ICLR 2025 Papers

Despite the impressive performance of general large language models(LLMs), many of their applications in specific domains (e.g., low-data and knowledge-intensive) still confront significant challenges. Supervised fine-tuning (SFT)---where a general LLM is further trained on a small labeled dataset to adapt for specific tasks or domains---has shown great power for developing domain-specific LLMs. However, existing SFT data primarily consist of Question and Answer (Q&A) pairs, which poses a significant challenge for LLMs to comprehend the correlation and logic of knowledge underlying the Q&A. To address this challenge, we propose a conceptually flexible and general framework to boost SFT, namely Knowledge Graph-Driven Supervised Fine-Tuning (KG-SFT). The key idea of KG-SFT is to generate high-quality explanations for each Q&A pair via a structured knowledge graph to enhance the knowledge comprehension and manipulation of LLMs. Specifically, KG-SFT consists of three components: Extractor, Generator, and Detector. For a given Q&A pair, (i) Extractor first identifies entities within Q&A pairs and extracts relevant reasoning subgraphs from external KGs, (ii) Generator then produces corresponding fluent explanations utilizing these reasoning subgraphs, and (iii) finally, Detector performs sentence-level knowledge conflicts detection on these explanations to guarantee the reliability. KG-SFT focuses on generating high-quality explanations to improve the quality of Q&A pair, which reveals a promising direction for supplementing existing data augmentation methods. Extensive experiments on fifteen different domains and six different languages demonstrate the effectiveness of KG-SFT, leading to an accuracy improvement of up to 18% and an average of 8.7% in low-data scenarios.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, applying ZO fine-tuning in memory-constrained settings such as mobile phones and laptops remains challenging since these settings often involve weight quantization, while ZO requires full-precision perturbation and update. In this study, we address this limitation by combining static sparse ZO fine-tuning with quantization. Our approach transfers a small, static subset (0.1%) of "sensitive" parameters from pre-training to downstream tasks, focusing fine-tuning on this sparse set of parameters. The remaining untuned parameters are quantized, reducing memory demands. Our proposed workflow enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8GB of memory while outperforming full model ZO fine-tuning performance and in-context learning.

We introduce DRESS, a novel approach for generating stylized large language model (LLM) responses through representation editing. Existing methods like prompting and fine-tuning are either insufficient for complex style adaptation or computationally expensive, particularly in tasks like NPC creation or character role-playing. Our approach leverages the over-parameterized nature of LLMs to disentangle a style-relevant subspace within the model's representation space to conduct representation editing, ensuring a minimal impact on the original semantics. By applying adaptive editing strengths, we dynamically adjust the steering vectors in the style subspace to maintain both stylistic fidelity and semantic integrity. We develop two stylized QA benchmark datasets to validate the effectiveness of DRESS, and the results demonstrate significant improvements compared to baseline methods such as prompting and ITI. In short, DRESS is a lightweight, train-free solution for enhancing LLMs with flexible and effective style control, making it particularly useful for developing stylized conversational agents. Codes and benchmark datasets are available at https://github.com/ArthurLeoM/DRESS-LLM.

Recent research shows that fine-tuning on benign instruction-following data can inadvertently undo the safety alignment process and increase a model's propensity to comply with harmful queries. While instruction-following fine-tuning is important, task-specific fine-tuning-where models are trained on datasets with clear ground truth answers (e.g., multiple choice questions)-can enhance model performance on specialized downstream tasks. Understanding and mitigating safety risks in the task-specific setting remains distinct from the instruction-following context due to structural differences in the data. Our work demonstrates how malicious actors can subtly manipulate the structure of almost *any* task-specific dataset to foster significantly more dangerous model behaviors, while maintaining an appearance of innocuity and reasonable downstream task performance. To address this issue, we propose a novel mitigation strategy that mixes in safety data which *mimics* the task format and prompting style of the user data, showing this is significantly more effective and efficient than existing baselines at re-establishing safety alignment while maintaining similar task performance.

