ICML 2025 Papers

While Foundation Models provide a general tool for rapid content creation, they regularly require task-specific adaptation. Traditionally, this exercise involves careful curation of datasets and repeated fine-tuning of the underlying model. Fine-tuning techniques enable practitioners to adapt foundation models for many new applications but require expensive and lengthy training while being notably sensitive to hyperparameter choices. To overcome these limitations, we introduce Text-to-LoRA (T2L), a model capable of adapting large language models (LLMs) on the fly solely based on a natural language description of the target task. T2L is a hypernetwork trained to construct LoRAs in a single inexpensive forward pass. After training T2L on a suite of 9 pre-trained LoRA adapters (GSM8K, Arc, etc.), we show that the ad-hoc reconstructed LoRA instances match the performance of task-specific adapters across the corresponding test sets. Furthermore, T2L can compress hundreds of LoRA instances and zero-shot generalize to entirely unseen tasks. This approach provides a significant step towards democratizing the specialization of foundation models and enables language-based adaptation with minimal compute requirements. Our code is available at https://github.com/SakanaAI/text-to-lora

Fine-tuning LLMs is both computationally and memory-intensive. While parameter-efficient fine-tuning methods, such as QLoRA and DoRA, reduce the number of trainable parameters and lower memory usage, they do not decrease computational cost. In some cases, they may even slow down fine-tuning. In this paper, we introduce SparseLoRA, a method that accelerates LLM fine-tuning through contextual sparsity. We propose a lightweight, training-free SVD sparsity estimator that dynamically selects a sparse subset of weights for loss and gradient computation. Also, we systematically analyze and address sensitivity across layers, tokens, and training steps. Our experimental results show that SparseLoRA reduces computational cost by up to $2.0\times$ and a measured speedup of up to $1.5\times$ while maintaining accuracy across various downstream tasks, including commonsense and arithmetic reasoning, code generation, and instruction following.

In order to streamline the fine-tuning of foundation models, Low-Rank Adapters (LoRAs) have been substantially adopted across various fields, including instruction tuning and domain adaptation. The underlying concept of LoRA involves decomposing a full-rank matrix into the product of two lower-rank matrices, which reduces storage consumption and accelerates the training process. Furthermore, to address the limited expressive capacity of LoRA, the Mixture-of-Expert (MoE) has been introduced for incorporating multiple LoRA adapters. The integration of LoRA experts leads to a visible improvement across several downstream scenes. However, the mixture of LoRAs (MoE-LoRA) still exhibits its low robustness during tuning and inferring. Inspired by the Riemannian Preconditioners which train LoRA as a sub-space projector, we propose a new training strategy for MoE-LoRA, to stabilize and boost its feature learning by gate-rescaled multi-space projections. We provide both a theoretical solution as well as an alternative engineering strategy. Examinations on SGD and AdamW optimizers demonstrate the effectiveness of our methodology. Source code is available at https://github.com/THUDM/MoELoRA_Riemannian.

Pre-trained foundation models (FMs) have shown exceptional performance in univariate time series forecasting tasks. However, several practical challenges persist, including managing intricate dependencies among features and quantifying uncertainty in predictions. This study aims to tackle these critical limitations by introducing **adapters**—feature-space transformations that facilitate the effective use of pre-trained univariate time series FMs for multivariate tasks. Adapters operate by projecting multivariate inputs into a suitable latent space and applying the FM independently to each dimension. Inspired by the literature on representation learning and partially stochastic Bayesian neural networks, we present a range of adapters and optimization/inference strategies. Experiments conducted on both synthetic and real-world datasets confirm the efficacy of adapters, demonstrating substantial enhancements in forecasting accuracy and uncertainty quantification compared to baseline methods. Our framework, **AdaPTS**, positions adapters as a modular, scalable, and effective solution for leveraging time series FMs in multivariate contexts, thereby promoting their wider adoption in real-world applications. We release the code at https://github.com/abenechehab/AdaPTS.

We introduce Interactive Bayesian Distributional Robustness (IBDR), a novel Bayesian inference framework that allows modeling the interactions between particles, thereby enhancing ensemble quality through increased particle diversity. IBDR is grounded in a generalized theoretical framework that connects the distributional population loss with the approximate posterior, motivating a practical dual optimization procedure that enforces distributional robustness while fostering particle diversity. We evaluate IBDR's performance against various baseline methods using the VTAB-1K benchmark and the common reasoning language task. The results consistently show that IBDR outperforms these baselines, underscoring its effectiveness in real-world applications.

