A key concern with the concept of *“alignment”* is the implicit question of *“alignment to what?”*. AI systems are increasingly used across the world, yet safety alignment is often focused on homogeneous monolingual settings. Additionally, preference training and safety measures often overfit to harms common in Western-centric datasets. Here, we explore the viability of different alignment approaches when balancing dual objectives: addressing and optimizing for a non-homogeneous set of languages and cultural preferences while minimizing both global and local harms. We collect the first human annotated red teaming prompts in different languages, distinguishing between global and local harm, which serve as a laboratory to understand the reliability of alignment techniques when faced with preference distributions that are non-stationary across geographies and languages. While this setting is seldom covered by the literature to date, which primarily centers on English harm mitigation, it captures real-world interactions with AI systems around the world. We establish a new precedent for state-of-the-art alignment techniques across 6 languages with minimal degradation in general performance. Our work provides important insights into cross-lingual transfer and novel optimization approaches to safeguard AI systems designed to serve global populations.
论文检索
输入标题、作者或关键词，从 2,164 篇学术成果中精准定位
Seeking answers effectively for long videos is essential to build video question answering (videoQA) systems. Previous methods adaptively select frames and regions from long videos to save computations. However, this fails to reason over the whole sequence of video, leading to sub-optimal performance. To address this problem, we introduce a state space layer (SSL) into multi-modal Transformer to efficiently integrate global semantics of the video, which mitigates the video information loss caused by frame and region selection modules. Our SSL includes a gating unit to enable controllability over the flow of global semantics into visual representations. To further enhance the controllability, we introduce a cross-modal compositional congruence objective to encourage global semantics aligned with the question. To rigorously evaluate long-form videoQA capacity, we construct two new benchmarks Ego-QA and MAD-QA featuring videos of considerably long length, i.e. 17.5 minutes and 1.9 hours, respectively. Extensive experiments demonstrate the superiority of our framework on these new as well as existing datasets.
With the growing popularity of general-purpose Large Language Models (LLMs), comes a need for more global explanations of model behaviors. Concept-based explanations arise as a promising avenue for explaining high-level patterns learned by LLMs. Yet their evaluation poses unique challenges, especially due to their non-local nature and high dimensional representation in a model’s hidden space. Current methods approach concepts from different perspectives, lacking a unified formalization. This makes evaluating the core measures of concepts, namely faithfulness or readability, challenging. To bridge the gap, we introduce a formal definition of concepts generalizing to diverse concept-based explanations’ settings. Based on this, we quantify the faithfulness of a concept explanation via perturbation. We ensure adequate perturbation in the high-dimensional space for different concepts via an optimization problem. Readability is approximated via an automatic and deterministic measure, quantifying the coherence of patterns that maximally activate a concept while aligning with human understanding. Finally, based on measurement theory, we apply a meta-evaluation method for evaluating these measures, generalizable to other types of explanations or tasks as well. Extensive experimental analysis has been conducted to inform the selection of explanation evaluation measures.
Optimization · Discrete and Combinatorial Optimization
Neural network pruning is a key technique towards engineering large yet scalable, interpretable, and generalizable models. Prior work on the subject has developed largely along two orthogonal directions: (1) differentiable pruning for efficiently and accurately scoring the importance of parameters, and (2) combinatorial optimization for efficiently searching over the space of sparse models. We unite the two approaches, both theoretically and empirically, to produce a coherent framework for structured neural network pruning in which differentiable pruning guides combinatorial optimization algorithms to select the most important sparse set of parameters. Theoretically, we show how many existing differentiable pruning techniques can be understood as nonconvex regularization for group sparse optimization, and prove that for a wide class of nonconvex regularizers, the global optimum is unique, group-sparse, and provably yields an approximate solution to a sparse convex optimization problem. The resulting algorithm that we propose, SequentialAttention++, advances the state of the art in large-scale neural network block-wise pruning tasks on the ImageNet and Criteo datasets.
