We propose BOOM, a model-based RL that uses a soft value-weighted likelihood-free alignment loss to bootstrap the policy from non-parametric planner with world model, achieving state-of-the-art performance.

Debugging today's chain-of-thought reasoning for video understanding and introducing a novel reward for MLLM post-training to ground video reasoning in visual evidence.

Camera conditioning as relative projective transformation in multi-view transformer

We introduce a method that allows adversarial optimization to be used in general-sum settings to train more robust and diverse policies.

A study on equivariant and non-equivariant convolution networks for point clouds

Scalable, simple, and practical algorithm for model-based RL with regret bounds across several RL settings and experiments on state-based, visual control and hardware tasks.

We find a small set of neurons whose activations can be redirected at test-time to mitigate high-norm artifacts in Vision Transformers.

We build a LM-based system that can outperfrom expert AI researchers in predicting the outcomes of empirical AI research ideas, without running actual experiments.

We introduce a diffusion-based video model that predicts egocentric futures from full-body 3D motion, enabling realistic and controllable first-person simulation.

We study the mechanism of chain of continuous thought on the graph reachability problem, and show it can reason by maintaining a superposition of multiple search traces both theoretically and empirically.

We provide a computational efficient algorithm to achieve $O(H)$ deployment cost with polynomial sample complexity.

We derive brain-like inference as natural gradient descent on free energy (FOND). The resulting spiking network (iP-VAE) outperforms amortized VAEs in reconstruction-sparsity trade-offs and out-of-distribution generalization.

Deep learning image filters lack normalization. We propose filter normalization with learnable scaling and shifting, achieving significant improvements in intensity variation benchmarks.

We propose to scale the number of interaction steps for agents as a new axis of test-time scaling and develop a curriculum-based online RL algorithm for training agents to scale interaction.

Our framework explains how users with inconsistent preferences can align an engagement-driven algorithm with their interests in equilibrium, highlighting both the challenges and potential remedies for achieving alignment.

We present a unified theoretical framework for standard and high-order flow matching and prove their minimax optimality.

We propose SimpleStrat for diversifying LLM generations and introduce CoverageQA a benchmark of multi-answer questions for evaluating coverage.

We establish the universal approximation capability of single-layer, single-head self- and cross-attention mechanisms

Introducing models that prove their own correctness via an Interactive Proof, and how to learn such models.

GradMetaNet is a neural architecture that efficiently processes gradients of other networks by exploiting their symmetries and rank-1 decomposition structure, enabling better learned optimizers, model editing, and loss curvature estimation

We present the first systematic study of lossy latency–quality trade-offs in LLM agents, introducing HFTBench and StreetFighter benchmarks, and proposing an adaptive mixed-precision framework for real-world latency-sensitive tasks.

We show that vision-language models can piece together scattered training image patches—a capability we call visual stitching—which enables generalization to unseen images but also introduces new safety risks.

We propose KRIS-Bench, a benchmark designed to evaluate knowledge-based reasoning in instruction-based image editing models.

We introduce IneqMath, an informal inequality proving benchmark, and an LLM-as-judge suite, revealing that top LLMs achieve <10% overall accuracy due to flawed step-wise reasoning.

GSO: SWE Agents Struggle at Reasoning and Engineering for Software Optimization

A training-free, generative approach that infers object removal order by exploiting statistical co-occurrence and asymmetry priors learned by generative models.

We discover that coherent value systems emerge with scale in LLMs and propose the research avenue of utility engineering to analyze and control these emergent value systems.

To achieve personalization in LLMs, we leverage the user model to incorporate a curiosity-based intrinsic reward into multi-turn RLHF.

We introduce Alternating Gradient Flows, a framework modeling feature learning in two-layer networks with small initialization as utility maximization and cost minimization—unifying saddle-to-saddle analyses and explaining Fourier feature emergence

We build a nearly-perfect predictive model for memorization in diffusion models using theory and controlled experiments.