Large Multimodal Models (LMMs), or Vision-Language Models (VLMs), have shown impressive capabilities in a wide range of visual tasks. However, they often struggle with fine-grained visual reasoning, failing to identify domain-specific objectives and provide justifiable explanations for their predictions. To address the above challenge, we propose a novel visual rejection sampling framework to improve the cognition and explainability of LMMs using self-synthesized data. Specifically, visual fine-tuning requires images, queries, and target answers. Our approach begins by synthesizing interpretable answers that include human-verifiable visual features. These features are based on expert-defined concepts, and carefully selected based on their alignment with the image content. After each round of fine-tuning, we apply a reward model-free filtering mechanism to select the highest-quality interpretable answers for the next round of tuning. This iterative process of synthetic data generation and fine-tuning progressively improves the model's ability to generate accurate and reasonable explanations. Experimental results demonstrate the effectiveness of our method in improving both the accuracy and explainability of specialized visual classification tasks.

The training and fine-tuning of large language models (LLMs) often involve diverse textual data from multiple sources, which poses challenges due to conflicting gradient directions, hindering optimization and specialization. These challenges can undermine model generalization across tasks, resulting in reduced downstream performance. Recent research suggests that fine-tuning LLMs on carefully selected, task-specific subsets of data can match or even surpass the performance of using the entire dataset. Building on these insights, we propose the Ensembles of Low-Rank Expert Adapters (ELREA) framework to improve the model's capability to handle diverse tasks. ELREA clusters the training instructions based on their gradient directions, representing different areas of expertise and thereby reducing conflicts during optimization. Expert adapters are then trained on these clusters, utilizing the low-rank adaptation (LoRA) technique to ensure training efficiency and model scalability. During inference, ELREA combines predictions from the most relevant expert adapters based on the input data's gradient similarity to the training clusters, ensuring optimal adapter selection for each task. Experiments show that our method outperforms baseline LoRA adapters trained on the full dataset and other ensemble approaches with similar training and inference complexity across a range of domain-specific tasks.

Large Language Models (LLMs) generating unsafe responses to toxic prompts is a significant issue in their applications. While various efforts aim to address this safety concern, previous approaches often demand substantial human data collection or rely on the less dependable option of using another LLM to generate corrective data. In this paper, we aim to take this problem and overcome limitations of requiring significant high-quality human data. Our method requires only a small set of unsafe responses to toxic prompts, easily obtained from the unsafe LLM itself. By employing a semantic cost combined with a negative Earth Mover Distance (EMD) loss, we guide the LLM away from generating unsafe responses. Additionally, we propose a novel lower bound for EMD loss, enabling more efficient optimization. Our results demonstrate superior performance and data efficiency compared to baselines, and we further examine the nuanced effects of over-alignment and potential degradation of language capabilities when using contrastive data.

Given a collection of neural networks, can we determine which are parent models and which are child models fine-tuned from the parents? In this work, we strive to answer this question via introducing a new task termed as neural phylogeny detection, aimed at identifying the existence and direction of the fine-tuning relationship. Specifically, neural phylogeny detection attempts to identify all parent-child model pairs and determine, within each pair, which model is the parent and which is the child. We present two approaches for neural phylogeny detection: a learning-free method and a learning-based method. First, we propose a metric that leverages the distance from network parameters to a fake initialization to infer fine-tuning directions. By integrating this metric with traditional clustering algorithms, we propose a series of efficient, learning-free neural phylogeny detection methods. Second, we introduce a transformer-based neural phylogeny detector, which significantly enhances detection accuracy through a learning-based manner. Extensive experiments, ranging from shallow fully-connected networks to open-sourced Stable Diffusion and LLaMA models, progressively validate the effectiveness of both methods. The results demonstrate the reliability of both the learning-free and the learning-based approaches across various learning tasks and network architectures, as well as their ability to detect cross-generational phylogeny between ancestor models and their fine-tuned descendants.

