We propose a constrained optimization framework for LLM unlearning that maximizes entropy to forget while preserving utility, achieving stable and effective results across benchmarks.

We theoretically and empirically analyze the effect of the Pooling Operation in the downstream performance.

We introduce a universal graph structure augmentor for cross-domain data scaling.

STSBench is a benchmark that evaluates the capabilities of Multi-modal Large Language Models to reason about spatio-temporal actions.

We introduce MIRAGE, a benchmark for multimodal expert consultation in agriculture featuring single-turn and multi-turn tasks.

We study the problem of computing an optimal large language model (LLM) policy for a constrained alignment problem.

We propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge.

This paper introduces the first unified platform for evaluating text-to-video and video-to-text models across five key dimensions: safety, hallucination, fairness, privacy, and adversarial robustness.

We introduce CLEVER, a hand-curated benchmark for verified code generation in Lean. It requires full formal specs and proofs. No few-shot method solves all stages, making it a strong testbed for synthesis and formal reasoning.

We propose a compute- and memory-efficient inference framework for handling extremely long inputs with pre-trained Transformers.

Enabling NeRFs to "see" in wireless multipath settings.

The paper introduces a novel latent diffusion model that uses local vicinity structures to achieve state-of-the-art domain adaptation, preserving privacy without needing source data during adaptation.

We characterize key factors for good synthetic priors and build a SOTA tabular foundation model by pretraining on these priors.

We introduce AgentTTS, an LLM-agent framework that efficiently allocates test-time compute across multi-stage tasks, achieving compute-optimal performance through insight-guided adaptive search.

We propose SAGE, a self-guided hierarchical exploration framework that trains generalist foundation model web agents without human supervision, achieving state-of-the-art performance

We propose GLVD, a fast and accurate hybrid method for 3D face reconstruction that combines per-vertex neural fields with dynamically predicted 3D landmarks to achieve state-of-the-art quality from just one or a few images.

We show that explicit reasoning via chain-of-thought can hurt instruction-following in LLMs by reducing constraint adherence, and propose four mitigation methods to recover or improve performance.

We propose Think-RM, a training framework for generative reward models that enables long-horizon reasoning, and introduce a pairwise RLHF pipeline that directly optimizes policies using pairwise preference rewards.

Neuron Chunking is a hardware-aware sparsification framework that abstracts access patterns into contiguity distributions to couple neuron selection with flash I/O behavior and improve I/O efficiency in VLM inference.