Reinforcement Learning · Everything Else
Learning from human preference data has emerged as the dominant paradigm for fine-tuning large language models (LLMs). The two most common families of techniques -- online reinforcement learning (RL) such as Proximal Policy Optimization (PPO) and offline contrastive methods such as Direct Preference Optimization (DPO) -- were positioned as equivalent in prior work due to the fact that both have to start from the same offline preference dataset. To further expand our theoretical understanding of the similarities and differences between online and offline techniques for preference fine-tuning, we conduct a rigorous analysis through the lens of *dataset coverage*, a concept that captures how the training data covers the test distribution and is widely used in RL. We prove that a global coverage condition is both necessary and sufficient for offline contrastive methods to converge to the optimal policy, but a weaker partial coverage condition suffices for online RL methods. This separation provides one explanation of why online RL methods can perform better than offline methods, especially when the offline preference data is not diverse enough. Finally, motivated by our preceding theoretical observations, we derive a hybrid preference optimization (HyPO) algorithm that uses offline data for contrastive-based preference optimization and online unlabeled data for KL regularization. Theoretically and empirically, we demonstrate that HyPO is more performant than its pure offline counterpart DPO, while still preserving its computation and memory efficiency.
Reinforcement Learning · Planning
Controlling the evolution of complex physical systems is a fundamental task across science and engineering. Classical techniques suffer from limited applicability or huge computational costs. On the other hand, recent deep learning and reinforcement learning-based approaches often struggle to optimize long-term control sequences under the constraints of system dynamics. In this work, we introduce Diffusion Physical systems Control (DiffPhyCon), a new class of method to address the physical systems control problem. DiffPhyCon excels by simultaneously minimizing both the learned generative energy function and the predefined control objectives across the entire trajectory and control sequence. Thus, it can explore globally and plan near-optimal control sequences. Moreover, we enhance DiffPhyCon with prior reweighting, enabling the discovery of control sequences that significantly deviate from the training distribution. We test our method on three tasks: 1D Burgers' equation, 2D jellyfish movement control, and 2D high-dimensional smoke control, where our generated jellyfish dataset is released as a benchmark for complex physical system control research. Our method outperforms widely applied classical approaches and state-of-the-art deep learning and reinforcement learning methods. Notably, DiffPhyCon unveils an intriguing fast-close-slow-open pattern observed in the jellyfish, aligning with established findings in the field of fluid dynamics. The project website, jellyfish dataset, and code can be found at https://github.com/AI4Science-WestlakeU/diffphycon.
Although Large Visual Language Models (LVLMs) have demonstrated exceptional abilities in understanding multimodal data, they invariably suffer from hallucinations, leading to a disconnection between the generated text and the corresponding images. Almost all current visual contrastive decoding methods attempt to mitigate these hallucinations by introducing visual uncertainty information that appropriately widens the contrastive logits gap between hallucinatory and targeted ones. However, due to uncontrollable nature of the global visual uncertainty, they struggle to precisely induce the hallucinatory tokens, which severely limits their effectiveness in mitigating hallucinations and may even lead to the generation of undesired hallucinations. To tackle this issue, we conducted the theoretical analysis to promote the effectiveness of contrast decoding. Building on this insight, we introduce a novel optimization strategy named Hallucination-Induced Optimization (HIO). This strategy seeks to amplify the contrast between hallucinatory and targeted tokens relying on a fine-tuned theoretical preference model (i.e., Contrary Bradley-Terry Model), thereby facilitating efficient contrast decoding to alleviate hallucinations in LVLMs. Extensive experimental research demonstrates that our HIO strategy can effectively reduce hallucinations in LVLMs, outperforming state-of-the-art methods across various benchmarks.
Generative Models · Diffusion models
Diffusion models have demonstrated empirical successes in various applications and can be adapted to task-specific needs via guidance. This paper studies a form of gradient guidance for adapting a pre-trained diffusion model towards optimizing user-specified objectives. We establish a mathematical framework for guided diffusion to systematically study its optimization theory and algorithmic design. Our theoretical analysis spots a strong link between guided diffusion models and optimization: gradient-guided diffusion models are essentially sampling solutions to a regularized optimization problem, where the regularization is imposed by the pre-training data. As for guidance design, directly bringing in the gradient of an external objective function as guidance would jeopardize the structure in generated samples. We investigate a modified form of gradient guidance based on a forward prediction loss, which leverages the information in pre-trained score functions and provably preserves the latent structure. We further consider an iteratively fine-tuned version of gradient-guided diffusion where guidance and score network are both updated with newly generated samples. This process mimics a first-order optimization iteration in expectation, for which we proved $\tilde{\mathcal{O}}(1/K)$ convergence rate to the global optimum when the objective function is concave. Our code is released at https://github.com/yukang123/GGDMOptim.git.
