ICLR 2025 Papers

Various parameter-efficient fine-tuning (PEFT) methods, including LoRA and its variants, have gained popularity for reducing computational costs. However, there is often an accuracy gap between PEFT approaches and full fine-tuning (FT), and this discrepancy has not yet been systematically explored. In this work, we introduce a method for selecting sparse sub-matrices that aims to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational costs and memory costs. We explored both gradient-based and activation-based parameter selection methods to identify the most significant sub-matrices for downstream tasks, updating only these blocks during fine-tuning. In our experiments, we demonstrated that SMT consistently surpasses other PEFT baselines (e.g., LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67% compared to FT. We also examine how the performance of LoRA and DoRA tends to plateau and decline as the number of trainable parameters increases, in contrast, our SMT method does not suffer from such issues.

As advancements in large language models (LLMs) continue and the demand for personalized models increases, parameter-efficient fine-tuning (PEFT) methods (e.g., LoRA) become essential due to their efficiency in reducing computation costs. However, recent studies have raised alarming concerns that LoRA fine-tuning could potentially compromise the safety alignment in LLMs, posing significant risks for the model owner. In this paper, we first investigate the underlying mechanism by analyzing the changes in safety alignment related features before and after fine-tuning. Then, we propose a fixed safety module calculated by safety data and a task-specific initialization for trainable parameters in low-rank adaptations, termed Safety-alignment preserved Low-Rank Adaptation (SaLoRA). Unlike previous LoRA methods and their variants, SaLoRA enables targeted modifications to LLMs without disrupting their original alignments. Our experiments show that SaLoRA outperforms various adapters-based approaches across various evaluation metrics in different fine-tuning tasks.

Humans excel at reusing prior knowledge to address new challenges and developing skills while solving problems. This paradigm becomes increasingly popular in the development of autonomous agents, as it develops systems that can self-evolve in response to new challenges like human beings. However, previous methods suffer from limited training efficiency when expanding new skills and fail to fully leverage prior knowledge to facilitate new task learning. We propose Parametric Skill Expansion and Composition (PSEC), a new framework designed to iteratively evolve the agents' capabilities and efficiently address new challenges by maintaining a manageable skill library. This library can progressively integrate skill primitives as plug-and-play Low-Rank Adaptation (LoRA) modules in parameter-efficient finetuning, facilitating efficient and flexible skill expansion. This structure also enables the direct skill compositions in parameter space by merging LoRA modules that encode different skills, leveraging shared information across skills to effectively program new skills. Based on this, we propose a context-aware modular to dynamically activate different skills to collaboratively handle new tasks. Empowering diverse applications including multi-objective composition, dynamics shift, and continual policy shift, the results on D4RL, DSRL benchmarks, and the DeepMind Control Suite show that PSEC exhibits superior capacity to leverage prior knowledge to efficiently tackle new challenges, as well as expand its skill libraries to evolve the capabilities. Project website: https://ltlhuuu.github.io/PSEC/.

Given a matrix $A \in \mathbb{R}^{m\times d}$ with singular values $\sigma_1\geq \cdots \geq \sigma_d$, and a random matrix $G \in \mathbb{R}^{m\times d}$ with iid $N(0,T)$ entries for some $T>0$, we derive new bounds on the Frobenius distance between subspaces spanned by the top-$k$ (right) singular vectors of $A$ and $A+G$. This problem arises in numerous applications in statistics where a data matrix may be corrupted by Gaussian noise, and in the analysis of the Gaussian mechanism in differential privacy, where Gaussian noise is added to data to preserve private information. We show that, for matrices $A$ where the gaps in the top-$k$ singular values are roughly $\Omega(\sigma_k-\sigma_{k+1})$ the expected Frobenius distance between the subspaces is $\tilde{O}(\frac{\sqrt{d}}{\sigma_k-\sigma_{k+1}} \times \sqrt{T})$, improving on previous bounds by a factor of $\frac{\sqrt{m}}{\sqrt{d}}$. To obtain our bounds we view the perturbation to the singular vectors as a diffusion process-- the Dyson-Bessel process-- and use tools from stochastic calculus to track the evolution of the subspace spanned by the top-$k$ singular vectors, which may be of independent interest.

Fine-tuning large language models (LLMs) is crucial for task specialization but often becomes resource-intensive due to redundant or uninformative data. Existing data selection methods typically rely either on computationally expensive gradient-based metrics or static embeddings that fail to adapt dynamically to the model’s evolving state, thus limiting their practical effectiveness. To address this, we propose DELIFT (Data Efficient Language model Instruction Fine-Tuning), leveraging a novel, computationally efficient utility metric inspired by In-Context Learning (ICL). Our ICL-based metric measures the informational value of each data sample by quantifying its effectiveness as an in-context example in improving model predictions for other samples, reflecting its actual contribution relative to the model’s current state. Integrated with tailored submodular optimization methods, DELIFT systematically selects diverse, informative subsets optimized specifically for each fine-tuning stage: instruction tuning, task-specific adaptation, and continual fine-tuning. Experimental results across multiple datasets and model scales show DELIFT reduces fine-tuning data requirements by up to 70% without compromising performance, consistently outperforming existing methods by up to 26% in effectiveness and efficiency.

Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to an unlabeled target domain with no access to the source data. Inspired by the success of large Vision-Language (ViL) models in many applications, the latest research has validated ViL's benefit for SFDA by using their predictions as pseudo supervision. However, we observe that ViL's supervision could be noisy and inaccurate at an unknown rate, potentially introducing additional negative effects during adaption. To address this thus-far ignored challenge, we introduce a novel Proxy Denoising (__ProDe__) approach. The key idea is to leverage the ViL model as a proxy to facilitate the adaptation process towards the latent domain-invariant space. Concretely, we design a proxy denoising mechanism to correct ViL's predictions. This is grounded on a proxy confidence theory that models the dynamic effect of proxy's divergence against the domain-invariant space during adaptation. To capitalize the corrected proxy, we further derive a mutual knowledge distilling regularization. Extensive experiments show that ProDe significantly outperforms the current state-of-the-art alternatives under both conventional closed-set setting and the more challenging open-set, partial-set, generalized SFDA, multi-target, multi-source, and test-time settings. Our code and data are available at https://github.com/tntek/source-free-domain-adaptation.

While large language models have demonstrated impressive performance across various domains and tasks, their security issues have become increasingly severe. Machine unlearning has emerged as a representative approach for model safety and security by removing the influence of undesired data on the target model. However, these methods do not sufficiently consider that unlearning requests in real-world scenarios are continuously emerging, especially in the context of LLMs, which may lead to accumulated model utility loss that eventually becomes unacceptable. Moreover, existing LLM unlearning methods often ignore previous data access limitations due to privacy concerns and copyright protection. Without previous data, the utility preservation during unlearning is much harder. To overcome these challenges, we propose the \OOO{} framework that includes an \underline{\textit{O}}rthogonal low-rank adapter (LoRA) for continually unlearning requested data and an \underline{\textit{O}}ut-\underline{\textit{O}}f-Distribution (OOD) detector to measure the similarity between input and unlearning data. The orthogonal LoRA achieves parameter disentanglement among continual unlearning requests. The OOD detector is trained with a novel contrastive entropy loss and utilizes a glocal-aware scoring mechanism. During inference, our \OOO{} framework can decide whether and to what extent to load the unlearning LoRA based on the OOD detector's predicted similarity between the input and the unlearned knowledge. Notably, \OOO{}'s effectiveness does not rely on any retained data. We conducted extensive experiments on \OOO{} and state-of-the-art LLM unlearning methods across three tasks and seven datasets. The results indicate that \OOO{} consistently achieves the best unlearning effectiveness and utility preservation, especially when facing continuous unlearning requests. The source codes can be found at \url{https://github.com/GCYZSL/O3-LLM-UNLEARNING}.

Powerful as they are, graph neural networks (GNNs) are known to be vulnerable to distribution shifts. Recently, test-time adaptation (TTA) has attracted attention due to its ability to adapt a pre-trained model to a target domain, without re-accessing the source domain. However, existing TTA algorithms are primarily designed for attribute shifts in vision tasks, where samples are independent. These methods perform poorly on graph data that experience structure shifts, where node connectivity differs between source and target graphs. We attribute this performance gap to the distinct impact of node attribute shifts versus graph structure shifts: the latter significantly degrades the quality of node representations and blurs the boundaries between different node categories. To address structure shifts in graphs, we propose Matcha, an innovative framework designed for effective and efficient adaptation to structure shifts by adjusting the htop-aggregation parameters in GNNs. To enhance the representation quality, we design a prediction-informed clustering loss to encourage the formation of distinct clusters for different node categories. Additionally, Matcha seamlessly integrates with existing TTA algorithms, allowing it to handle attribute shifts effectively while improving overall performance under combined structure and attribute shifts. We validate the effectiveness of Matcha on both synthetic and real-world datasets, demonstrating its robustness across various combinations of structure and attribute shifts. Our code is available at https://github.com/baowenxuan/Matcha.

Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning.