We propose a Competitive Low-Rank Adaptation (ComLoRA) framework to address the limitations of the LoRA method, which either lacks capacity with a single rank-$r$ LoRA or risks inefficiency and overfitting with a larger rank-$Kr$ LoRA, where $K$ is an integer larger than 1. The proposed ComLoRA method initializes $K$ distinct LoRA components, each with rank $r$, and allows them to compete during training. This competition drives each LoRA component to outperform the others, improving overall model performance. The best-performing LoRA is selected based on validation metrics, ensuring that the final model outperforms a single rank-$r$ LoRA and matches the effectiveness of a larger rank-$Kr$ LoRA, all while avoiding extra computational overhead during inference. To the best of our knowledge, this is the first work to introduce and explore competitive learning in the context of LoRA optimization. The ComLoRA's code is available at https://github.com/hqsiswiliam/comlora.

We investigate the statistical and computational limits of prompt tuning for transformer-based foundation models. Our key contributions are that prompt tuning on *single-head* transformers with only a *single* self-attention layer: (i) is universal, and (ii) supports efficient (even almost-linear time) algorithms under the Strong Exponential Time Hypothesis (SETH). Statistically, we prove that prompt tuning on such the simplest possible transformers are universal approximators for sequence-to-sequence Lipschitz functions. In addition, we provide an exponential-in-$dL$ and -in-$(1/\epsilon)$ lower bound on the required soft-prompt tokens for prompt tuning to memorize any dataset with 1-layer, 1-head transformers. Computationally, we identify a phase transition in the efficiency of prompt tuning, determined by the norm of the *soft-prompt-induced* keys and queries, and provide an upper bound criterion. Beyond this criterion, no sub-quadratic (efficient) algorithm for prompt tuning exists under SETH. Within this criterion, we showcase our theory by proving the existence of almost-linear time prompt tuning inference algorithms. These fundamental limits provide important necessary conditions for designing expressive and efficient prompt tuning methods for practitioners.

Low-Rank Adaptation (LoRA) has emerged as a popular technique for fine-tuning large language models (LLMs) to various domains due to its modular design and widespread availability on platforms like Huggingface. This modularity has sparked interest in combining multiple LoRAs to significantly enhance LLM capabilities. However, existing methods for LoRA composition primarily focus on task-specific adaptations that require additional training, and current model merging techniques often fail to fully leverage LoRA's modular nature, leading to parameter interference and performance degradation. In this paper, we explore the possibility of disassembling and reassembling multiple LoRAs at a finer granularity, much like assembling LEGO blocks. We introduce the concept of Minimal Semantic Units (MSUs), where the parameters corresponding to each rank in LoRA function as independent units. These MSUs exhibit properties such as permutation invariance and concatenation-summation equivalence, allowing for flexible combinations to form new LoRAs. Building on these insights, we propose the LoRA-LEGO framework. This framework conducts rank-wise parameter clustering by grouping MSUs from different LoRAs into $k$ clusters. The centroid of each cluster serves as a representative MSU, enabling the assembly of a merged LoRA with an adjusted rank of $k$. Additionally, we apply a dual reweighting strategy to optimize the scale of the merged LoRA. Experiments across various benchmarks demonstrate that our method outperforms existing approaches in LoRA merging.

Diffusion Language Models (DLMs) have emerged as a promising new paradigm for text generative modeling, potentially addressing limitations of autoregressive (AR) models. However, current DLMs have been studied at a smaller scale compared to their AR counterparts and lack fair comparison on language modeling benchmarks. Additionally, training diffusion models from scratch at scale remains challenging. Given the prevalence of open-source AR language models, we propose adapting these models to build text diffusion models. We demonstrate connections between AR and diffusion modeling objectives and introduce a simple continual pre-training approach for training diffusion models. Through systematic evaluation on language modeling, reasoning, and commonsense benchmarks, we show that we can convert AR models ranging from 127M to 7B parameters (GPT2 and LLaMA) into diffusion models DiffuGPT and DiffuLLaMA, using less than 200B tokens for training. Our experimental results reveal that these models outperform earlier DLMs and are competitive with their AR counterparts. We release a suite of DLMs (127M-355M-7B) capable of generating fluent text, performing in-context learning, filling in the middle without prompt re-ordering, and following instructions.

