ICLR 2025 Papers

Large Scale Pre-Trained Language Models (PTMs) have demonstrated unprecedented capabilities across diverse natural language processing tasks. Adapting such models to downstream tasks is computationally intensive and time-consuming, particularly in black-box scenarios common in Language-Model-as-a-Service (LMaaS) environments, where model parameters and gradients are inaccessible. Recently, black-box prompt learning using zeroth-order gradients has emerged as a promising approach to address these challenges by optimizing learnable continuous prompts in embedding spaces, starting with \textit{randomly initialized discrete text prompts}. However, its reliance on randomly initialized discrete prompts limits adaptability to diverse downstream tasks or models. To address this limitation, this paper introduces ZO-PoG, a novel framework that optimizes prompts through a collaborative approach, combining Policy Gradient optimization for initial discrete text prompts and Zeroth-Order optimization for continuous prompts in embedding space. By optimizing collaboratively between discrete and continuous prompts, ZO-PoG maximizes adaptability to downstream tasks, achieving superior results without direct access to the model’s internal structures. Importantly, we establish the sub-linear convergence of ZO-PoG under mild assumptions. The experiments on different datasets demonstrate significant improvements in various tasks compared to the baselines.

Offline model-based optimization (MBO) aims to identify a design that maximizes a black-box function using only a fixed, pre-collected dataset of designs and their corresponding scores. This problem has garnered significant attention from both scientific and industrial domains. A common approach in offline MBO is to train a regression-based surrogate model by minimizing mean squared error (MSE) and then find the best design within this surrogate model by different optimizers (e.g., gradient ascent). However, a critical challenge is the risk of out-of-distribution errors, i.e., the surrogate model may typically overestimate the scores and mislead the optimizers into suboptimal regions. Prior works have attempted to address this issue in various ways, such as using regularization techniques and ensemble learning to enhance the robustness of the model, but it still remains. In this paper, we argue that regression models trained with MSE are not well-aligned with the primary goal of offline MBO, which is to \textit{select} promising designs rather than to predict their scores precisely. Notably, if a surrogate model can maintain the order of candidate designs based on their relative score relationships, it can produce the best designs even without precise predictions. To validate it, we conduct experiments to compare the relationship between the quality of the final designs and MSE, finding that the correlation is really very weak. In contrast, a metric that measures order-maintaining quality shows a significantly stronger correlation. Based on this observation, we propose learning a ranking-based model that leverages learning to rank techniques to prioritize promising designs based on their relative scores. We show that the generalization error on ranking loss can be well bounded. Empirical results across diverse tasks demonstrate the superior performance of our proposed ranking-based method than twenty existing methods. Our implementation is available at \url{https://github.com/lamda-bbo/Offline-RaM}.

Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity. To understand this inconsistency, we unify existing methods into a standard framework, showing they are equivalent to solving a common optimization problem: minimize average loss subject to a method-specific mixing law---an implicit assumption on the relationship between loss and mixture proportions. This framework suggests that measuring the fidelity of a method's mixing law can offer insights into its performance. Empirically, we find that existing methods set their mixing law parameters inaccurately, resulting in the inconsistent mixing performance we observe. Using this insight, we derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Empirically, Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.27 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.012 test perplexity points.

Optimization in deep learning remains poorly understood. A key difficulty is that optimizers exhibit complex oscillatory dynamics, referred to as "edge of stability," which cannot be captured by traditional optimization theory. In this paper, we show that the path taken by an oscillatory optimizer can often be captured by a _central flow_: a differential equation which directly models the time-averaged (i.e. smoothed) optimization trajectory. We empirically show that these central flows can predict long-term optimization trajectories for generic neural networks with a high degree of numerical accuracy. By interpreting these flows, we are able to understand how gradient descent makes progress even as the loss sometimes goes up; how adaptive optimizers ``adapt'' to the local loss landscape; and how adaptive optimizers implicitly seek out regions of weight space where they can take larger steps. These insights (and others) are not apparent from the optimizers' update rules, but are revealed by the central flows. Therefore, we believe that central flows constitute a promising tool for reasoning about optimization in deep learning.

Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty. Following the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE. On the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability.

In finite-sum optimization problems, the sample orders for parameter updates can significantly influence the convergence rate of optimization algorithms. While numerous sample ordering techniques have been proposed in the context of single-objective optimization, the problem of sample ordering in finite-sum multi-objective optimization has not been thoroughly explored. To address this gap, we propose a sample ordering method called JoGBa, which finds the sample orders for multiple objectives by jointly performing online vector balancing on the gradients of all objectives. Our theoretical analysis demonstrates that this approach outperforms the standard baseline of random ordering and accelerates the convergence rate for the MGDA algorithm. Empirical evaluation across various datasets with different multi-objective optimization algorithms further demonstrates that JoGBa can achieve faster convergence and superior final performance than other data ordering strategies.

