论文检索

Probabilistic Methods · Gaussian Processes

Siavash Ameli, Chris van der Heide, Liam Hodgkinson, Michael Mahoney

Computing eigenvalues of very large matrices is a critical task in many machine learning applications, including the evaluation of log-determinants, the trace of matrix functions, and other important metrics. As datasets continue to grow in scale, the corresponding covariance and kernel matrices become increasingly large, often reaching magnitudes that make their direct formation impractical or impossible. Existing techniques typically rely on matrix-vector products, which can provide efficient approximations, if the matrix spectrum behaves well. However, in settings like distributed learning, or when the matrix is defined only indirectly, access to the full data set can be restricted to only very small sub-matrices of the original matrix. In these cases, the matrix of nominal interest is not even available as an implicit operator, meaning that even matrix-vector products may not be available. In such settings, the matrix is "impalpable", in the sense that we have access to only masked snapshots of it. We draw on principles from free probability theory to introduce a novel method of "free decompression" to estimate the spectrum of such matrices. Our method can be used to extrapolate from the empirical spectral densities of small submatrices to infer the eigenspectrum of extremely large (impalpable) matrices (that we cannot form or even evaluate with full matrix-vector products). We demonstrate the effectiveness of this approach through a series of examples, comparing its performance against known limiting distributions from random matrix theory in synthetic settings, as well as applying it to submatrices of real-world datasets, matching them with their full empirical eigenspectra.

Social Aspects · Trustworthy Machine Learning

Zhe Tao, Aditya V Thakur

In explainable AI, DNN gradients are used to interpret the prediction; in safety-critical control systems, gradients could encode safety constraints; in scientific-computing applications, gradients could encode physical invariants. While recent work on provable editing of DNNs has focused on input-output constraints, the problem of enforcing hard constraints on DNN gradients remains unaddressed. We present ProGrad, the first efficient approach for editing the parameters of a DNN to provably enforce hard constraints on the DNN gradients. Given a DNN $\mathcal{N}$ with parameters $\theta$, and a set $\mathcal{S}$ of pairs $(\mathrm{x}, \mathrm{Q})$ of input $\mathrm{x}$ and corresponding linear gradient constraints $\mathrm{Q}$, ProGrad finds new parameters $\theta'$ such that $\bigwedge_{(\mathrm{x}, \mathrm{Q}) \in \mathcal{S}} \frac{\partial}{\partial \mathrm{x}}\mathcal{N}(\mathrm{x}; \theta') \in \mathrm{Q}$ while minimizing the changes $\lVert\theta' - \theta\rVert$. The key contribution is a novel *conditional variable gradient* of DNNs, which relaxes the NP-hard provable gradient editing problem to a linear program (LP), enabling ProGrad to use an LP solver to efficiently and effectively enforce the gradient constraints. We experimentally evaluated ProGrad via enforcing (i) hard Grad-CAM constraints on ImageNet ResNet DNNs; (ii) hard Integrated Gradients constraints on Llama 3 and Qwen 3 LLMs; (iii) hard gradient constraints in training a DNN to approximate a target function as a proxy for safety constraints in control systems and physical invariants in scientific applications. The results highlight the unique capability of ProGrad in enforcing hard constraints on DNN gradients.

General Machine Learning · Everything Else

Andong Wang, Yuning Qiu, Haonan Huang, Zhong Jin, Guoxu Zhou, Qibin Zhao

Despite the growing success of transform-based tensor models such as the t-product, their underlying geometric principles remain poorly understood. Classical differential geometry, built on real-valued function spaces, is not well suited to capture the algebraic and spectral structure induced by transform-based tensor operations. In this work, we take an initial step toward a geometric framework for tensors equipped with tube-wise multiplication via orthogonal transforms. We introduce the notion of smooth t-manifolds, defined as topological spaces locally modeled on structured tensor modules over a commutative t-scalar ring. This formulation enables transform-consistent definitions of geometric objects, including metrics, gradients, Laplacians, and geodesics, thereby bridging discrete and continuous tensor settings within a unified algebraic-geometric perspective. On this basis, we develop a statistical procedure for testing whether tensor data lie near a low-dimensional t-manifold, and provide nonasymptotic guarantees for manifold fitting under noise. We further establish approximation bounds for tensor neural networks that learn smooth functions over t-manifolds, with generalization rates determined by intrinsic geometric complexity. This framework offers a theoretical foundation for geometry-aware learning in structured tensor spaces and supports the development of models that align with transform-based tensor representations.