Adapting foundation models for specific purposes has become a standard approach to build machine learning systems for downstream applications. Yet, it is an open question which mechanisms take place during adaptation. Here we develop a new Sparse Autoencoder (SAE) for the CLIP vision transformer, named PatchSAE, to extract interpretable concepts at granular levels (e.g., shape, color, or semantics of an object) and their patch-wise spatial attributions. We explore how these concepts influence the model output in downstream image classification tasks and investigate how recent state-of-the-art prompt-based adaptation techniques change the association of model inputs to these concepts. While activations of concepts slightly change between adapted and non-adapted models, we find that the majority of gains on common adaptation tasks can be explained with the existing concepts already present in the non-adapted foundation model. This work provides a concrete framework to train and use SAEs for Vision Transformers and provides insights into explaining adaptation mechanisms.

Prior parameter-efficient fine-tuning (PEFT) algorithms reduce memory usage and computational costs of fine-tuning large neural network models by training only a few additional adapter parameters, rather than the entire model. However, the reduction in computational costs due to PEFT does not necessarily translate to a reduction in training time; although the computational costs of the adapter layers are much smaller than the pretrained layers, it is well known that those two types of layers are processed sequentially on GPUs, resulting in significant latency overhead. LoRA and its variants avoid this latency overhead by merging the low-rank adapter matrices with the pretrained weights during inference. However, those layers cannot be merged during training since the pretrained weights must remain frozen while the low-rank adapter matrices are updated continuously over the course of training. Furthermore, LoRA and its variants do not reduce activation memory, as the first low-rank adapter matrix still requires the input activations to the pretrained weights to compute weight gradients. To mitigate this issue, we propose **Pa**rtial **C**onnection **A**daptation (**PaCA**), which fine-tunes randomly selected partial connections within the pretrained weights instead of introducing adapter layers in the model. PaCA not only enhances training speed by eliminating the time overhead due to the sequential processing of the adapter and pretrained layers but also reduces activation memory since only partial activations, rather than full activations, need to be stored for gradient computation. Compared to LoRA, PaCA reduces training time by 22% and total memory usage by 16%, while maintaining comparable accuracy across various fine-tuning scenarios, such as fine-tuning on the MMLU dataset and instruction tuning on the Oasst1 dataset. PaCA can also be combined with quantization, enabling the fine-tuning of large models such as LLaMA3.1-70B. In addition, PaCA enables training with 23% longer sequence and improves throughput by 16\% on both NVIDIA A100 GPU and INTEL Gaudi2 HPU compared to LoRA. The code is available at [https://github.com/WooSunghyeon/paca](https://github.com/WooSunghyeon/paca).

We investigate LoRA in federated learning through the lens of the asymmetry analysis of the learned $A$ and $B$ matrices. In doing so, we uncover that $A$ matrices are responsible for learning general knowledge, while $B$ matrices focus on capturing client-specific knowledge. Based on this finding, we introduce Federated Share-A Low-Rank Adaptation (FedSA-LoRA), which employs two low-rank trainable matrices $A$ and $B$ to model the weight update, but only $A$ matrices are shared with the server for aggregation. Moreover, we delve into the relationship between the learned $A$ and $B$ matrices in other LoRA variants, such as rsLoRA and VeRA, revealing a consistent pattern. Consequently, we extend our FedSA-LoRA method to these LoRA variants, resulting in FedSA-rsLoRA and FedSA-VeRA. In this way, we establish a general paradigm for integrating LoRA with FL, offering guidance for future work on subsequent LoRA variants combined with FL. Extensive experimental results on natural language understanding and generation tasks demonstrate the effectiveness of the proposed method. Our code is available at https://github.com/Pengxin-Guo/FedSA-LoRA.

Adapting large language models (LLMs) to specialized domains typically requires domain-specific corpora for continual pre-training to facilitate knowledge memorization and related instructions for fine-tuning to apply this knowledge. However, this method may lead to inefficient knowledge memorization due to a lack of awareness of knowledge utilization during the continual pre-training and demands LLMs to simultaneously learn knowledge utilization and format alignment with divergent training objectives during the fine-tuning. To enhance the domain adaptation of LLMs, we revise this process and propose a new domain adaptation framework including domain knowledge learning and general format alignment, called \emph{Mix-CPT}. Specifically, we first conduct a knowledge mixture continual pre-training that concurrently focuses on knowledge memorization and utilization. To avoid catastrophic forgetting, we further propose a logit swap self-distillation constraint. By leveraging the knowledge and capabilities acquired during continual pre-training, we then efficiently perform instruction tuning and alignment with a few general training samples to achieve format alignment. Extensive experiments show that our proposed \emph{Mix-CPT} framework can simultaneously improve the task-solving capabilities of LLMs on the target and general domains.