The need to explain decisions made by AI systems is driven by both recent regulation and user demand. The decisions are often explainable only post hoc. In counterfactual explanations, one may ask what constitutes the best counterfactual explanation. Clearly, multiple criteria must be taken into account, although "distance from the sample" is a key criterion. Recent methods that consider the plausibility of a counterfactual seem to sacrifice this original objective. Here, we present a system that provides high-likelihood explanations that are, at the same time, close and sparse. We show that the search for the most likely explanations satisfying many common desiderata for counterfactual explanations can be modeled using Mixed-Integer Optimization (MIO). We use a Sum-Product Network (SPN) to estimate the likelihood of a counterfactual. To achieve that, we propose an MIO formulation of an SPN, which can be of independent interest. The source code with examples is available at https://github.com/Epanemu/LiCE.

Training Large Language Models (LLMs) poses significant memory challenges, making low-rank adaptation methods an attractive solution. Previously, Low-Rank Adaptation (LoRA) addressed this by adding a trainable low-rank matrix to the frozen pre-trained weights in each layer, reducing the number of trainable parameters and optimizer states. GaLore, which compresses the gradient matrix instead of the weight matrix, has demonstrated superior performance to LoRA with faster convergence and reduced memory consumption. Despite their empirical success, the performance of these methods has not been fully understood or explained theoretically. In this paper, we analyze the optimization landscapes of LoRA, GaLore, and full-rank methods, revealing that GaLore benefits from fewer spurious local minima and a larger region that satisfies the \pl, a variant of Polyak-Łojasiewicz (PL) condition, leading to faster convergence. Our analysis leads to a novel method, GaRare, which further improves GaLore by using gradient random projection to reduce computational overhead. Practically, GaRare achieves strong performance in both pre-training and fine-tuning tasks, offering a more efficient approach to large-scale model adaptation.

Learning to sample from complex unnormalized distributions over discrete domains emerged as a promising research direction with applications in statistical physics, variational inference, and combinatorial optimization. Recent work has demonstrated the potential of diffusion models in this domain. However, existing methods face limitations in memory scaling and thus the number of attainable diffusion steps since they require backpropagation through the entire generative process. To overcome these limitations we introduce two novel training methods for discrete diffusion samplers, one grounded in the policy gradient theorem and the other one leveraging Self-Normalized Neural Importance Sampling (SN-NIS). These methods yield memory-efficient training and achieve state-of-the-art results in unsupervised combinatorial optimization. Numerous scientific applications additionally require the ability of unbiased sampling. We introduce adaptations of SN-NIS and Neural Markov Chain Monte Carlo that enable for the first time the application of discrete diffusion models to this problem. We validate our methods on Ising model benchmarks and find that they outperform popular autoregressive approaches. Our work opens new avenues for applying diffusion models to a wide range of scientific applications in discrete domains that were hitherto restricted to exact likelihood models.

The spectral risk has wide applications in machine learning, especially in real-world decision-making, where people are concerned with more than just average model performance. By assigning different weights to the losses of different sample points, rather than the same weights as in the empirical risk, it allows the model's performance to lie between the average performance and the worst-case performance. In this paper, we propose SOREL, the first stochastic gradient-based algorithm with convergence guarantees for spectral risks minimization. Previous approaches often rely on smoothing the spectral risk by adding a strongly concave function, thereby lacking convergence guarantees for the original spectral risk. We theoretically prove that our algorithm achieves a near-optimal rate of $\widetilde{O}(1/\sqrt{\epsilon})$ to obtain an $\epsilon$-optimal solution in terms $\epsilon$. Experiments on real datasets show that our algorithm outperforms existing ones in most cases, both in terms of runtime and sample complexity.

