We prove that Private Evolution, a practical algorithm for differentially private synthetic data generation, converges in Wasserstein distance as the number of samples increases under a new, realistic theoretical model

We present Group-MATES, an efficient group-level data selection method that optimizes the speed-quality frontier for LLM pretraining. On DCLM, Group-MATES achieves up to 9.4% relative accuracy gain and 1.75× faster training than random selection.

We introduce a novel offline RL algorithm that leverages shortcut models to scale both training and inference.

A post-training quantization method for point cloud 3D detection

This paper presents a stylized model of the regulator's choice of an open-source definition in order to evaluate which standards will establish appropriate economic incentives for foundation model developers.

We introduce a principled framework for validating LLM-as-a-judge systems under rating indeterminacy, where multiple ratings can be "correct."

We propose Conformal Mixed-Integer Constraint Learning, a novel framework that provides probabilistic feasibility guarantees for data-driven constraints in optimization problems.

OMiSO (Online MicroStimulation Optimization): a brain stimulation framework to drive neural population activity toward specified states by adaptively choosing stimulation parameters based on brain state information

We introduce a new probabilistic method and task for quantifying the privacy risk of a document with personal attributes using large language models.

This paper is to jointly optimize mutiple modules in complex RAG pipeline using multi-agent reinforcement learning.

We consider the computational complexity of computing stationary points in min-max optimization, with a particular focus on the special case of computing Nash equilibria in (two-)team zero-sum games.

A Challenging Benchmark for Detecting Fake Images Generated by Advanced Neural Rendering Methods

A report of the Big ANN competition in NeurIPS'23

We propose FCGrad, a conflict-aware gradient adjustment method that improves social welfare while ensuring fairness in multi-agent mixed-motive settings by resolving conflicts between individual and collective objectives.

GC-xLSTM jointly optimizes a sparse feature selector and predictive xLSTM to uncover a graph of Granger Causal dependencies

Video Recognition models are not smooth as a function of time, but smoothing them improves accuracy

We identify dimensional collapse in VQVAEs, where the codebook's effective dimensionality is surprisingly low, and investigate its implications and potential remedies.

Feed-forward 3D reconstruction with flexibility to rig constraints

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.

We propose a sequential test for auditing differential privacy that detects violations with sample sizes that are orders of magnitude smaller than existing methods.

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 give improved approximation algorithms for fair monotone matroid submodular maximization.

A novel intrinsic motivation method based on world-model memory mismatch enables embodied agents to exhibit robust autonomous behaviors that closely match whole-brain neural data from zebrafish.

Propose a vehicle trajectory learning model capable of transferring across regions and tasks without retraining.

A new method for predicting PDE solutions is proposed based on Hamiltonian mechanics and functional approximation.

We present an algorithm to learn parameterized skills from demonstration data and show that our approach results in improved adaptation to unseen tasks.

AgMMU is a challenging real‑world benchmark for evaluating and advancing vision-language models (VLMs) in the knowledge‑intensive domain of agriculture.

We introduce a benchmark to measure safety of general computer use agents across diverse categories of harm

Machine learning researchers must urgently work with policymakers to address growing risks from embodied AI by plugging gaps in existing frameworks.

Theoretical inconsistencies among Responsible AI metrics aren't a problem but a benefit – they enable both pluralistic approaches to alignment that respect diverse values, help conceptual understanding, and create more robust, adaptable models.

AI testing will become less reliable as AIs develop better understanding of when they're being tested; we propose using game theory to help mitigate this.

a small audio-language model for audio reasoning that achieves SoTA performance with 50 times fewer parameters and 60 times fewer audio hours.

We propose thought communication for agents to interact directly mind-to-mind, prove its nonparametric identifiability, and develop a framework for communication beyond language.

We use a novel projection-based Lyapunov function to prove the first asymptotic optimality theorem for fully heterogeneous weakly-coupled Markov Decision Processes.

We propose a conformalized framework for constructing prediction sets in stream networks, a specialized class of spatiotemporal graphs.

We introduce the Martingale Score, an unsupervised metric from Bayesian statistics, to show that reasoning in LLMs often leads to belief entrenchment rather than truth-seeking, and shows this score predicts ground-truth accuracy.

We introduce GLM-Transformer, a hybrid deep learning and GLM framework that identifies interpretable cross-area neural interactions while accounting for trial-varying individual-neuron dynamics via a Transformer-based VAE..

