Autoformalization, the process of automatically translating natural language mathematics into machine-verifiable formal language, has demonstrated advancements with the progress of large language models (LLMs). However, a key obstacle to further advancements is the scarcity of paired datasets that align natural language with formal language. To address this challenge, we introduce ATLAS (Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data), an iterative data generation framework designed to produce large-scale, high-quality parallel theorem statements. With the proposed ATLAS running for 10 iterations, we construct an undergraduate-level dataset comprising 300k theorem statements and develop the ATLAS translator, achieving accuracies of 80.59% (pass@8) and 92.99% (pass@128) on ProofNet, significantly outperforming the base model (23.99% and 47.17%) and InternLM2-Math-Plus-7B (50.94% and 80.32%). Furthermore, the ATLAS translator also achieves state-of-the-art performance on both the high-school-level miniF2F dataset and the graduate-level MathQual dataset introduced in this work. The datasets, model, and code will be released to the public soon.

We then introduce a novel hierarchical knowledge distillation strategy that incorporates the similarity matrix, feature representation, and response map-based distillation to guide the learning of the student Transformer network. We also enhance the model's ability to capture temporal dependencies by applying the temporal Fourier transform to establish temporal relationships between video frames. We adapt the network model to specific target objects during testing via a newly proposed test-time tuning strategy to achieve high performance and flexibility in target tracking. Recognizing the limitations of existing event-based tracking datasets, which are predominantly low-resolution, we propose EventVOT, the first large-scale high-resolution event-based tracking dataset. It comprises 1141 videos spanning diverse categories such as pedestrians, vehicles, UAVs, ping pong, etc. Extensive experiments on both low-resolution (FE240hz, VisEvent, FELT), and our newly proposed high-resolution EventVOT dataset fully validated the effectiveness of our proposed method. Both the benchmark dataset and source code have been released on https://github.com/Event-AHU/EventVOT_Benchmark

Large Language Models (LLMs) exhibit potential artificial generic intelligence recently, however, their usage is costly with high response latency. Given mixed LLMs with their own strengths and weaknesses, LLM routing aims to identify the most suitable model for each query in the stream to maximize response quality and minimize cost and latency. However, the challenges involve: (1) dynamic trade-offs among quality, cost, and latency; (2) enabling continual learning in deployed systems; and (3) navigating a varying (e.g., new LLM addition or old LLM removal) set of LLM candidates over time. To bridge these gaps, we develop MixLLM, a dynamic contextual-bandit-based routing system for query-LLM assignment. Specifically, we first leverage query tags to enhance query embeddings for the routing task. Next, we design lightweight prediction models to estimate the response qualities and costs of queries over LLMs. We then devise a meta-decision maker to choose the query-LLM assignments to best tradeoff response quality, cost, and latency. Finally, the system benefits from continual training, allowing it to adapt to evolving queries and user feedback over time. Our extensive experiments show that MixLLM achieves the best trade-offs in response quality, cost, and latency (97.25% of GPT-4's quality at 24.18% of the cost under the time constraint).

Vision-Language-Action (VLA) models have shown substantial potential in real-world robotic manipulation. However, fine-tuning these models through supervised learning struggles to achieve robust performance due to limited, inconsistent demonstrations, especially in contact-rich environments. In this paper, we propose a reinforced fine-tuning approach for VLA models, named ConRFT, which consists of offline and online fine-tuning with a unified consistency-based training objective, to address these challenges. In the offline stage, our method integrates behavior cloning and Q-learning to effectively extract policy from a small set of demonstrations and stabilize value estimating. In the online stage, the VLA model is further fine-tuned via consistency policy, with human interventions to ensure safe exploration and high sample efficiency. We evaluate our approach on eight diverse real-world manipulation tasks. It achieves an average success rate of 96.3% within 45-90 minutes of online fine-tuning, outperforming prior supervised methods with a 144% improvement in success rate and 1.9x shorter episode length. This work highlights the potential of integrating reinforcement learning to enhance the performance of VLA models for real-world robotic applications.