The LoRA method has achieved notable success in reducing GPU memory usage by applying low-rank updates to weight matrices. Yet, one simple question remains: can we push this reduction even further? Furthermore, is it possible to achieve this while improving performance and reducing computation time? Answering these questions requires moving beyond the conventional weight-centric approach. In this paper, we present a state-based fine-tuning framework that shifts the focus from weight adaptation to optimizing forward states, with LoRA acting as a special example. Specifically, state-based tuning introduces parameterized perturbations to the states within the computational graph, allowing us to control states across an entire residual block. A key advantage of this approach is the potential to avoid storing large intermediate states in models like transformers. Empirical results across multiple architectures—including ViT, RoBERTa, LLaMA2-7B, and LLaMA3-8B—show that our method further reduces memory consumption and computation time while simultaneously improving performance. Moreover, as a result of memory reduction, we explore the feasibility to train 7B/8B models on consumer-level GPUs like Nvidia 3090, without model quantization. The code is available at an anonymous GitHub repository

Machine unlearning presents a promising approach to mitigating privacy and safety concerns in large language models (LLMs) by enabling the selective removal of targeted data or knowledge while preserving model utility. However, existing unlearning methods remain over-sensitive to downstream fine-tuning, which can rapidly recover what is supposed to be unlearned information even when the fine-tuning task is entirely unrelated to the unlearning objective. To enhance robustness, we introduce the concept of `invariance' into unlearning for the first time from the perspective of invariant risk minimization (IRM), a principle for environment-agnostic training. By leveraging IRM, we develop a new invariance-regularized LLM unlearning framework, termed invariant LLM unlearning (ILU). We show that the proposed invariance regularization, even using only a single fine-tuning dataset during ILU training, can enable unlearning robustness to generalize effectively across diverse and new fine-tuning tasks at test time. A task vector analysis is also provided to further elucidate the rationale behind ILU's effectiveness. Extensive experiments on the WMDP benchmark, which focuses on removing an LLM's hazardous knowledge generation capabilities, reveal that ILU significantly outperforms state-of-the-art unlearning methods, including negative preference optimization (NPO) and representation misdirection for unlearning (RMU). Notably, ILU achieves superior unlearning robustness across diverse downstream fine-tuning scenarios (e.g., math, paraphrase detection, and sentiment analysis) while preserving the fine-tuning performance.

Modern machine learning pipelines are increasingly combining and mixing data from diverse and disparate sources, e.g., pre-training large language models. Yet, finding the optimal data mixture is a challenging and open problem. We formalize this data mixing problem as a bi-level objective: the best mixture is the one that would lead to the best model for a downstream objective. Unfortunately, this objective is generally intractable. In this paper, we make the observation that the bi-level data mixing objective becomes convex as our model class becomes larger. We develop and study a gradient-based approach for optimizing this convex objective, which we call MixMin, and test it on language modeling and chemistry tasks. MixMin was the only method that uniformly improved the data mixture in all our experiments. With MixMin, we improved the data mixture using less than 0.2% additional compute for a pythia-$410M$ model trained on $8.2B$ tokens, resulting between 1-5% relative improvement to negative log likelihood on PIQA, ARC Easy, SciQ, and OpenWebMath. Crucially, we found that MixMin mixtures for smaller models improved training of larger models, suggesting that MixMin mixtures may be scale-invariant. When mixing bioassay data to train an XGBoost model, we saw improvements to average precision scores of $0.03-0.15$.

Modern methods for fine-tuning a Vision Transformer (ViT) like Low-Rank Adaptation (LoRA) and its variants demonstrate impressive performance. However, these methods ignore the high-dimensional nature of Multi-Head Attention (MHA) weight tensors. To address this limitation, we propose Canonical Rank Adaptation (CaRA). CaRA leverages tensor mathematics, first by tensorising the transformer into two different tensors; one for projection layers in MHA and the other for feed-forward layers. Second, the tensorised formulation is fine-tuned using the low-rank adaptation in Canonical-Polyadic Decomposition (CPD) form. Employing CaRA efficiently minimizes the number of trainable parameters. Experimentally, CaRA outperforms existing Parameter-Efficient Fine-Tuning (PEFT) methods in visual classification benchmarks such as Visual Task Adaptation Benchmark (VTAB)-1k and Fine-Grained Visual Categorization (FGVC).