Federated Learning (FL) offers a promising framework for collaborative and privacy-preserving machine learning across distributed data sources. However, the substantial communication costs associated with FL significantly challenge its efficiency. Specifically, in each communication round, the communication costs scale linearly with the model's dimension, which presents a formidable obstacle, especially in large model scenarios. Despite various communication-efficient strategies, the intrinsic dimension-dependent communication cost remains a major bottleneck for current FL implementations. This paper proposes a novel dimension-free communication algorithm - DeComFL, which leverages the zeroth-order optimization techniques and reduces the communication cost from $\mathcal{O}(d)$ to $\mathcal{O}(1)$ by transmitting only a constant number of scalar values between clients and the server in each round, regardless of the dimension $d$ of the model parameters. Theoretically, in non-convex functions, we prove that our algorithm achieves state-of-the-art rates, which show a linear speedup of the number of clients and local steps under standard assumptions. With additional low effective rank assumption, we can further show that the convergence rate is independent of the model dimension $d$ as well. Empirical evaluations, encompassing both classic deep learning training and large language model fine-tuning, demonstrate significant reductions in communication overhead. Notably, DeComFL achieves this by transmitting only around 1MB of data in total between the server and a client to fine-tune a model with billions of parameters. The code is available at https://github.com/ZidongLiu/DeComFL.

Direct Preference Optimization (DPO) has been widely adopted for preference alignment of Large Language Models (LLMs) due to its simplicity and effectiveness. However, DPO is derived as a bandit problem in which the whole response is treated as a single arm, ignoring the importance differences between tokens, which may affect optimization efficiency and make it difficult to achieve optimal results. In this work, we propose that the optimal data for DPO has equal expected rewards for each token in winning and losing responses, as there is no difference in token importance. However, since the optimal dataset is unavailable in practice, we propose using the original dataset for importance sampling to achieve unbiased optimization. Accordingly, we propose a token-level importance sampling DPO objective named TIS-DPO that assigns importance weights to each token based on its reward. Inspired by previous works, we estimate the token importance weights using the difference in prediction probabilities from a pair of contrastive LLMs. We explore three methods to construct these contrastive LLMs: (1) guiding the original LLM with contrastive prompts, (2) training two separate LLMs using winning and losing responses, and (3) performing forward and reverse DPO training with winning and losing responses. Experiments show that TIS-DPO significantly outperforms various baseline methods on harmlessness and helpfulness alignment and summarization tasks. We also visualize the estimated weights, demonstrating their ability to identify key token positions.

How to promote the robustness of existing deep learning models is a challenging problem for many practical classification tasks. Recently, Distributionally Robust Optimization (DRO) methods have shown promising potential to tackle this problem. These methods aim to construct reliable models by minimizing the worst-case risk within a local region (called ''uncertainty set'') around the empirical data distribution. However, conventional DRO methods tend to be overly pessimistic, leading to certain discrepancy between the real data distribution and the uncertainty set, which can degrade the classification performance. To address this issue, we propose a manifold-based DRO method that takes the geometric structure of training data into account for constructing the uncertainty set. Specifically, our method employs a carefully designed ''game'' that integrates contrastive learning with Jacobian regularization to capture the manifold structure, enabling us to solve DRO problems constrained by the data manifold. By utilizing a novel idea for approximating geodesic distance on manifolds, we also provide the theoretical guarantees for its robustness. Moreover, our proposed method is easy to implement in practice. We conduct a set of experiments on several popular benchmark datasets, where the results demonstrate our advantages in terms of accuracy and robustness.

Stochastic gradient descent with momentum, also known as Stochastic Heavy Ball method (SHB), is one of the most popular algorithms for solving large-scale stochastic optimization problems in various machine learning tasks. In practical scenarios, tuning the step-size and momentum parameters of the method is a prohibitively expensive and time-consuming process. In this work, inspired by the recent advantages of stochastic Polyak step-size in the performance of stochastic gradient descent (SGD), we propose and explore new Polyak-type variants suitable for the update rule of the SHB method. In particular, using the Iterate Moving Average (IMA) viewpoint of SHB, we propose and analyze three novel step-size selections: MomSPSmax, MomDecSPS, and MomAdaSPS. For MomSPSmax, we provide convergence guarantees for SHB to a neighborhood of the solution for convex and smooth problems (without assuming interpolation). If interpolation is also satisfied, then using MomSPSmax, SHB converges to the true solution at a fast rate matching the deterministic HB. The other two variants, MomDecSPS and MomAdaSPS, are the first adaptive step-size for SHB that guarantee convergence to the exact minimizer - without a priori knowledge of the problem parameters and without assuming interpolation. Our convergence analysis of SHB is tight and obtains the convergence guarantees of stochastic Polyak step-size for SGD as a special case. We supplement our analysis with experiments validating our theory and demonstrating the effectiveness and robustness of our algorithms.