Deep Learning · Algorithms
Second-order optimization algorithms, such as the Newton method and the natural gradient descent (NGD) method exhibit excellent convergence properties for training deep neural networks, but the high computational cost limits its practical application. In this paper, we focus on the NGD method and propose a novel layer-wise natural gradient descent (LNGD) method to further reduce computational costs and accelerate the training process. Specifically, based on the block diagonal approximation of the Fisher information matrix, we first propose the layer-wise sample method to compute each block matrix without performing a complete back-propagation. Then, each block matrix is approximated as a Kronecker product of two smaller matrices, one of which is a diagonal matrix, while keeping the traces equal before and after approximation. By these two steps, we provide a new approximation for the Fisher information matrix, which can effectively reduce the computational cost while preserving the main information of each block matrix. Moreover, we propose a new adaptive layer-wise learning rate to further accelerate training. Based on these new approaches, we propose the LNGD optimizer. The global convergence analysis of LNGD is established under some assumptions. Experiments on image classification and machine translation tasks show that our method is quite competitive compared to the state-of-the-art methods.
Deep Learning · Everything Else
Due to the lack of state dimension optimization methods, deep state space models (SSMs) have sacrificed model capacity, training search space, or stability to alleviate computational costs caused by high state dimensions. In this work, we provide a structured pruning method for SSMs, Layer-Adaptive STate pruning (LAST), which reduces the state dimension of each layer in minimizing model-level output energy loss by extending modal truncation for a single system. LAST scores are evaluated using the $\mathcal{H}_{\infty}$ norms of subsystems and layer-wise energy normalization. The scores serve as global pruning criteria, enabling cross-layer comparison of states and layer-adaptive pruning. Across various sequence benchmarks, LAST optimizes previous SSMs, revealing the redundancy and compressibility of their state spaces. Notably, we demonstrate that, on average, pruning 33\% of states still maintains performance with 0.52\% accuracy loss in multi-input multi-output SSMs without retraining. Code is available at https://github.com/msgwak/LAST.
Causality
Existing approaches to differentiable structure learning of directed acyclic graphs (DAGs) rely on strong identifiability assumptions in order to guarantee that global minimizers of the acyclicity-constrained optimization problem identifies the true DAG. Moreover, it has been observed empirically that the optimizer may exploit undesirable artifacts in the loss function. We explain and remedy these issues by studying the behavior of differentiable acyclicity-constrained programs under general likelihoods with multiple global minimizers. By carefully regularizing the likelihood, it is possible to identify the sparsest model in the Markov equivalence class, even in the absence of an identifiable parametrization. We first study the Gaussian case in detail, showing how proper regularization of the likelihood defines a score that identifies the sparsest model. Assuming faithfulness, it also recovers the Markov equivalence class. These results are then generalized to general models and likelihoods, where the same claims hold. These theoretical results are validated empirically, showing how this can be done using standard gradient-based optimizers (without resorting to approximations such as Gumbel-Softmax), thus paving the way for differentiable structure learning under general models and losses. Open-source code is available at \url{https://github.com/duntrain/dagrad}.
Reinforcement Learning · Multi-agent
In this paper, we propose a provably efficient natural policy gradient algorithm called Spectral Dynamic Embedding Policy Optimization (\SDEPO) for two-player zero-sum stochastic Markov games with continuous state space and finite action space. In the policy evaluation procedure of our algorithm, a novel kernel embedding method is employed to construct a finite-dimensional linear approximations to the state-action value function. We explicitly analyze the approximation error in policy evaluation, and show that \SDEPO\ achieves an $\tilde{O}(\frac{1}{(1-\gamma)^3\epsilon})$ last-iterate convergence to the $\epsilon-$optimal Nash equilibrium, which is independent of the cardinality of the state space. The complexity result matches the best-known results for global convergence of policy gradient algorithms for single agent setting. Moreover, we also propose a practical variant of \SDEPO\ to deal with continuous action space and empirical results demonstrate the practical superiority of the proposed method.