Optimization · Stochastic & Convex Optimization

Aleksey Minabutdinov, Patrick Cheridito

We introduce a novel deep learning algorithm for computing convex conjugates of differentiable convex functions, a fundamental operation in convex analysis with various applications in different fields such as optimization, control theory, physics and economics. While traditional numerical methods suffer from the curse of dimensionality and become computationally intractable in high dimensions, more recent neural network-based approaches scale better, but have mostly been studied with the aim of solving optimal transport problems and require the solution of complicated optimization or max-min problems. Using an implicit Fenchel formulation of convex conjugation, our approach facilitates an efficient gradient-based framework for the minimization of approximation errors and, as a byproduct, also provides a posteriori error estimates for the approximation quality. Numerical experiments demonstrate our method's ability to deliver accurate results across different high-dimensional examples. Moreover, by employing symbolic regression with Kolmogorov–Arnold networks, it is able to obtain the exact convex conjugates of specific convex functions.

Computer Vision · Everything Else

Zeqin Yu, Haotao Xie, Jian Zhang, Jiangqun Ni, Wenkang Su, Jiwu Huang

Existing Text Image Forgery Localization (T-IFL) methods often suffer from poor generalization due to the limited scale of real-world datasets and the distribution gap caused by synthetic data that fails to capture the complexity of real-world tampering. To tackle this issue, we propose Fourier Series-based Tampering Synthesis (FSTS), a structured and interpretable framework for synthesizing tampered text images. FSTS first collects 16,750 real-world tampering instances from five representative tampering types, using a structured pipeline that records human-performed editing traces via multi-format logs (e.g., video, PSD, and editing logs). By analyzing these collected parameters and identifying recurring behavioral patterns at both individual and population levels, we formulate a hierarchical modeling framework. Specifically, each individual tampering parameter is represented as a compact combination of basis operation–parameter configurations, while the population-level distribution is constructed by aggregating these behaviors. Since this formulation draws inspiration from the Fourier series, it enables an interpretable approximation using basis functions and their learned weights. By sampling from this modeled distribution, FSTS synthesizes diverse and realistic training data that better reflect real-world forgery traces. Extensive experiments across four evaluation protocols demonstrate that models trained with FSTS data achieve significantly improved generalization on real-world datasets. Dataset is available at \href{https://github.com/ZeqinYu/FSTS}{Project Page}.

Applications · Physics

Rishal Aggarwal, Jacky Chen, Nicholas Boffi, David Koes

Efficient sampling from the Boltzmann distribution given its energy function is a key challenge for modeling complex physical systems such as molecules. Boltzmann Generators address this problem by leveraging continuous normalizing flows to transform a simple prior into a distribution that can be reweighted to match the target using sample likelihoods. Despite the elegance of this approach, obtaining these likelihoods requires computing costly Jacobians during integration, which is impractical for large molecular systems. To overcome this difficulty, we train an energy-based model (EBM) to approximate likelihoods using both noise contrastive estimation (NCE) and score matching, which we show outperforms the use of either objective in isolation. On 2D synthetic systems where failure can be easily visualized, NCE improves mode weighting relative to score matching alone. On alanine dipeptide, our method yields free energy profiles and energy distributions that closely match those obtained using exact likelihoods while achieving $100\times$ faster inference. By training on multiple dipeptide systems, we show that our approach also exhibits effective transfer learning, generalizing to new systems at inference time and achieving at least a $6\times$ speedup over standard MD with only a bit of fine-tuning. While many recent efforts in generative modeling have prioritized models with fast sampling, our work demonstrates the design of models with accelerated likelihoods, enabling the application of reweighting schemes that ensure unbiased Boltzmann statistics at scale. Our code is available at https://github.com/RishalAggarwal/BoltzNCE.

