ICLR 2025 论文列表

This paper systematically explores neural functional networks (NFN) for transformer architectures. NFN are specialized neural networks that treat the weights, gradients, or sparsity patterns of a deep neural network (DNN) as input data and have proven valuable for tasks such as learnable optimizers, implicit data representations, and weight editing. While NFN have been extensively developed for MLP and CNN, no prior work has addressed their design for transformers, despite the importance of transformers in modern deep learning. This paper aims to address this gap by providing a systematic study of NFN for transformers. We first determine the maximal symmetric group of the weights in a multi-head attention module as well as a necessary and sufficient condition under which two sets of hyperparameters of the multi-head attention module define the same function. We then define the weight space of transformer architectures and its associated group action, which leads to the design principles for NFN in transformers. Based on these, we introduce Transformer-NFN, an NFN that is equivariant under this group action. Additionally, we release a dataset of more than 125,000 Transformers model checkpoints trained on two datasets with two different tasks, providing a benchmark for evaluating Transformer-NFN and encouraging further research on transformer training and performance.

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning. Our results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.

Harmful fine-tuning attack poses serious safety concerns for large language models' fine-tuning-as-a-service. While existing defenses have been proposed to mitigate the issue, their performances are still far away from satisfactory, and the root cause of the problem has not been fully recovered. To this end, we in this paper show that \textit{harmful perturbation} over the model weights could be a probable cause of alignment-broken. In order to attenuate the negative impact of harmful perturbation, we propose an alignment-stage solution, dubbed Booster. Technically, along with the original alignment loss, we append a loss regularizer in the alignment stage's optimization. The regularizer ensures that the model's harmful loss reduction after the simulated harmful perturbation is attenuated, thereby mitigating the subsequent fine-tuning risk. Empirical results show that Booster can effectively reduce the harmful score of the fine-tuned models while maintaining the performance of downstream tasks. Our code is available at https://github.com/git-disl/Booster

Knowledge distillation (KD) is widely used to train small, high-performing student language models (LMs) using large teacher LMs. While effective in fine-tuning, KD during pre-training faces efficiency, flexibility, and effectiveness issues. Existing methods either incur high computational costs due to online teacher inference, require tokenization matching between teacher and student LMs, or risk losing the difficulty and diversity of the teacher-generated training data. In this work, we propose **MiniPLM**, a KD framework for pre-training LMs by refining the training data distribution with the teacher LM's knowledge. For efficiency, MiniPLM performs offline teacher inference, allowing KD for multiple student LMs without adding training costs. For flexibility, MiniPLM operates solely on the training corpus, enabling KD across model families. For effectiveness, MiniPLM leverages the differences between large and small LMs to enhance the training data difficulty and diversity, helping student LMs acquire versatile and sophisticated knowledge. Extensive experiments demonstrate that MiniPLM boosts the student LMs' performance on 9 common downstream tasks, improves language modeling capabilities, and reduces pre-training computation. The benefit of MiniPLM extends to larger training scales, evidenced by the scaling curve extrapolation. Further analysis reveals that MiniPLM supports KD across model families and enhances the pre-training data utilization. Our code, data, and models can be found at https://github.com/thu-coai/MiniPLM.

Transformer models have been widely investigated in different domains by providing long-range dependency handling and global contextual awareness, driving the development of popular AI applications such as ChatGPT, Gemini, and Alexa. State Space Models (SSMs) have emerged as strong contenders in the field of sequential modeling, challenging the dominance of Transformers. SSMs incorporate a selective mechanism that allows for dynamic parameter adjustment based on input data, enhancing their performance. However, this mechanism also comes with increasing computational complexity and bandwidth demands, posing challenges for deployment on resource-constraint mobile devices. To address these challenges without sacrificing the accuracy of the selective mechanism, we propose a sparse learning framework that integrates architecture-aware compiler optimizations. We introduce an end-to-end solution--$\mathbf{C}_4^n$ kernel sparsity, which prunes $n$ elements from every four contiguous weights, and develop a compiler-based acceleration solution to ensure execution efficiency for this sparsity on mobile devices. Based on the kernel sparsity, our framework generates optimized sparse models targeting specific sparsity or latency requirements for various model sizes. We further leverage pruned weights to compensate for the remaining weights, enhancing downstream task performance. For practical hardware acceleration, we propose $\mathbf{C}_4^n$-specific optimizations combined with a layout transformation elimination strategy. This approach mitigates inefficiencies arising from fine-grained pruning in linear layers and improves performance across other operations. Experimental results demonstrate that our method achieves superior task performance compared to other semi-structured pruning methods and achieves up-to 7$\times$ speedup compared to llama.cpp framework on mobile devices.