Learning the individualized treatment effects under the assumption of rank preservation.

We propose a novel hybrid of position embedding to improve the length generalization ability of Vision-Language Models.

PartCrafter is a structured 3D generative model that jointly generates multiple parts and objects from a single RGB image in one shot.

Rigor in AI remains largely understood in terms of methodological rigor. This common, yet narrow conceptualization of rigor has contributed to many responsible AI concerns. Here, we argue for a broader view of what rigorous AI research should entail.

We present a fairshare training data pricing framework for large language models where data pricing depends on the data valuation (which estimates data contribution towards model learning), and benefits both sellers and buyers.

Scaling Transformers for brain-computer interfaces show gains restricted to low data regimes.

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

We provide a principled approach for incorporating fully synthetic samples in downstream statistical inference.

Antidistillation sampling strategically modifyies a model's next-token probability distribution to poison reasoning traces, rendering them significantly less effective for distillation while preserving the model's practical utility.

We show that using checklists to automatically grade responses for reinforcement learning leads to improved instruction following

Two closed-form empirical Bernstein inequalities that are as sharp as the non-empirical ("oracle") Bernstein inequality for matrices.

We design a novel algorithm to stabilize an unknown partially observable LTI system with a complexity bound independent of the state dimension.

We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

We train LLMs to reason efficiently using RL

A novel approach to perform distributionally robust feature selection in a model-agnostic manner

We present a data-centric pretraining framework that builds safety into the model from the start

We address the challenge of predicting where a person's visual perception will focus next within their three-dimensional environment.

An Early Assessment for Multimodal Output of Omni Foundation Models

We scale the influence-function-based data valuation method to recent LLMs and their massive training datasets.

Encouraging model in model-based reinforcement learning to converge to flatter minima in the loss landscape will result in better downstream policies

A model-free speculative decoding method that accelerates agentic AI workloads using suffix trees. Achieves 5.3x speedup on multi-agent tasks.

Task-optimized convolutional recurrent neural networks trained on realistic tactile inputs align with rodent somatosensory data, suggesting brain tactile processing uses temporally-precise representations shaped by categorization-driven optimization.

We propose Refer-Judge, a Jury-and-Judge Chain-of-Thought framework that leverages MLLMs to uncover toxic data in 3D visual grounding

We show that transformers achieve length generalization when training on shorter main task and longer auxiliary tasks together.

We show that diffusion language models are a lot more sample-efficient than standard autoregressive language models, due to their ability to learn from different token orderings.

We reveal that Edge of Stability dynamics enhance the alignment between the Neural Tangent Kernel and the target.

We show that even exact unlearning, the gold standard for data removal in large language models, can leak sensitive information by using guidance between pre- and post-unlearning checkpoints.

We propose an online reinforcement learning technique to fine-tune a family of flow matching policies for robot learning.

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

We introduce Roboflow100-VL, a large-scale dataset with diverse concepts not typically found in internet-scale pre-training.

We introduce ParallelPrompt, the first benchmark and dataset for studying intra-query parallelism in real-world LLM queries, enabling reproducible evaluation of structured execution strategies.

Researchers and reviewers in AI for Social Impact should adopt broader but more rigorous standards for impact

We analyze what kind of LLMs have large improvement in RL finetuning and propose behavior injection augmentation to prepare the LLMs for RL.

This paper introduces a new simple but efficient learning mechanism for improving the robust alignment between visual and textual modalities by solving shuffling problems.

We propose a parameter-efficient fine-tuning method that improves the performance of LoRA in the presence of computational and data heterogeneity in FL.

We develop an adaptive image tokenizer that compresses images into variable-sized latent features based on its content complexity.

We propose an assume-guarantee compositional framework, CoVeNN, to scale verification to large networks. CoVeNN can verify nearly 7 times more problems than state-of-the-art verifiers.

Long-term feed-forward 3D point tracking in persistent 3D point maps.

We reliably predict the behavior of black-box language models by training predcitors on their responses to follow-up questions.

Learning likelihoods instead of calculating expensive jacobians on generative ODEs to calculate free energy differences and obtain the Boltzmann distribution.

AirRep is a text representation model optimized for TDA, offering performance comparable to gradient-based methods while being significantly more efficient.

Open-source framework for LLM unlearning supporting multiple benchmarks and methods.