A widespread strategy to obtain a language model that performs well on a target domain is to finetune a pretrained model to perform unsupervised next-token prediction on data from that target domain. Finetuning presents two challenges: \textit{(i)} if the amount of target data is limited, as in most practical applications, the model will quickly overfit, and \textit{(ii)} the model will drift away from the original model, forgetting the pretraining data and the generic knowledge that comes with it. Our goal is to derive scaling laws that quantify these two phenomena for various target domains, amounts of available target data, and model scales. We measure the efficiency of injecting pretraining data into the finetuning data mixture to avoid forgetting and mitigate overfitting. A key practical takeaway from our study is that injecting as little as $1\%$ of pretraining data in the finetuning data mixture prevents the model from forgetting the pretraining set.

Motivated by the recent success of time-series foundation models for zero-shot forecasting, we present a methodology for _in-context fine-tuning_ of a time-series foundation model. In particular, we design a pretrained foundation model that can be prompted (at inference time) with multiple time-series examples, in order to forecast a target time-series into the future. Our foundation model is specifically trained to utilize examples from multiple related time-series in its context window (in addition to the history of the target time-series) to help it adapt to the specific distribution of the target domain at inference time. We show that such a foundation model that uses in-context examples at inference time can obtain much better performance on popular forecasting benchmarks compared to supervised deep learning methods, statistical models, and other time series foundation models. Interestingly, our in-context fine-tuning approach even matches the performance of a foundation model that is explicitly fine-tuned on the target domain.

Machine learning has demonstrated transformative potential for database operations, such as query optimization and in-database data analytics. However, dynamic database environments, characterized by frequent updates and evolving data distributions, introduce concept drift, which leads to performance degradation for learned models and limits their practical applicability. Addressing this challenge requires efficient frameworks capable of adapting to shifting concepts while minimizing the overhead of retraining or fine-tuning. In this paper, we propose FLAIR, an online adaptation framework that introduces a new paradigm called \textit{in-context adaptation} for learned database operations. FLAIR leverages the inherent property of data systems, i.e., immediate availability of execution results for predictions, to enable dynamic context construction. By formalizing adaptation as $f:(\mathbf{x} | \mathcal{C}_t) \to \mathbf{y}$, with $\mathcal{C}_t$ representing a dynamic context memory, FLAIR delivers predictions aligned with the current concept, eliminating the need for runtime parameter optimization. To achieve this, FLAIR integrates two key modules: a Task Featurization Module for encoding task-specific features into standardized representations, and a Dynamic Decision Engine, pre-trained via Bayesian meta-training, to adapt seamlessly using contextual information at runtime. Extensive experiments across key database tasks demonstrate that FLAIR outperforms state-of-the-art baselines, achieving up to $5.2\times$ faster adaptation and reducing error by 22.5\% for cardinality estimation.

Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant, uninformative, or even harmful. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves downstream performance. Code is available at https://github.com/UCSC-REAL/TokenCleaning.

*In silico* modeling of transcriptional responses to perturbations is crucial for advancing our understanding of cellular processes and disease mechanisms. We present PertEval-scFM, a standardized framework designed to evaluate models for perturbation effect prediction. We apply PertEval-scFM to benchmark zero-shot single-cell foundation model (scFM) embeddings against baseline models to assess whether these contextualized representations enhance perturbation effect prediction. Our results show that scFM embeddings offer limited improvement over simple baseline models in the zero-shot setting, particularly under distribution shift. Overall, this study provides a systematic evaluation of zero-shot scFM embeddings for perturbation effect prediction, highlighting the challenges of this task and the limitations of current-generation scFMs. Our findings underscore the need for specialized models and high-quality datasets that capture a broader range of cellular states. Source code and documentation can be found at: https://github.com/aaronwtr/PertEval.