Recent advances in style and appearance transfer are impressive, but most methods isolate global style and local appearance transfer, neglecting semantic correspondence. Additionally, image and video tasks are typically handled in isolation, with little focus on integrating them for video transfer. To address these limitations, we introduce a novel task, *Semantic Style Transfer*, which involves transferring style and appearance features from a reference image to a target visual content based on semantic correspondence. We subsequently propose a training-free method, *Semantix*, an energy-guided sampler designed for Semantic Style Transfer that simultaneously guides both style and appearance transfer based on semantic understanding capacity of pre-trained diffusion models. Additionally, as a sampler, *Semantix* can be seamlessly applied to both image and video models, enabling semantic style transfer to be generic across various visual media. Specifically, once inverting both reference and context images or videos to noise space by SDEs, *Semantix* utilizes a meticulously crafted energy function to guide the sampling process, including three key components: *Style Feature Guidance*, *Spatial Feature Guidance* and *Semantic Distance* as a regularisation term. Experimental results demonstrate that *Semantix* not only effectively accomplishes the task of semantic style transfer across images and videos, but also surpasses existing state-of-the-art solutions in both fields.

Large Language Models (LLMs) have shown immense potential in enhancing various aspects of our daily lives, from conversational AI to search and AI assistants. However, their growing capabilities come at the cost of extremely large model sizes, making deployment on edge devices challenging due to memory and computational constraints. This paper introduces a novel approach to LLM weight pruning that directly optimizes for approximating the attention matrix, a core component of transformer architectures. Unlike existing methods that focus on linear approximations, our approach accounts for the non-linear nature of the Softmax attention mechanism. We provide theoretical guarantees for the convergence of our Gradient Descent-based optimization method to a near-optimal pruning mask solution. Our empirical results demonstrate the effectiveness of our non-linear pruning approach in maintaining model performance while significantly reducing computational costs, which is beyond the current state-of-the-art methods, i.e., SparseGPT and Wanda, by a large margin. This work establishes a new theoretical foundation for pruning algorithm design in LLMs, potentially paving the way for more efficient LLM inference on resource-constrained devices.

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.

The dominant paradigm for learning on graphs is message passing. Despite being a strong inductive bias, the local message passing mechanism faces challenges such as over-smoothing, over-squashing, and limited expressivity. To address these issues, we introduce Bundle Neural Networks (BuNNs), a novel graph neural network architecture that operates via *message diffusion* on *flat vector bundles* — geometrically inspired structures that assign to each node a vector space and an orthogonal map. A BuNN layer evolves node features through a diffusion-type partial differential equation, where its discrete form acts as a special case of the recently introduced Sheaf Neural Network (SNN), effectively alleviating over-smoothing. The continuous nature of message diffusion enables BuNNs to operate at larger scales, reducing over-squashing. We establish the universality of BuNNs in approximating feature transformations on infinite families of graphs with injective positional encodings, marking the first positive expressivity result of its kind. We support our claims with formal analysis and synthetic experiments. Empirically, BuNNs perform strongly on heterophilic and long-range tasks, which demonstrates their robustness on a diverse range of challenging real-world tasks.