As language models (LMs) approach human-level performance, a comprehensive understanding of their behavior becomes crucial. This includes evaluating capabilities, biases, task performance, and alignment with societal values. Extensive initial evaluations, including red teaming and diverse benchmarking, can establish a model’s behavioral profile. However, subsequent fine-tuning or deployment modifications may alter these behaviors in unintended ways. We present an efficient statistical test to tackle Behavioral Shift Auditing (BSA) in LMs, which we define as detecting distribution shifts in qualitative properties of the output distributions of LMs. Our test compares model generations from a baseline model to those of the model under scrutiny and provides theoretical guarantees for change detection while controlling false positives. The test features a configurable tolerance parameter that adjusts sensitivity to behavioral changes for different use cases. We evaluate our approach using two case studies: monitoring changes in (a) toxicity and (b) translation performance. We find that the test is able to detect meaningful changes in behavior distributions using just hundreds of examples.

As large language models (LLMs) increasingly shape the AI landscape, fine-tuning pretrained models has become more popular than in the pre-LLM era for achieving optimal performance in domain-specific tasks. However, pretrained LLMs such as ChatGPT are periodically evolved (i.e., model parameters are frequently updated), making it challenging for downstream users with limited resources to keep up with fine-tuning the newest LLMs for their domain application. Even though fine-tuning costs have nowadays been reduced thanks to the innovations of parameter-efficient fine-tuning such as LoRA, not all downstream users have adequate computing for frequent personalization. Moreover, access to fine-tuning datasets, particularly in sensitive domains such as healthcare, could be time-restrictive, making it crucial to retain the knowledge encoded in earlier fine-tuned rounds for future adaptation. In this paper, we present PORTLLM, a training-free framework that (i) creates an initial lightweight model update patch to capture domain-specific knowledge, and (ii) allows a subsequent seamless plugging for the continual personalization of evolved LLM at minimal cost. Our extensive experiments cover seven representative datasets, from easier question-answering tasks {BoolQ, SST2} to harder reasoning tasks {WinoGrande, GSM8K}, and models including {Mistral-7B,Llama2, Llama3.1, and Gemma2}, validating the portability of our designed model patches and showcasing the effectiveness of our proposed framework. For instance, PORTLLM achieves comparable performance to LoRA fine-tuning with reductions of up to 12.2× in GPU memory usage. Finally, we provide theoretical justifications to understand the portability of our model update patches, which offers new insights into the theoretical dimension of LLMs’ personalization.

Robust fine-tuning aims to adapt large foundation models to downstream tasks while preserving their robustness to distribution shifts. Existing methods primarily focus on constraining and projecting current model towards the pre-trained initialization based on the magnitudes between fine-tuned and pre-trained weights, which often require extensive hyper-parameter tuning and can sometimes result in underfitting. In this work, we propose $\textbf{Di}$rectional $\textbf{Gra}$dient $\textbf{P}$rojection (DiGraP), a novel layer-wise trainable method that incorporates directional information from gradients to bridge regularization and multi-objective optimization. Besides demonstrating our method on image classification, as another contribution we generalize this area to the multi-modal evaluation settings for robust fine-tuning. Specifically, we first bridge the uni-modal and multi-modal gap by performing analysis on Image Classification reformulated Visual Question Answering (VQA) benchmarks and further categorize ten out-of-distribution (OOD) VQA datasets by distribution shift types and degree (i.e. near versus far OOD). Experimental results show that DiGraP consistently outperforms existing baselines across Image Classfication and VQA tasks with discriminative and generative backbones, improving both in-distribution (ID) generalization and OOD robustness.