Theory · Learning Theory

Fanghui Liu, Leello Tadesse Dadi, Volkan Cevher

Recent studies show that a reproducing kernel Hilbert space (RKHS) is not a suitable space to model functions by neural networks as the curse of dimensionality (CoD) cannot be evaded when trying to approximate even a single ReLU neuron. In this paper, we study a suitable function space for over-parameterized two-layer neural networks with bounded norms (e.g., the path norm, the Barron norm) in the perspective of sample complexity and generalization properties. First, we show that the path norm (as well as the Barron norm) is able to obtain width-independence sample complexity bounds, which allows for uniform convergence guarantees. Based on this result, we derive the improved result of metric entropy for $\epsilon$-covering up to $O(\epsilon^{-\frac{2d}{d+2}})$ ($d$ is the input dimension and the depending constant is at most polynomial order of $d$) via the convex hull technique, which demonstrates the separation with kernel methods with $\Omega(\epsilon^{-d})$ to learn the target function in a Barron space. Second, this metric entropy result allows for building a sharper generalization bound under a general moment hypothesis setting, achieving the rate at $O(n^{-\frac{d+2}{2d+2}})$. Our analysis is novel in that it offers a sharper and refined estimation for metric entropy (with a clear dependence relationship on the dimension $d$) and unbounded sampling in the estimation of the sample error and the output error.

Theory · Learning Theory

Hongrui Chen, Jihao Long, Lei Wu

"We consider the problem of learning functions within the $\mathcal{F}_{p,\pi}$ and Barron spaces, which play crucial roles in understanding random feature models (RFMs), two-layer neural networks, as well as kernel methods. Leveraging tools from information-based complexity (IBC), we establish a dual equivalence between approximation and estimation, and then apply it to study the learning of the preceding function spaces. The duality allows us to focus on the more tractable problem between approximation and estimation. To showcase the efficacy of our duality framework, we delve into two important but under-explored problems: \begin{itemize} \item \textbf{Random feature learning beyond kernel regime:} We derive sharp bounds for learning $\cF_{p,\pi}$ using RFMs. Notably, the learning is efficient without the curse of dimensionality for $p>1$. This underscores the extended applicability of RFMs beyond the traditional kernel regime, since $\mathcal{F}_{p,\pi}$ with $p<2$ is strictly larger than the corresponding reproducing kernel Hilbert space (RKHS) where $p=2$. \item \textbf{The $L^\infty$ learning of RKHS:} We establish sharp, spectrum-dependent characterizations for the convergence of $L^\infty$ learning error in both noiseless and noisy settings. Surprisingly, we show that popular kernel ridge regression can achieve near-optimal performance in $L^\infty$ learning, despite it primarily minimizing square loss. \end{itemize} To establish the aforementioned duality, we introduce a type of IBC, termed $I$-complexity, to measure the size of a function class. Notably, $I$-complexity offers a tight characterization of learning in noiseless settings, yields lower bounds comparable to Le Cam's in noisy settings, and is versatile in deriving upper bounds. We believe that our duality framework holds potential for broad application in learning analysis across more scenarios."