We provide finite-particle convergence rates for the Stein Variational Gradient Descent (SVGD) algorithm in the Kernelized Stein Discrepancy ($\KSD$) and Wasserstein-2 metrics. Our key insight is that the time derivative of the relative entropy between the joint density of $N$ particle locations and the $N$-fold product target measure, starting from a regular initial distribution, splits into a dominant 'negative part' proportional to $N$ times the expected $\KSD^2$ and a smaller 'positive part'. This observation leads to $\KSD$ rates of order $1/\sqrt{N}$, in both continuous and discrete time, providing a near optimal (in the sense of matching the corresponding i.i.d. rates) double exponential improvement over the recent result by~\cite{shi2024finite}. Under mild assumptions on the kernel and potential, these bounds also grow polynomially in the dimension $d$. By adding a bilinear component to the kernel, the above approach is used to further obtain Wasserstein-2 convergence in continuous time. For the case of `bilinear + Mat\'ern' kernels, we derive Wasserstein-2 rates that exhibit a curse-of-dimensionality similar to the i.i.d. setting. We also obtain marginal convergence and long-time propagation of chaos results for the time-averaged particle laws.

The burgeoning capabilities of large language models (LLMs) have underscored the need for alignment to ensure these models act in accordance with human values and intentions. Existing alignment frameworks present constraints either in the form of expensive human effort or high computational costs. This paper explores a promising middle ground, where we employ a weak LLM that is significantly less resource-intensive than top-tier models, yet offers more automation than purely human feedback. We present a systematic study to evaluate and understand weak LLM's ability to generate feedback for alignment. Our empirical findings demonstrate that weak LLMs can provide feedback that rivals or even exceeds that of fully human-annotated data. Our study indicates a minimized impact of model size on feedback efficacy, shedding light on a scalable and sustainable alignment strategy. To deepen our understanding of alignment under weak LLM feedback, we conduct a series of qualitative and quantitative analyses, offering novel insights into the quality discrepancies between human feedback vs. weak LLM feedback. Code is publicly available at https://github.com/deeplearning-wisc/weak_llm_teacher.

Positional encodings (PEs) are essential for building powerful and expressive graph neural networks and graph transformers, as they effectively capture the relative spatial relationships between nodes. Although extensive research has been devoted to PEs in undirected graphs, PEs for directed graphs remain relatively unexplored. This work seeks to address this gap. We first introduce the notion of *Walk Profile*, a generalization of walk-counting sequences for directed graphs. A walk profile encompasses numerous structural features crucial for directed graph-relevant applications, such as program analysis and circuit performance prediction. We identify the limitations of existing PE methods in representing walk profiles and propose a novel *Multi-q Magnetic Laplacian PE*, which extends the Magnetic Laplacian eigenvector-based PE by incorporating multiple potential factors. The new PE can provably express walk profiles. Furthermore, we generalize prior basis-invariant neural networks to enable the stable use of the new PE in the complex domain. Our numerical experiments validate the expressiveness of the proposed PEs and demonstrate their effectiveness in solving sorting network satisfiability and performing well on general circuit benchmarks. Our code is available at https://github.com/Graph-COM/Multi-q-Maglap.

Existing benchmarks for language agents do not set them up to interact with human users or follow domain-specific rules, both of which are vital to safe and realistic deployment. We propose $\tau$-bench, a benchmark with two domains (retail and airline) emulating dynamic conversations between a user (simulated by language models) and a customer service agent provided with domain-specific API tools and policy guidelines. We employ a efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state, and propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (gpt-4o) succeed on $<50\%$ of the tasks, and are terribly inconsistent (pass^8 < 25\% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and reliably follow rules.

We study the optimal batch-regret tradeoff for batch linear contextual bandits. For this problem, we design batch learning algorithms and prove that they achieve the optimal regret bounds (up to logarithmic factors) for any batch number $M$, number of actions $K$, time horizon $T$, and dimension $d$. Therefore, we establish the \emph{full-parameter-range} (almost) optimal batch-regret tradeoff for the batch linear contextual bandit problem. Along our analysis, we also prove a new matrix concentration inequality with dependence on their dynamic upper bounds, which, to the best of our knowledge, is the first of its kind in literature and maybe of independent interest.

