We introduce the Birkhoff Extension, a polynomial-time, almost-everywhere differentiable relaxation of permutation objectives to doubly stochastic matrices, enabling efficient combinatorial optimization.

We introduce a new SGD-based algorithm with delayed projection for training kernel machines that achieves comparable or superior performance while reducing training time from days to under an hour.

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.

This paper presents a lightweight MCMC-based approach for constrained decoding in LLMs

We adopt reinforcement learning to train LLMs to generate quality code with rewards derived from program analysis.

This paper demonstrates that language models possess metacognitive-like abilities to monitor and control their internal neural activations.

Staggered resets fix harmful nonstationarity in massively parallel RL's short synchronous rollouts by varying environment start times, improving state coverage and boosting sample efficiency, performance, and scalability.

a new learning scheme for multi-modal LLM, LLAVA

We construct generalized Lyapunov certificates by augmenting value functions with residual terms to certify the stability of control policies from optimal control or reinforcement learning.

conformal prediction for time series using switching dynamics systems

An algorithm with theoretically guaranteed convergence rate to optimize a preference function over the Pareto set of a given set of objectives.

Zero Shot Metric depth from a pre-trained monocular depth diffusion model by incorporating optical defocus cues at Inference time. Outperforms some of the existing Zero Shot Metric depth baselines, while also

We enable vision-language models to efficiently encode and utilize multimodal multilingual knowledge through a lightweight, continuous memory mechanism.

We rethink the roles of descriptions and narrations in role-playing and build a system to initialize the characters' behavior modes as executable codes with descriptions and refine with narrations.

A principled and scalable method for model-agnostic data valuation

we theoretically study when and how unlabeled data can help in multi-objective learning

Transformer, Mamba, and RWKV language models show consistent patterns of change in behavior over the course of training

A Metropolis–Hastings driven adaptive sampling framework for 3D Gaussian Splatting that unifies densification and pruning to achieve more efficient, high-quality real-time scene reconstructions with fewer Gaussians.

We identify several factors that lead to token premium effects in monolingual tokenizers and provide two interventions which significantly reduce tokenizer inequities.

We contribute provable guarantees that regularized policy gradient methods converge in approximate Nash equilibria in imperfect-information extensive-form zero-sum games.

We propose grafting, a simple approach to materialize new architectures by editing pretrained diffusion transformers. It enables architectural exploration under small compute budgets.

We conducted a systematic review of CS papers that collect/use nude images and found widespread nonconsensual collection, as well as concerning practices like intentional collection of abuse content and the nonconsensual distribution of nude images.

We study replicable distribution testing.

We present a general algorithm for learning real-valued Multi-Index Models and matching Statistical Query lower bounds.

Trainable sparse attention for video diffusion model.

A sequence-level EDM-derived diffusion model using progressively increasing noise levels is designed for probabilistic forecasting of complex dynamics like weather.

Low-precision quantized variants of Oja’s algorithm can provably converge in streaming PCA under suitably fine discretization while drastically reducing memory and compute requirements.

We prove that black-box variational inference with the mean-field Gaussian variational family converges in a rate with an explicit dimension dependence of only $O(\log d)$.

NFL-BA enhances tracking and mapping in neural rendering-based SLAM by explicitly modeling dynamic near-field lighting, with applications in endoscopy.

TIDMAD is the first dataset and benchmark from a dark matter physics experiment, providing ultra-long time series data and comprehensive benchmarking tools that enable models to directly advance the fundamental physics search for dark matter.

We introduce TalkCuts, a large-scale dataset for multi-shot speech video generation, and demonstrate its utility through a simple LLM-guided generation baseline.

System identification from videos during collisions with arbitrarily shaped rigid bodies.

Statistically and computationally efficient reduction from approximate DP to pure DP.

Inference-time approach to mitigate verbatim generation in large language models

We present scPilot, a framework for omics-native reasoning in which a large language model converses, invokes bioinformatics tools, and iteratively explains its decisions to for single-cell analysis.

demonstrates reasoning settings where scaling the length of chain-of-thought offers an exponential advantage over parallel scaling and aggregation.

We present an end-to-end framework for self supervised combinatorial optimization under various constraints (cardinality, matroid, independent set, etc.).