Qixin Zhang, Yan Sun, Can Jin, Xikun Zhang, Yao SHU, Puning Zhao, Li Shen, Dacheng Tao

In this paper, we present two effective policy learning algorithms for multi-agent online coordination(MA-OC) problem. The first one, **MA-SPL**, not only can achieve the optimal $(1-\frac{c}{e})$-approximation guarantee for the MA-OC problem with submodular objectives but also can handle the unexplored $\alpha$-weakly DR-submodular and $(\gamma,\beta)$-weakly submodular scenarios, where $c$ is the curvature of the investigated submodular functions, $\alpha$ denotes the diminishing-return(DR) ratio and the tuple$(\gamma,\beta)$ represents the submodularity ratios. Subsequently, in order to reduce the reliance on the unknown parameters $\alpha,\gamma,\beta$ inherent in the **MA-SPL** algorithm, we then introduce the second online algorithm named **MA-MPL**. This **MA-MPL** algorithm is entirely *parameter-free* and simultaneously can maintain the same approximation ratio as the first **MA-SPL** algorithm. The core of our **MA-SPL** and **MA-MPL** algorithms is a novel continuous-relaxation technique term as policy-based continuous extension. Compared with the well-established multi-linear extension, a notable advantage of this new policy-based continuous extension is its ability to provide a lossless rounding scheme for any set function, thereby enabling us to tackle the challenging weakly submodular objective functions. Finally, extensive simulations are conducted to demonstrate the effectiveness of our proposed algorithms.

Reinforcement Learning · Everything Else

Hon Tik Tse, Siddarth Chandrasekar, Marlos C. Machado

In recent years, the successor representation (SR) has attracted increasing attention in reinforcement learning (RL), and it has been used to address some of its key challenges, such as exploration, credit assignment, and generalization. The SR can be seen as representing the underlying credit assignment structure of the environment by implicitly encoding its induced transition dynamics. However, the SR is reward-agnostic. In this paper, we discuss a similar representation that also takes into account the reward dynamics of the problem. We study the default representation (DR), a recently proposed representation with limited theoretical (and empirical) analysis. Here, we lay some of the theoretical foundation underlying the DR in the tabular case by (1) deriving dynamic programming and (2) temporal-difference methods to learn the DR, (3) characterizing the basis for the vector space of the DR, and (4) formally extending the DR to the function approximation case through default features. Empirically, we analyze the benefits of the DR in many of the settings in which the SR has been applied, including (1) reward shaping, (2) option discovery, (3) exploration, and (4) transfer learning. Our results show that, compared to the SR, the DR gives rise to qualitatively different, reward-aware behaviour and quantitatively better performance in several settings.

Reinforcement Learning · Everything Else

Shayan Karimi, Xiaoqi Tan

This paper investigates the computational complexity of reinforcement learning within a novel linear function approximation regime, termed partial $q^{\pi}$-realizability. In this framework, the objective is to learn an $\epsilon$-optimal policy with respect to a predefined policy set $\Pi$, under the assumption that all value functions corresponding to policies in $\Pi$ are linearly realizable. This framework adopts assumptions that are weaker than those in the $q^{\pi}$-realizability setting yet stronger than those in the q*-realizability setup. As a result, it provides a more practical model for reinforcement learning scenarios where function approximation naturally arise. We prove that learning an $\epsilon$-optimal policy in this newly defined setting is computationally hard. More specifically, we establish NP-hardness under a parameterized greedy policy set (i.e., argmax) and, further, show that—unless NP = RP—an exponential lower bound (exponential in feature vector dimension) holds when the policy set contains softmax policies, under the Randomized Exponential Time Hypothesis. Our hardness results mirror those obtained in the $q^*$-realizability settings, and suggest that computational difficulty persists even when the policy class $ \Pi $ is expanded beyond the optimal policy, reinforcing the unbreakable nature of the computational hardness result regarding partial $ q^{\pi} $-realizability under two important policy sets. To establish our negative result, our primary technical contribution is a reduction from two complexity problems, $\delta$-Max-3SAT and $\delta$-Max-3SAT($b$), to instances of our problem settings: GLinear-$\kappa$-RL (under the greedy policy set) and SLinear-$\kappa$-RL (under the softmax policy set), respectively. Our findings indicate that positive computational results are generally unattainable in the context of partial $ q^{\pi} $-realizability, in sharp contrast to the $ q^{\pi} $-realizability setting under a generative access model.