Diffusion priors have been used for blind face restoration (BFR) by fine-tuning diffusion models (DMs) on restoration datasets to recover low-quality images. However, the naive application of DMs presents several key limitations. (i) The diffusion prior has inferior semantic consistency (e.g., ID, structure and color.), increasing the difficulty of optimizing the BFR model; (ii) reliance on hundreds of denoising iterations, preventing the effective cooperation with perceptual losses, which is crucial for faithful restoration. Observing that the latent consistency model (LCM) learns consistency noise-to-data mappings on the ODE-trajectory and therefore shows more semantic consistency in the subject identity, structural information and color preservation, we propose $\textit{InterLCM}$ to leverage the LCM for its superior semantic consistency and efficiency to counter the above issues. Treating low-quality images as the intermediate state of LCM, $\textit{InterLCM}$ achieves a balance between fidelity and quality by starting from earlier LCM steps. LCM also allows the integration of perceptual loss during training, leading to improved restoration quality, particularly in real-world scenarios. To mitigate structural and semantic uncertainties, $\textit{InterLCM}$ incorporates a Visual Module to extract visual features and a Spatial Encoder to capture spatial details, enhancing the fidelity of restored images. Extensive experiments demonstrate that $\textit{InterLCM}$ outperforms existing approaches in both synthetic and real-world datasets while also achieving faster inference speed. Code and models will be publicly available.

Spiking Neural Networks (SNNs) offer a promising, biologically inspired approach for processing spatiotemporal data, particularly for time series forecasting. However, conventional neuron models like the Leaky Integrate-and-Fire (LIF) struggle to capture long-term dependencies and effectively process multi-scale temporal dynamics. To overcome these limitations, we introduce the Temporal Segment Leaky Integrate-and-Fire (TS-LIF) model, featuring a novel dual-compartment architecture. The dendritic and somatic compartments specialize in capturing distinct frequency components, providing functional heterogeneity that enhances the neuron's ability to process both low- and high-frequency information. Furthermore, the newly introduced direct somatic current injection reduces information loss during intra-neuronal transmission, while dendritic spike generation improves multi-scale information extraction. We provide a theoretical stability analysis of the TS-LIF model and explain how each compartment contributes to distinct frequency response characteristics. Experimental results show that TS-LIF outperforms traditional SNNs in time series forecasting, demonstrating better accuracy and robustness, even with missing data. TS-LIF advances the application of SNNs in time-series forecasting, providing a biologically inspired approach that captures complex temporal dynamics and offers potential for practical implementation in diverse forecasting scenarios.

Understanding world dynamics is crucial for planning in autonomous driving. Recent methods attempt to achieve this by learning a 3D occupancy world model that forecasts future surrounding scenes based on current observation. However, 3D occupancy labels are still required to produce promising results. Considering the high annotation cost for 3D outdoor scenes, we propose a semi-supervised vision-centric 3D occupancy world model, **PreWorld**, to leverage the potential of 2D labels through a novel two-stage training paradigm: the self-supervised pre-training stage and the fully-supervised fine-tuning stage. Specifically, during the pre-training stage, we utilize an attribute projection head to generate different attribute fields of a scene (e.g., RGB, density, semantic), thus enabling temporal supervision from 2D labels via volume rendering techniques. Furthermore, we introduce a simple yet effective state-conditioned forecasting module to recursively forecast future occupancy and ego trajectory in a direct manner. Extensive experiments on the nuScenes dataset validate the effectiveness and scalability of our method, and demonstrate that PreWorld achieves competitive performance across 3D occupancy prediction, 4D occupancy forecasting and motion planning tasks.

Recently, data-driven simulators based on graph neural networks have gained attention in modeling physical systems on unstructured meshes. However, they struggle with long-range dependencies in fluid flows, particularly in refined mesh regions. This challenge, known as the 'over-squashing' problem, hinders information propagation. While existing graph rewiring methods address this issue to some extent, they only consider graph topology, overlooking the underlying physical phenomena. We propose Physics-Informed Ollivier--Ricci Flow (PIORF), a novel rewiring method that combines physical correlations with graph topology. PIORF uses Ollivier--Ricci curvature (ORC) to identify bottleneck regions and connects these areas with nodes in high-velocity gradient nodes, enabling long-range interactions and mitigating over-squashing. Our approach is computationally efficient in rewiring edges and can scale to larger simulations. Experimental results on 3 fluid dynamics benchmark datasets show that PIORF consistently outperforms baseline models and existing rewiring methods, achieving up to 26.2\% improvement.

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.