The \textbf{i}nexact \textbf{s}tochastic \textbf{p}roximal \textbf{p}oint \textbf{a}lgorithm (isPPA) is popular for solving stochastic composite optimization problems with many applications in machine learning. While the convergence theory of the (inexact) PPA has been well established, the known convergence guarantees of isPPA require restrictive assumptions. In this paper, we establish the stability and almost sure convergence of isPPA under mild assumptions, where smoothness and (restrictive) strong convexity of the objective function are not required. Imposing a local Lipschitz condition on component functions and a quadratic growth condition on the objective function, we establish last-iterate iteration complexity bounds of isPPA regarding the distance to the solution set and the Karush–Kuhn–Tucker (KKT) residual. Moreover, we show that the established iteration complexity bounds are tight up to a constant by explicitly analyzing the bounds for the regularized Fr\'echet mean problem. We further validate the established convergence guarantees of isPPA by numerical experiments.

Incremental gradient and incremental proximal methods are a fundamental class of optimization algorithms used for solving finite sum problems, broadly studied in the literature. Yet, without strong convexity, their convergence guarantees have primarily been established for the ergodic (average) iterate. We establish the first nonasymptotic convergence guarantees for the last iterate of both incremental gradient and incremental proximal methods, in general convex smooth (for both) and convex Lipschitz (for the proximal variants) settings. Our oracle complexity bounds for the last iterate nearly match (i.e., match up to a square-root-log or a log factor) the best known oracle complexity bounds for the average iterate, for both classes of methods. We further obtain generalizations of our results to weighted averaging of the iterates with increasing weights and for randomly permuted ordering of updates. We study last iterate convergence of the incremental proximal method as a mathematical abstraction of forgetting in continual learning and prove a lower bound that certifies that a large amount of regularization is crucial to mitigating catastrophic forgetting---one of the key considerations in continual learning. Our results generalize last iterate guarantees for incremental methods compared to state of the art, as such results were previously known only for overparameterized linear models, which correspond to convex quadratic problems with infinitely many solutions.

We analyze two variants of Local Gradient Descent applied to distributed logistic regression with heterogeneous, separable data and show convergence at the rate $O(1/KR)$ for $K$ local steps and sufficiently large $R$ communication rounds. In contrast, all existing convergence guarantees for Local GD applied to any problem are at least $\Omega(1/R)$, meaning they fail to show the benefit of local updates. The key to our improved guarantee is showing progress on the logistic regression objective when using a large stepsize $\eta \gg 1/K$, whereas prior analysis depends on $\eta \leq 1/K$.

Decentralized training of deep neural networks has attracted significant attention for its theoretically superior scalability compared to synchronous data-parallel methods like All-Reduce. However, realizing this potential in multi-node training is challenging due to the complex design space that involves communication topologies, computation patterns, and optimization algorithms. This paper identifies three key factors that can lead to speedups over All-Reduce training and constructs a runtime model to determine when and how decentralization can shorten the per-iteration runtimes. To support the decentralized training of transformer-based models, we introduce a decentralized Adam algorithm that overlaps communications with computations, prove its convergence, and propose an accumulation technique to mitigate the high variance caused by small local batch sizes. We deploy our solution in clusters with up to 64 GPUs, demonstrating its practical advantages in both runtime and generalization performance under a fixed iteration budget. The experiment code is open-source at [https://github.com/WangZesen/Decentralized-Training-Exp](https://github.com/WangZesen/Decentralized-Training-Exp), and the extension code is open-source at [https://github.com/WangZesen/Decent-DP](https://github.com/WangZesen/Decent-DP).

While Transformers and other sequence-parallelizable neural network architectures seem like the current state of the art in sequence modeling, they specifically lack state-tracking capabilities. These are important for time-series tasks and logical reasoning. Traditional RNNs like LSTMs and GRUs, as well as modern variants like sLSTM do have these capabilities at the cost of strictly sequential processing. While this is often seen as a strong limitation, we show how fast these networks can get with our hardware-optimization FlashRNN in Triton and CUDA, optimizing kernels to the register level on modern GPUs. We extend traditional RNNs with a parallelization variant that processes multiple RNNs of smaller hidden state in parallel, similar to the head-wise processing in Transformers. To enable flexibility on different GPU variants, we introduce a new optimization framework for hardware-internal cache sizes, memory and compute handling. It models the hardware in a setting using polyhedral-like constraints, including the notion of divisibility. This speeds up the solution process in our ConstrINT library for general integer constraint satisfaction problems (integer CSPs). We show that our kernels can achieve 50x speed-ups over a vanilla PyTorch implementation and allow 40x larger hidden sizes compared to our Triton implementation. We have open-sourced our kernels and the optimization library to boost research in the direction of state-tracking enabled RNNs and sequence modeling here: https://github.com/NX-AI/flashrnn