General Machine Learning · Transfer, Multitask and Meta-learning

Haomiao Qiu, Miao Zhang, Ziyue Qiao, Liqiang Nie

Continual Learning (CL) aims to enable models to continuously acquire new knowledge from a sequence of tasks with avoiding the forgetting of learned information. However, existing CL methods only rely on the parameters of the most recent task for inference, which makes them susceptible to catastrophic forgetting. Inspired by the recent success of model merging techniques, we propose Perturb-and-Merge (P\&M), a novel continual learning framework that integrates model merging into the CL paradigm to mitigate forgetting. Specifically, after training on each task, P\&M constructs a new model by forming a convex combination of the previous model and the newly trained task-specific model. Through theoretical analysis, We minimize the total loss increase across all tasks and derive a closed-form solution for the merging coefficient under mild assumptions. To further improve the performance of the merged model, we observe that the degradation introduced during merging can be alleviated by a regularization term composed of the task vector and the Hessian matrix of the loss function. Interestingly, we show that this term can be efficiently approximated using second-order symmetric finite differences, and a stochastic perturbation strategy along the task vector direction is accordingly devised which incurs no additional forward or backward passes while providing an effective approximation of the regularization term. Finally, we combine P\&M with LoRA, a parameter-efficient fine-tuning method, to reduce memory overhead. Our proposed approach achieves state-of-the-art performance on several continual learning benchmark datasets. The code is available at \url{https://github.com/qhmiao/P-M-for-Continual-Learning}.

Reinforcement Learning · Everything Else

Andre Barreto, Vincent Dumoulin, Yiran Mao, Mark Rowland, Nicolas Perez-Nieves, Bobak Shahriari, Yann Dauphin, Doina Precup, Hugo Larochelle

Reinforcement learning from human feedback usually models preferences using a reward function that does not distinguish between people. We argue that this is unlikely to be a good design choice in contexts with high potential for disagreement, like in the training of large language models. We formalise and analyse the problem of learning a reward model that can be specialised to a user. Using the principle of empirical risk minimisation, we derive a probably approximately correct (PAC) bound showing the dependency of the approximation error on the number of training examples, as usual, and also on the number of human raters who provided feedback on them. Based on our theoretical findings, we discuss how to best collect pairwise preference data and argue that adaptive reward models should be beneficial when there is considerable disagreement among users. We also propose a concrete architecture for an adaptive reward model. Our approach leverages the observation that individual preferences can be captured as a linear combination of a set of general reward features. We show how to learn such features and subsequently use them to quickly adapt the reward model to a specific individual, even if their preferences are not reflected in the training data. We present experiments with large language models illustrating our theoretical results and comparing the proposed architecture with a non-adaptive baseline. Consistent with our analysis, the benefits provided by our model increase with the number of raters and the heterogeneity of their preferences. We also show that our model compares favourably to adaptive counterparts, including those performing in-context personalisation.

General Machine Learning · Everything Else

Zixun Wang, Ben Dai

Semantic segmentation labels each pixel in an image with its corresponding class, and is typically evaluated using the Intersection over Union (IoU) and Dice metrics to quantify the overlap between predicted and ground-truth segmentation masks. In the literature, most existing methods estimate pixel-wise class probabilities, then apply argmax or thresholding to obtain the final prediction. These methods have been shown to generally lead to inconsistent or suboptimal results, as they do not directly maximize segmentation metrics. To address this issue, a novel consistent segmentation framework, RankSEG, has been proposed, which includes RankDice and RankIoU specifically designed to optimize the Dice and IoU metrics, respectively. Although RankSEG almost guarantees improved performance, it suffers from two major drawbacks. First, it is its computational expense—RankDice has a complexity of $\mathcal{O}(d \log d)$ with a substantial constant factor (where $d$ represents the number of pixels), while RankIoU exhibits even higher complexity $\mathcal{O}(d^2)$, thus limiting its practical application. For instance, in LiTS, prediction with RankSEG takes 16.33 seconds compared to just 0.01 seconds with the argmax rule. Second, RankSEG is only applicable to overlapping segmentation settings, where multiple classes can occupy the same pixel, which contrasts with standard benchmarks that typically assume non-overlapping segmentation. In this paper, we overcome these two drawbacks via a \textit{reciprocal moment approximation} (RMA) of RankSEG with the following contributions: (i) we improve RankSEG using RMA, namely RankSEG-RMA, reduces the complexity of both algorithms to $\mathcal{O}(d)$ while maintaining comparable performance; (ii) inspired by RMA, we develop a pixel-wise score function that allows efficient implementation for non-overlapping segmentation settings. We illustrate the effectiveness of our method across various datasets and state-of-the-art models. The code of our method is available in: \url{https://github.com/ZixunWang/RankSEG-RMA}.

