequency, response length, and accuracy, and surpassing ZeroRL and SFT baselines on challenging math benchmarks.

We introduce UniGen, a unified multimodal large language model capable of image understanding and generation.

a novel approach to enable training high resolution sparse voxel diffusion model for 3D generation

We propose explicit and interpretable one-step generation framework that retains the advantages of traditional diffusion models, such as access to intermediate states and fine-grained control, while enabling fast sampling.

The paper introduces a novel two-stage sketching framework for federated LoRA that enables differentially private and effective updating of low-rank matrices in large language models.

We propose GRIFFIN to accelerate the inference speed of LLM by addressing the token misalignment issue in speculative decoding.

We develop a CLIP model that is SotA on both image and video zero-shot recognition. Using its strong, general features we further create SotA encoders for language and spatial tasks.

We propose a bilevel framework that jointly learns layer-wise sparsity and low-rank structure via truncated Gaussian sampling and efficient matrix approximation, achieving better performance for compressing LLMs.

A report of the Big ANN competition in NeurIPS'23

We propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand.

This paper introduces 4D-VLA, a pretraining framework that enhances spatial and temporal reasoning in robotics by aligning coordinate systems with RGB-D sequences and efficient memory sampling.

We propose GOOD, a training-free framework that leverages off-the-shelf classifiers to guide diffusion models for generating diverse, informative OOD samples—improving outlier exposure without requiring external datasets or embedding alignment.

We propose a hybrid physics-AI model incorporating urban surface energy processes and physical priors for improved local urban climate emulation.

We provide confidence intervals, prediction intervals, and hypothesis tests for variable importance for boosting with dropout and parallel training.

we propose a scalable pre-training method named POLicy DiscriminAtive LeaRning (POLAR), which trains a reward model (RM) to discern identical policies and discriminate different ones.

A sparse attention mechanism balances efficiency, long-range random access flexibility and length generalization ability

We present Fact-R1, a reasoning-enhanced framework for video misinformation detection, along with FakeVV, a large-scale benchmark of over 100,000 annotated video-text pairs.

FAPEX is a fractional neural frame operator-enhanced model designed for subject-agnostic seizure prediction.

We present 3EED, the first large-scale benchmark for 3D visual grounding across vehicles, drones, and quadrupeds, with over 134K 3D objects and 25K human-verified expressions in diverse outdoor scenes.

We introduce ORIGAMISPACE, a new origami dataset and benchmark to evaluate MLLMs in Multi-Step Spatial Reasoning with Mathematical Constraints.

We propose an information-preserving and data-free pipeline for ultra-efficient delta compression, achieving ultra-high compression ratios and strong performance across various models.

GoRA improves LoRA by adaptively assigning ranks and initializing weights using gradient info, achieving better performance while staying efficient. It outperforms LoRA and full fine-tuning on GLUE (+5.88) and GSM8k (+5.13) benchmarks.

The first deep-learning model for sub-daily ocean forecast

This paper introduces EventMG, a novel model built on multilevel Mamba-Graph architecture with specialized event scanning (SEST), delivering lightweight, efficient, and high-performance spatiotemporal representation for event camera data.

We design relative positional encoding methods for spiking Transformers

We propose Retrieval-Augmented Diagnosis (RAD), a holistic method that explicitly integrates disease-centered medical knowledge into multimodal diagnosis models.

We employ contrastive learning to extract complete point cloud structures from partial (incomplete) point clouds for guiding point cloud completion, achieving state-of-the-art (SOTA) results in the field of self-supervised point cloud completion.

We introduce EvaLearn, a pioneering benchmark designed to evaluate LLMs on their learning capability and efficiency in challenging tasks.

We introduced Multi-agent KTO, a method that trains LLM to play Werewolf through direct gameplay. Our approach outperforms GPT-4o and RL+LLM methods, achieving human-competitive performance.

We introduce a novel approach which leverages a convex potential for efficient sampling of EBMs.

Our proposed RH-Partial2Global adopts a jackknife conformal prediction-guided selection strategy to construct reliable alternative sets and a covering design-based sampling strategy to ensure holistic coverage of local pairwise preferences.

Reinforcement Learning Successfully Attempt for End-to-End Autonomous Driving in CARLA v2

Propose a novel approach that redefines pseudo-labeling and confidence estimation in the latent space.

This paper introduces an adaptive Riemannian hypergradient descent algorithm for solving Riemannian bilevel optimization problems.

We introduce a sample-efficient Diffusion-Inspired Multi-Agent world model (DIMA) for multi-agent control environments--MAMuJoCo and Bi-DexHands

We have decoupled the math atomic capabilities of large language models and explored their interaction relationships in mathematical reasoning tasks.

We present MICS, a multi-model collaborative search strategy that facilitates the generation of effective step-by-step CoT data.

we introduce the use of the memory buffer into the video stereo matching task, aiming to construct a memory buffer that is both compact and informative.

We present StreamBridge, a framework that transforms offline Video-LLMs into streaming models.

We address the problem of temporal text-attributed graph (TTAG) modeling and propose CROSS. CROSS is the first framework designed to unify text semantics and graph structures with LLMs for TTAG modeling.

JanusDNA, the first bidirectional DNA foundation model built upon a novel pretraining paradigm, integrating the optimization efficiency of autoregressive modeling with the bidirectional comprehension capability of masked modeling.

In this work, we introduce \textit{BackdoorLLM}, the first comprehensive benchmark for studying backdoor attacks and defenses on LLMs.

we propose a new local acceleration framework based on inexact proximal operators (e.g., Catalyst acceleration).

We propose Spatial Adversarial Alignment (SAA), which leverages a witness model to fine-tune the surrogate model via spatial-aware and adversarial-aware alignment to generate highly transferable adversarial examples.

We develop sublinear-time algorithms for computing and optimizing opinion-based quantities in the FJ model using partial rooted forest sampling.

We propose CADMorph, an inference-time editing method for parametric CAD models using the signal of geometry changes via a Plan–Generate–Verify Loop over pretrained priors, namely a LDM and an LLM, bypassing the need of non-existent editing data.

OrderMind is a spatial-aware framework for manipulation ordering in cluttered environments. It learns object manipulation orderings by encoding spatial relationships. It outperforms existing methods in both simulated and real-world tasks.

Spiral is a new type of LiDAR generation model that enables semantic awareness and progressive diffusion.

This paper the first unified multimodal CoT-based reward model, capable of multi-dimensional, step-by-step long-chain reasoning for both visual understanding and generation reward tasks.

This paper presents HumanoidGen, an automated task creation and demonstration collection framework that leverages atomic dexterous skills and LLM reasoning to generate relational constraints.

RetrievalAttention improves decoding speed and reduces GPU memory usage in Transformer-based LLMs by using pre-built, attention-aware KV vector indexes stored in CPU memory, achieving significant efficiency gains without compromising accuracy.

We introduce SPAR-7M and SPAR-Bench, a large-scale 2D-supervised dataset and benchmark for spatial reasoning, built using 3D ground-truth.