Diffusion models have demonstrated their capabilities in modeling trajectories of multi-tasks. However, existing multi-task planners or policies typically rely on task-specific demonstrations via multi-task imitation, or require task-specific reward labels to facilitate policy optimization via Reinforcement Learning (RL). They are costly due to the substantial human efforts required to collect expert data or design reward functions. To address these challenges, we aim to develop a versatile diffusion planner capable of leveraging large-scale inferior data that contains task-agnostic sub-optimal trajectories, with the ability to fast adapt to specific tasks. In this paper, we propose SODP, a two-stage framework that leverages Sub-Optimal data to learn a Diffusion Planner, which is generalizable for various downstream tasks. Specifically, in the pre-training stage, we train a foundation diffusion planner that extracts general planning capabilities by modeling the versatile distribution of multi-task trajectories, which can be sub-optimal and has wide data coverage. Then for downstream tasks, we adopt RL-based fine-tuning with task-specific rewards to quickly refine the diffusion planner, which aims to generate action sequences with higher task-specific returns. Experimental results from multi-task domains including Meta-World and Adroit demonstrate that SODP outperforms state-of-the-art methods with only a small amount of data for reward-guided fine-tuning.

To address the challenge of scarce computational resources in genomic modeling, we introduce GERM, a genomic foundation model optimized for accessibility and adaptability. GERM improves upon models like DNABERT-2 by eliminating outliers that hinder low-rank adaptation and post-training quantization, enhancing both efficiency and robustness. We replace the vanilla attention layer with an outlier-free mechanism inspired by associative memory models. By removing outliers during both pre-training and fine-tuning, this approach accelerates adaptation, reduces computational costs, and enhances quantization robustness within acceptable loss margins. Additionally, we propose GERM-T, a strategy that employs small-step continual learning within the outlier-free framework, leveraging original checkpoints to avoid retraining from scratch. Empirically, GERM improves fine-tuning performance by 37.98% and quantization by 64.34% over the baseline model. It also reduces average kurtosis by 92.14% and maximum infinity norm by 82.77%. Compared to leading methods, GERM consistently delivers superior performance, offering a practical solution for genomic modeling in resource-constrained settings.

We introduce ProLoRA, enabling zero-shot adaptation of parameter-efficient fine-tuning in text-to-image diffusion models. ProLoRA transfers pre-trained low-rank adjustments (e.g., LoRA) from a source to a target model without additional training data. This overcomes the limitations of traditional methods that require retraining when switching base models, often challenging due to data constraints. ProLoRA achieves this via projection of source adjustments into the target model's weight space, leveraging subspace and null space similarities and selectively targeting aligned layers. Evaluations on established text-to-image models demonstrate successful knowledge transfer and comparable performance without retraining.

Low-rank adaptation (LoRA) has become a standard approach for fine-tuning large foundation models. However, our theoretical understanding of LoRA remains limited as prior analyses of LoRA's training dynamics either rely on linearization arguments or consider highly simplified setups. In this work, we analyze the LoRA loss landscape without such restrictive assumptions. We define two regimes: a "special regime", which includes idealized setups where linearization arguments hold, and a "generic regime" representing more realistic setups where linearization arguments do not hold. In the generic regime, we show that LoRA training converges to a global minimizer with low rank and small magnitude, or a qualitatively distinct solution with high rank and large magnitude. Finally, we argue that the zero-initialization and weight decay in LoRA training induce an implicit bias toward the low-rank, small-magnitude region of the parameter space—where global minima lie—thus shedding light on why LoRA training usually succeeds in finding global minima.

Inspired by the success of unsupervised pre-training paradigms, researchers have applied these approaches to DNA pre-training. However, we argue that these approaches alone yield suboptimal results because pure DNA sequences lack sufficient information, since their functions are regulated by genomic profiles like chromatin accessibility. Here, we demonstrate that supervised training for genomic profile prediction serves as a more effective alternative to pure sequence pre-training. Furthermore, considering the multi-species and multi-profile nature of genomic profile prediction, we introduce our **S**pecies-**P**rofile **A**daptive **C**ollaborative **E**xperts (SPACE) that leverages Mixture of Experts (MoE) to better capture the relationships between DNA sequences across different species and genomic profiles, thereby learning more effective DNA representations. Through extensive experiments across various tasks, our model achieves state-of-the-art performance, establishing that DNA models trained with supervised genomic profiles serve as powerful DNA representation learners.