Deep Learning · Theory

Jingpu Cheng, Ting Lin, Zuowei Shen, Qianxiao Li

We investigate the universal approximation property (UAP) of transformer-type architectures, providing a unified theoretical framework that extends prior results on residual networks to models incorporating attention mechanisms. Our work identifies token distinguishability as a fundamental requirement for UAP and introduces a general sufficient condition that applies to a broad class of architectures. Leveraging an analyticity assumption on the attention layer, we can significantly simplify the verification of this condition, providing a non-constructive approach in establishing UAP for such architectures. We demonstrate the applicability of our framework by proving UAP for transformers with various attention mechanisms, including kernel-based and sparse ones. The corollaries of our results either generalize prior works or establish UAP for architectures not previously covered. Furthermore, our framework offers a principled foundation for designing novel transformer architectures with inherent UAP guarantees, including those with specific functional symmetries. We propose examples to illustrate these insights.

Applications · Physics

Yuze Hao, Linchao Zhu, Yi Yang

Inverse design aims to design the input variables of a physical system to optimize a specified objective function, typically formulated as a search or optimization problem. However, in 3D domains, the design space grows exponentially, rendering exhaustive grid-based searches infeasible. Recent advances in deep learning have accelerated inverse design by providing powerful generative priors and differentiable surrogate models. Nevertheless, current methods tend to approximate the 3D design space using 2D projections or fine-tune existing 3D shapes. These approaches sacrifice volumetric detail and constrain design exploration, preventing true 3D design from scratch. In this paper, we propose a 3D Inverse Design (3DID) framework that directly navigates the 3D design space by coupling a continuous latent representation with a physics-aware optimization strategy. We first learn a unified physics–geometry embedding that compactly captures shape and physical field data in a continuous latent space. Then, we introduce a two-stage strategy to perform physics-aware optimization. In the first stage, a gradient-guided diffusion sampler explores the global latent manifold. In the second stage, an objectivedriven, topology-preserving refinement further sculpts each candidate toward the target objective. This enables 3DID to generate high-fidelity 3D geometries, outperforming existing methods in both solution quality and design versatility.

General Machine Learning · Representation Learning

Youqi WU, Jingwei Zhang, Farzan Farnia

State-of-the-art embeddings often capture distinct yet complementary discriminative features: For instance, one image embedding model may excel at distinguishing fine-grained textures, while another focuses on object-level structure. Motivated by this observation, we propose a principled approach to fuse such complementary representations through *kernel multiplication*. Multiplying the kernel similarity functions of two embeddings allows their discriminative structures to interact, producing a fused representation whose kernel encodes the union of the clusters identified by each parent embedding. This formulation also provides a natural way to construct *joint kernels* for paired multi-modal data (e.g., image–text tuples), where the product of modality-specific kernels inherits structure from both domains. We highlight that this kernel product is mathematically realized via the *Kronecker product* of the embedding feature maps, yielding our proposed *KrossFuse* framework for embedding fusion. To address the computational cost of the resulting high-dimensional Kronecker space, we further develop *RP-KrossFuse*, a scalable variant that leverages random projections for efficient approximation. As a key application, we use this framework to bridge the performance gap between cross-modal embeddings (e.g., CLIP, BLIP) and unimodal experts (e.g., DINOv2, E5). Experiments show that RP-KrossFuse effectively integrates these models, enhancing modality-specific performance while preserving cross-modal alignment. The project code is available at https://github.com/yokiwuuu/KrossFuse.