Efficient fine-tuning methods are critical to address the high computational and parameter complexity while adapting large pre-trained models to downstream tasks. Our study is inspired by prior research that represents each convolution filter as a linear combination of a small set of filter subspace elements, referred to as filter atoms. In this paper, we propose to fine-tune pre-trained models by adjusting only filter atoms, which are responsible for spatial-only convolution, while preserving spatially-invariant channel combination knowledge in atom coefficients. In this way, we bring a new filter subspace view for model tuning. Furthermore, each filter atom can be recursively decomposed as a combination of another set of atoms, which naturally expands the number of tunable parameters in the filter subspace. By only adapting filter atoms constructed by a small number of parameters, while maintaining the rest of model parameters constant, the proposed approach is highly parameter-efficient. It effectively preserves the capabilities of pre-trained models and prevents overfitting to downstream tasks. Extensive experiments show that such a simple scheme surpasses previous tuning baselines for both discriminate and generative tasks.

Neural networks are often challenging to work with due to their large size and complexity. To address this, various methods aim to reduce model size by sparsifying or decomposing weight matrices, such as magnitude pruning and low-rank or block-diagonal factorization. In this work, we present Double Sparse Factorization (DSF), where we factorize each weight matrix into two sparse matrices. Although solving this problem exactly is computationally infeasible, we propose an efficient heuristic based on alternating minimization via ADMM that achieves state-of-the-art results, enabling unprecedented sparsification of neural networks. For instance, in a one-shot pruning setting, our method can reduce the size of the LLaMA2-13B model by 50% while maintaining better performance than the dense LLaMA2-7B model. We also compare favorably with Optimal Brain Compression, the state-of-the-art layer-wise pruning approach for convolutional neural networks. Furthermore, accuracy improvements of our method persist even after further model fine-tuning. Code available at: https://github.com/usamec/double_sparse

An increasing interest in learning to forecast for time-series of high-dimensional observations is the ability to adapt to systems with diverse underlying dynamics. Access to observations that define a stationary distribution of these systems is often unattainable, as the underlying dynamics may change over time. Naively training or retraining models at each shift may lead to catastrophic forgetting about previously-seen systems. We present a new continual meta-learning (CML) framework to realize continual slow-and fast adaptation of latent dynamics (CoSFan). We leverage a feed-forward meta-model to infer *what* the current system is and *how* to adapt a latent dynamics function to it, enabling *fast adaptation* to specific dynamics. We then develop novel strategies to automatically detect *when* a shift of data distribution occurs, with which to identify its underlying dynamics and its relation with previously-seen dynamics. In combination with fixed-memory experience replay mechanisms, this enables continual *slow update* of the *what-how* meta-model. Empirical studies demonstrated that both the meta- and continual-learning component was critical for learning to forecast across non-stationary distributions of diverse dynamics systems, and the feed-forward meta-model combined with task-aware/-relational continual learning strategies significantly outperformed existing CML alternatives.

The recent paradigm shift to large-scale foundation models has brought about a new era for deep learning that, while has found great success in practice, has also been plagued by prohibitively expensive costs in terms of high memory consumption and compute. To mitigate these issues, there has been a concerted effort in post-hoc neural network pruning techniques that do not require costly retraining. Despite the considerable progress being made, existing methods often exhibit a steady drop in model performance as the compression increases. In this paper, we present a novel approach to compressing large transformers, coined OATS, that compresses the model weights by approximating each weight matrix as the sum of a sparse matrix and a low-rank matrix. Prior to the decomposition, the weights are first scaled by the second moment of their input embeddings, so as to ensure the preservation of outlier features recently observed in large transformer models. Without retraining, OATS achieves state-of-the-art performance when compressing large language models, such as Llama-3 and Phi-3, and vision transformers, such as Google's ViT and DINOv2, by up to $60\\%$, all while speeding up the model's inference on a CPU by up to $1.37\times$ compared to prior pruning methods.

We consider the problem of *low probability estimation*: given a machine learning model and a formally-specified input distribution, how can we estimate the probability of a binary property of the model's output, even when that probability is too small to estimate by random sampling? This problem is motivated by the need to improve worst-case performance, which distribution shift can make much more likely. We study low probability estimation in the context of argmax sampling from small transformer language models. We compare two types of methods: importance sampling, which involves searching for inputs giving rise to the rare output, and activation extrapolation, which involves extrapolating a probability distribution fit to the model's logits. We find that importance sampling outperforms activation extrapolation, but both outperform naive sampling. Finally, we explain how minimizing the probability estimate of an undesirable behavior generalizes adversarial training, and argue that new methods for low probability estimation are needed to provide stronger guarantees about worst-case performance.