Optimization · Zero-order and Black-box Optimization
High-dimensional Bayesian optimization (BO) tasks such as molecular design often require $>10,$$000$ function evaluations before obtaining meaningful results. While methods like sparse variational Gaussian processes (SVGPs) reduce computational requirements in these settings, the underlying approximations result in suboptimal data acquisitions that slow the progress of optimization. In this paper we modify SVGPs to better align with the goals of BO: targeting informed data acquisition over global posterior fidelity. Using the framework of utility-calibrated variational inference (Lacoste–Julien et al., 2011), we unify GP approximation and data acquisition into a joint optimization problem, thereby ensuring optimal decisions under a limited computational budget. Our approach can be used with any decision-theoretic acquisition function and is readily compatible with trust region methods like TuRBO (Eriksson et al., 2019). We derive efficient joint objectives for the expected improvement (EI) and knowledge gradient (KG) acquisition functions in both the standard and batch BO settings. On a variety of recent high dimensional benchmark tasks in control and molecular design, our approach significantly outperforms standard SVGPs and is capable of achieving comparable rewards with up to $10\times$ fewer function evaluations.
In modern chip design, placement aims at placing millions of circuit modules, which is an essential step that significantly influences power, performance, and area (PPA) metrics. Recently, reinforcement learning (RL) has emerged as a promising technique for improving placement quality, especially macro placement. However, current RL-based placement methods suffer from long training times, low generalization ability, and inability to guarantee PPA results. A key issue lies in the problem formulation, i.e., using RL to place from scratch, which results in limits useful information and inaccurate rewards during the training process. In this work, we propose an approach that utilizes RL for the refinement stage, which allows the RL policy to learn how to adjust existing placement layouts, thereby receiving sufficient information for the policy to act and obtain relatively dense and precise rewards. Additionally, we introduce the concept of regularity during training, which is considered an important metric in the chip design industry but is often overlooked in current RL placement methods. We evaluate our approach on the ISPD 2005 and ICCAD 2015 benchmark, comparing the global half-perimeter wirelength and regularity of our proposed method against several competitive approaches. Besides, we test the PPA performance using commercial software, showing that RL as a regulator can achieve significant PPA improvements. Our RL regulator can fine-tune placements from any method and enhance their quality. Our work opens up new possibilities for the application of RL in placement, providing a more effective and efficient approach to optimizing chip design. Our code is available at \url{https://github.com/lamda-bbo/macro-regulator}.
Neural Collapse (NC) was a recently discovered phenomenon that the output features and the classifier weights of the neural network converge to optimal geometric structures at the Terminal Phase of Training (TPT) under various losses. However, the relationship between these optimal structures at TPT and the classification performance remains elusive, especially in imbalanced learning. Even though it is noticed that fixing the classifier to an optimal structure can mitigate the minority collapse problem, the performance is still not comparable to the classical imbalanced learning methods with a learnable classifier. In this work, we find that the optimal structure can be designed to represent a better classification rule, and thus achieve better performance. In particular, we justify that, to achieve better classification, the features from the minor classes should align with more directions. This justification then yields a decision rule called the Generalized Classification Rule (GCR) and we also term these directions as the centers of the classes. Then we study the NC under an MSE-type loss via the Unconstrained Features Model (UFM) framework where (1) the features from a class tend to collapse to the mean of the corresponding centers of that class (named Neural Collapse to Multiple Centers (NCMC)) at the global optimum, and (2) the original classifier approximates a surrogate to GCR when NCMC occurs. Based on the analysis, we develop a strategy for determining the number of centers and propose a Cosine Loss function for the fixed classifier that induces NCMC. Our experiments have shown that the Cosine Loss can induce NCMC and has performance on long-tail classification comparable to the classical imbalanced learning methods.
Theory · Reinforcement Learning and Planning
Occupancy functions play an instrumental role in reinforcement learning (RL) for guiding exploration, handling distribution shift, and optimizing general objectives beyond the expected return. Yet, computationally efficient policy optimization methods that use (only) occupancy functions are virtually non-existent. In this paper, we establish the theoretical foundations of model-free policy gradient (PG) methods that compute the gradient through the occupancy for both online and offline RL, without modeling value functions. Our algorithms reduce gradient estimation to squared-loss regression and are computationally oracle-efficient. We characterize the sample complexities of both local and global convergence, accounting for both finite-sample estimation error and the roles of exploration (online) and data coverage (offline). Occupancy-based PG naturally handles arbitrary offline data distributions, and, with one-line algorithmic changes, can be adapted to optimize any differentiable objective functional.