Novel method for 3D mesh interior painting

We propose Q-chunking, a simple, effective offline-to-online RL method that uses action chunking to improve value propagation and exploration via temporally coherent actions.

We introduce CQN-AS, a value-based RL algorithm that learns a critic network that outputs Q-values over a sequence of actions.

We propose a simple robot learning recipe leveraging privileged information to train active perception policies on real robots.

We identify and explore a prominent layer-level correction mechanism found in a variety of open-weight LLMs

We use random matrix theory to estimate the spectral density of matrices too large to fit into memory.

A systematic review of benchmarks for large language models focusing on issues of construct validity, which results in a set of recommendations for best practices.

A dataset of multi-agent system traces, and a systematic analysis of failures in multi-agent LLM systems, featuring a structured taxonomy and an automated evaluation pipeline.

The AI data value chain strips economic value from its generators, benefiting only aggregators. We identify missing provenance, power imbalances, and non-dynamic pricing as causes and envision a framework for a fairer data exchange that benefits all.

We propose score-based diffusion framework with explicit G-modeling, a nonparametric empirical Bayes method that achieves near-parametric score estimation guarantees and state-of-the-art denoising for multivariate heteroscedastic Gaussian mixtures.

SHAP zero estimates Shapley values and interactions with a near-zero marginal cost for future queried sequences through the Fourier transform.

a sparse attention with $\mathcal O(n \log n)$ complexity for long video generation

We develop an architecture, EddyFormer, to accelerate the resolution of three-dimensional turbulent fluid flows.

We solve the learning dynamics of (a close approximation of) word2vec in closed form, revealing what semantic features are learned.

Our paper shows that ensemble learning via majority voting achieves exponentially faster risk decay, improving base learners with slow rates.

We tackle prediction of partially observable systems with a multiscale inference scheme for diffusion models, improving accuracy and stability on solar dynamics without increasing computational cost.

We introduce Woodbury's matrix identity, momentum-like SPRING and randomization to make energy natural gradient descent 75 times faster for PINNs.

We introduce REVERSE, the first framework to integrate generation adjustment with online post-hoc verification within a single VLM architecture. REVERSE detects, backtracks, and corrects hallucinations during the decoding process.

We create a unified benchmark for evaluating secure code generation, vulnerability detection and poc generation

We argue that there are important gaps between what unlearning methods can do and claims about/hopes for what these methods can do for broader impact.

An adaptive flow-based generative model for RGB-Thermal Image Translation, which can adapt to different sensors, viewpoints, and environments.

This paper introduces a novel Multimodal Attention-based Normalizing Flow approach to developing explicit, interpretable, and tractable multimodal fusion learning

We analyze optimal budget allocation strategies for iterative bootstrapping with synthetic data to maximize model performance.

Trainable sparse attention for video diffusion model.

Long-RL enables RL on hour-long videos on a single A100; LongVILA-R1-7B supports 8,192 frames and scores 65.1%/71.1% on VideoMME.

We develop private mechanisms for Euclidean Jordan algebras and private algorithms for symmetric cone programming, this leads to private algorithms for semidefinite programming.

We show how to learn representations of temporal distances that exploit quasimetric architectures in offline GCRL.

We distill a slow, unlearning-based data attribution method to a feature embedding space for efficient retrieval of highly influential training images.

We propose a method to speedup video diffusion generation through efficient attention.

We propose a method which exploit KV cache sparsity efficiently and dynamically through Top-P sampling.

We introduce a theoretical model for the co-occurrence statistics of words that quantitatively explains the emergence of linear analogies of word embedding in models like Word2Vec and GloVe.

We introduce a framework for evaluating & improving LLM consistency in simulated human dialogue. Our metrics correlate with human judgments and when used with multi-turn RL, reduce inconsistency across chit-chat, teaching and mental health dialogue.

We introduce Backward Conformal Prediction, a new method that adapts coverage levels to enforce interpretable, data-dependent prediction set sizes, with provable guarantees and practical estimators.