Vision-language models (VLMs) have revolutionized machine learning by leveraging large pre-trained models to tackle various downstream tasks. Although label, training, and data efficiency have improved, many state-of-the-art VLMs still require task-specific hyperparameter tuning and fail to fully exploit test samples. To overcome these challenges, we propose a graph-based approach for label-efficient adaptation and inference. Our method dynamically constructs a graph over text prompts, few-shot examples, and test samples, using label propagation for inference without task-specific tuning. Unlike existing zero-shot label propagation techniques, our approach requires no additional unlabeled support set and effectively leverages the test sample manifold through dynamic graph expansion. We further introduce a context-aware feature re-weighting mechanism to improve task adaptation accuracy. Additionally, our method supports efficient graph expansion, enabling real-time inductive inference. Extensive evaluations on downstream tasks, such as fine-grained categorization and out-of-distribution generalization, demonstrate the effectiveness of our approach. The source code is available at https://github.com/Yushu-Li/ECALP.

The non-stationary nature of electroencephalography (EEG) introduces distribution shifts across domains (e.g., days and subjects), posing a significant challenge to EEG-based neurotechnology generalization. Without labeled calibration data for target domains, the problem is a source-free unsupervised domain adaptation (SFUDA) problem. For scenarios with constant label distribution, Riemannian geometry-aware statistical alignment frameworks on the symmetric positive definite (SPD) manifold are considered state-of-the-art. However, many practical scenarios, including EEG-based sleep staging, exhibit label shifts. Here, we propose a geometric deep learning framework for SFUDA problems under specific distribution shifts, including label shifts. We introduce a novel, realistic generative model and show that prior Riemannian statistical alignment methods on the SPD manifold can compensate for specific marginal and conditional distribution shifts but hurt generalization under label shifts. As a remedy, we propose a parameter-efficient manifold optimization strategy termed SPDIM. SPDIM uses the information maximization principle to learn a single SPD-manifold-constrained parameter per target domain. In simulations, we demonstrate that SPDIM can compensate for the shifts under our generative model. Moreover, using public EEG-based brain-computer interface and sleep staging datasets, we show that SPDIM outperforms prior approaches.

Utilitarian algorithm configuration is a general-purpose technique for automatically searching the parameter space of a given algorithm to optimize its performance, as measured by a given utility function, on a given set of inputs. Recently introduced utilitarian configuration procedures offer optimality guarantees about the returned parameterization while provably adapting to the hardness of the underlying problem. However, the applicability of these approaches is severely limited by the fact that they only search a finite, relatively small set of parameters. They cannot effectively search the configuration space of algorithms with continuous or uncountable parameters. In this paper we introduce a new procedure, which we dub COUP (Continuous, Optimistic Utilitarian Procrastination). COUP is designed to search infinite parameter spaces efficiently to find good configurations quickly. Furthermore, COUP maintains the theoretical benefits of previous utilitarian configuration procedures when applied to finite parameter spaces but is significantly faster, both provably and experimentally.

The degree-corrected block model (DCBM), latent space model (LSM), and $\beta$-model are all popular network models. We combine their modeling ideas and propose the logit-DCBM as a new model. Similar as the $\beta$-model and LSM, the logit-DCBM contains nonlinear factors, where fitting the parameters is a challenging open problem. We resolve this problem by introducing a cancellation trick. We also propose R-SCORE as a recursive community detection algorithm, where in each iteration, we first use the idea above to update our parameter estimation, and then use the results to remove the nonlinear factors in the logit-DCBM so the renormalized model approximately satisfies a low-rank model, just like the DCBM. Our numerical study suggests that R-SCORE significantly improves over existing spectral approaches in many cases. Also, theoretically, we show that the Hamming error rate of R-SCORE is faster than that of SCORE in a specific sparse region, and is at least as fast outside this region.

KV cache has become a *de facto* technique for the inference of large language models (LLMs), where tensors of shape (layer number, head number, sequence length, feature dimension) are introduced to cache historical information for self-attention. As the size of the model and data grows, the KV cache can, yet, quickly become a bottleneck within the system in both storage and memory transfer. To address this, prior studies usually focus on the first three axes of the cache tensors for compression. This paper supplements them, focusing on the feature dimension axis, by utilizing low-rank projection matrices to transform the cache features into spaces with reduced dimensions. We begin by investigating the canonical orthogonal projection method for data compression through principal component analysis (PCA). We identify the drawback of PCA projection that model performance degrades rapidly under relatively low compression rates (less than 60%). This phenomenon is elucidated by insights derived from the principles of attention mechanisms. To bridge the gap, we propose to directly tune the orthogonal projection matrix on the continual pre-training or supervised fine-tuning datasets with an elaborate Matryoshka learning strategy. Thanks to such a strategy, we can adaptively search for the optimal compression rates for various layers and heads given varying compression budgets. Compared to Multi-head Latent Attention (MLA), our method can easily embrace pre-trained LLMs and hold a smooth tradeoff between performance and compression rate. We witness the high data efficiency of our training procedure and find that our method can sustain over 90\% performance with an average KV cache compression rate of 60% (and up to 75% in certain extreme scenarios) for popular LLMs like LLaMA2 and Mistral.