One important property of DIstribution Correction Estimation (DICE) methods is that the solution is the optimal stationary distribution ratio between the optimized and data collection policy. In this work, we show that DICE-based methods can be viewed as a transformation from the behavior distribution to the optimal policy distribution. Based on this, we propose a novel approach, Diffusion-DICE, that directly performs this transformation using diffusion models. We find that the optimal policy's score function can be decomposed into two terms: the behavior policy's score function and the gradient of a guidance term which depends on the optimal distribution ratio. The first term can be obtained from a diffusion model trained on the dataset and we propose an in-sample learning objective to learn the second term. Due to the multi-modality contained in the optimal policy distribution, the transformation in Diffusion-DICE may guide towards those local-optimal modes. We thus generate a few candidate actions and carefully select from them to achieve global-optimum. Different from all other diffusion-based offline RL methods, the \textit{guide-then-select} paradigm in Diffusion-DICE only uses in-sample actions for training and brings minimal error exploitation in the value function. We use a didatic toycase example to show how previous diffusion-based methods fail to generate optimal actions due to leveraging these errors and how Diffusion-DICE successfully avoid that. We then conduct extensive experiments on benchmark datasets to show the strong performance of Diffusion-DICE.
Optimization · Large Scale, Parallel and Distributed
Due to the heterogeneous architectures and class skew, the global representation models training in resource-adaptive federated self-supervised learning face with tricky challenges: $\textit{deviated representation abilities}$ and $\textit{inconsistent representation spaces}$. In this work, we are the first to propose a multi-teacher knowledge distillation framework, namely $\textit{FedMKD}$, to learn global representations with whole class knowledge from heterogeneous clients even under extreme class skew. Firstly, the adaptive knowledge integration mechanism is designed to learn better representations from all heterogeneous models with deviated representation abilities. Then the weighted combination of the self-supervised loss and the distillation loss can support the global model to encode all classes from clients into a unified space. Besides, the global knowledge anchored alignment module can make the local representation spaces close to the global spaces, which further improves the representation abilities of local ones. Finally, extensive experiments conducted on two datasets demonstrate the effectiveness of $\textit{FedMKD}$ which outperforms state-of-the-art baselines 4.78\% under linear evaluation on average.
Theory · Reinforcement Learning and Planning
This paper studies first-order policy optimization for robust average cost Markov decision processes (MDPs). Specifically, we focus on ergodic Markov chains. For robust average cost MDPs, the goal is to optimize the worst-case average cost over an uncertainty set of transition kernels. We first develop a sub-gradient of the robust average cost. Based on the sub-gradient, a robust policy mirror descent approach is further proposed. To characterize its iteration complexity, we develop a lower bound on the difference of robust average cost between two policies and further show that the robust average cost satisfies the PL-condition. We then show that with increasing step size, our robust policy mirror descent achieves a linear convergence rate in the optimality gap, and with constant step size, our algorithm converges to an $\epsilon$-optimal policy with an iteration complexity of $\mathcal{O}(1/\epsilon)$. The convergence rate of our algorithm matches with the best convergence rate of policy-based algorithms for robust MDPs. Moreover, our algorithm is the first algorithm that converges to the global optimum with general uncertainty sets for robust average cost MDPs. We provide simulation results to demonstrate the performance of our algorithm.
Reinforcement Learning · Function Approximation
We study episodic linear mixture MDPs with the unknown transition and adversarial rewards under full-information feedback, employing *dynamic regret* as the performance measure. We start with in-depth analyses of the strengths and limitations of the two most popular methods: occupancy-measure-based and policy-based methods. We observe that while the occupancy-measure-based method is effective in addressing non-stationary environments, it encounters difficulties with the unknown transition. In contrast, the policy-based method can deal with the unknown transition effectively but faces challenges in handling non-stationary environments. Building on this, we propose a novel algorithm that combines the benefits of both methods. Specifically, it employs (i) an *occupancy-measure-based global optimization* with a two-layer structure to handle non-stationary environments; and (ii) a *policy-based variance-aware value-targeted regression* to tackle the unknown transition. We bridge these two parts by a novel conversion. Our algorithm enjoys an $\widetilde{\mathcal{O}}(d \sqrt{H^3 K} + \sqrt{HK(H + \bar{P}_K)})$ dynamic regret, where $d$ is the feature mapping dimension, $H$ is the episode length, $K$ is the number of episodes, $\bar{P}_K$ is the non-stationarity measure. We show it is minimax optimal up to logarithmic factors by establishing a matching lower bound. To the best of our knowledge, this is the **first** work that achieves **near-optimal** dynamic regret for adversarial linear mixture MDPs with the unknown transition without prior knowledge of the non-stationarity measure.