We thank the reviewers for their thoughtful reviews, helpful suggestions, and the consistent feedback that the ideas are well presented and the results demonstrate a significant advance. Overall We believe that our presentation of a novel application is well suited to NIPS. NIPS has a culture and history of pushing forward both theory and application, and each makes the other stronger. Indeed, the NIPS call for papers specifically cites applications as one of the 10 technical areas of interest. As one recent example, Krizhevsky et al 2012 focused largely on one application, but has been transformative to the fields of computer vision and deep learning.

Summary and Contributions: This work shows that there is a class that is privately PAC learnable in polynomial time, but not efficiently learnable (assuming the existence of one-way functions) in the online setting, i.e., there is no polynomial time algorithm with a polynomial mistake bound. A line of recent work (focused on sample complexity) has demonstrated various equivalences between private PAC and online learning; this result focuses on the question of efficiency, proving that efficient private learnability does not imply efficient online learnability if one-way functions exist. The cryptographic assumption is standard for such results and is needed when ruling out general polynomial time learners. The class of functions that cannot be efficiently learned in the online model was given by [Blum 1994], which showed a separation between distribution free PAC learning and online learning. Much of the work toward separating the models of learning, which involves working out the cryptographic construction and giving an efficient PAC algorithm, was done there -- the technical contribution here is to give a private version of that algorithm.

Weaknesses: I have many reservation against the claims of the paper. I would appreciate it if the authors can clarify some of these issues during their rebuttal. First, the proof of their main theorem about logistic regression has many issues. One key issue is that the authors make assumptions within the proof that are not clearly stated or justified upfront. For example, in Line 440 in the supplementary materials, the proof assumes that theta Tv .1.

The paper studies the following interesting phenomenon (observed in the previous literature): when trained on the dataset with incorrectly labeled points (i.e. "label noise"), DNNs first learn the benign ("correctly labeled") points and once this is done they start "memorizing" the noisy points. It was previously shown in the literature (empirically) that the second "memorization" phase hurts the generalization. The authors make 2 Contributions: (Contribution 1) They demonstrate (empirically and theoretically) that similar phenomenon can be observed in the simpler setting of the over-parametrized (dimensionality number of points) linear two-class logistic regression, when the class distributions are isotropic Gaussian with fixed means \pm mu and vanishing variance (see Theorem 1 and Figure A.1). (Contribution 2) Motivated by the theory of contribution 1, the authors propose a novel regularizer. When used in the vanilla DNN training with the cross-entropy loss, this regularizer successfully prevents the networks from falling to the "memorization phase" (as evidenced by Figure 1). All the reviewers agree that the topic and the focus of this paper is very timely.

The paper presents some extensions to the Pentina and Lampert's PAC-Bayesian analysis of "Lifelong Learning" problems (ICML 2014), where a learner must adapt to various tasks exploiting knowledge from previously seen ones. The main contributions are risk bounds dedicated to two scenarios where the observed task are not sampled independently from each other. Roughly speaking, the first scenario share similarities with domain adaptation (albeit the risk bound is given on an average of all possible domains, instead of on a specific target domain) and the second is quasi-identical (up to my knowledge) to distribution drift. In the first setting (Section 3), the authors cleverly reuse Ralaivola et al.'s chromatic PAC-Bayesian theory to represent dependencies between tasks. However, this result alone let me unsatisfied. I wonder to which extent this result can be useful to the ambitious "lifelong learning" problem the authors are interested in.

This paper is concerned with optimization via a "ball optimization oracle", which returns the minimizer of a function restricted to an L2 ball of radius r around a query point x. The authors demonstrate an oracle complexity of roughly (R/r) {2/3} when combined with a Monteiro-Svaiter acceleration scheme. The authors show that this oracle can be implemented on a variety of important machine learning problems. The ideas in this paper are elegant and surprising, despite arising from a "deceptively simple" oracle. The reviewers were unanimously positive about this work, and everyone agrees it is an important theoretical contribution to the optimization community.