Zhanhao Xie, Yuexiao Ma, Xiawu Zheng, Fei Chao, Wanchen Sui, Yong Li, Shen Li, Rongrong Ji

The Mixture of Experts (MoE) architecture enables efficient model scaling through conditional computation, where only subset of parameters are activated per input. However, this distributed architecture poses unprecedented challenges for model compression, as conventional quantization methods optimized for dense networks prove inadequate. This paper introduces a specialized quantization framework for MoE architectures, motivated by our discovery that weight matrices across expert networks exhibit distinctive channel-wise outlier distributions, necessitating a more nuanced compression approach. Through theoretical analysis incorporating Fisher Information matrices and condition number characteristics, we establish a fundamental relationship between layer functionality and quantization sensitivity, demonstrating that down-projection layers inherently demand higher precision compared to up-projection layers. Leveraging these insights, we develop an automated channel-wise quantization framework that dynamically determines optimal bit-width allocations while maintaining minimal computational overhead through efficient statistical approximations. When evaluated on the Mixtral-8x7b-v0.1 architecture, our methodology demonstrates a 3.96% improvement over existing state-of-the-art approaches across natural language understanding benchmarks, while achieving superior compression ratios.

Xuemiao Zhang, Xu Liangyu, Feiyu Duan, Yongwei Zhou, Sirui Wang, Rongxiang Weng, Jingang Wang, Xunliang Cai

Large language models (LLMs) generally utilize a consistent data distribution throughout the pretraining process. However, as the model’s capability improves, it is intuitive that its data preferences dynamically change, indicating the need for pretraining with different data at various training stages. To achieve it, we propose the Perplexity Difference (PD) based Preference Curriculum learning (PDPC) framework, which always perceives and uses the data preferred by LLMs to train and boost them. First, we introduce the PD metric to quantify the difference in how challenging a sample is for weak versus strong models. Samples with high PD are more challenging for weak models to learn and are more suitable to be arranged in the later stage of pretraining. Second, we propose the preference function to approximate and predict the data preference of the LLM at any training step, so as to complete the arrangement of the dataset offline and ensure continuous training without interruption. Experimental results on 1.3B and 3B models demonstrate that PDPC significantly surpasses baselines. Notably, the 3B model trained on 1T tokens achieves an increased average accuracy of over 8.1% across MMLU and CMMLU.

Alessio Galatolo, Zhenbang Dai, Katie Winkle, Meriem Beloucif

Fine-tuning Large Language Models (LLMs) with first-order methods like back-propagation is computationally intensive. Zeroth-Order (ZO) optimisation uses function evaluations instead of gradients, reducing memory usage, but suffers from slow convergence in high-dimensional models. As a result, ZO research in LLMs has mostly focused on classification, overlooking more complex generative tasks. In this paper, we introduce ZOPrO, a novel ZO algorithm designed for *Preference Optimisation* in LLMs. We begin by analysing the interplay between policy and reward models during traditional (first-order) Preference Optimisation, uncovering patterns in their relative updates. Guided by these insights, we adapt Simultaneous Perturbation Stochastic Approximation (SPSA) with a targeted sampling strategy to accelerate convergence. Through experiments on summarisation, machine translation, and conversational assistants, we demonstrate that our method consistently enhances reward signals while achieving convergence times comparable to first-order methods. While it falls short of some state-of-the-art methods, our work is the first to apply Zeroth-Order methods to Preference Optimisation in LLMs, going beyond classification tasks and paving the way for a largely unexplored research direction. Code and visualisations are available at https://github.com/alessioGalatolo/VisZOPrO.