ICLR 2025 Papers

Prompt Tuning (PT) enables the adaptation of Pre-trained Large Language Models (PLMs) to downstream tasks by optimizing a small amount of soft virtual tokens, which are prepended to the input token embeddings. Recently, Decomposed Prompt Tuning (DePT) has demonstrated superior adaptation capabilities by decomposing the soft prompt into a shorter soft prompt and a pair of low-rank matrices. The product of the pair of low-rank matrices is added to the input token embeddings to offset them. Additionally, DePT achieves faster inference compared to PT due to the shorter soft prompt. However, in this paper, we find that the position-based token embedding offsets of DePT restricts its ability to generalize across diverse model inputs, and that the shared embedding offsets across many token embeddings result in sub-optimization. To tackle these issues, we introduce \textbf{A}daptive \textbf{De}composed \textbf{P}rompt \textbf{T}uning (ADePT), which is composed of a short soft prompt and a shallow token-shared feed-forward neural network. ADePT utilizes the token-shared feed-forward neural network to learn the embedding offsets for each token, enabling adaptive embedding offsets that vary according to the model input and better optimization of token embedding offsets. This enables ADePT to achieve superior adaptation performance without requiring more inference time or additional trainable parameters compared to vanilla PT and its variants. In comprehensive experiments across 23 natural language processing tasks and 4 typical PLMs of different scales, ADePT consistently surpasses the leading parameter-efficient fine-tuning methods, and even outperforms the full fine-tuning in certain scenarios. We also provide a theoretical analysis towards ADePT. Code is available at https://github.com/HungerPWAY/ADePT.

Direct Preference Optimization (DPO) has emerged as a more computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) with Proximal Policy Optimization (PPO), eliminating the need for reward models and online sampling. Despite these benefits, DPO and its variants remain sensitive to hyper-parameters and prone to instability, particularly on mathematical datasets. We argue that these issues arise from the unidirectional likelihood-derivative negative feedback inherent in the log-likelihood loss function. To address this, we propose a novel LLM alignment loss that establishes a stable Bidirectional Negative Feedback (BNF) during optimization. Our proposed BNF loss eliminates the need for pairwise contrastive losses and does not require any extra tunable hyper-parameters or pairwise preference data, streamlining the alignment pipeline to be as simple as supervised fine-tuning. We conduct extensive experiments across two challenging QA benchmarks and four reasoning benchmarks. The experimental results show that BNF achieves comparable performance to the best methods on QA benchmarks, while its performance decrease on the four reasoning benchmarks is significantly lower compared to the best methods, thus striking a better balance between value alignment and reasoning ability. In addition, we further validate the performance of BNF on non-pairwise datasets, and conduct in-depth analysis of log-likelihood and logit shifts across different preference optimization methods. We will release all the source code, checkpoints, and datasets on GitHub.

How to align large language models (LLMs) with user preferences from a static general dataset has been frequently studied. However, user preferences are usually personalized, changing, and diverse. This leads to the problem that the actual user preferences often do not coincide with those trained by the model developers in the practical use of LLMs. Since we cannot collect enough data and retrain for every demand, researching efficient real-time preference adaptation methods based on the backbone LLMs during test time is important. To this end, we introduce **Amulet**, a novel, training-free framework that formulates the decoding process of every token as a separate online learning problem with the guidance of simple user-provided prompts, thus enabling real-time optimization to satisfy users' personalized preferences. To reduce the computational cost brought by this optimization process for each token, we additionally provide a closed-form solution for each iteration step of the optimization process, thereby reducing the computational time cost to a negligible level. The detailed experimental results demonstrate that Amulet can achieve significant performance improvements in rich settings with combinations of different LLMs, datasets, and user preferences, while maintaining acceptable computational efficiency.

The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources to effectively explore the experiment space. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. The code used for the experiments can be found here: https://github.com/instructlab/training. Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, allowing for early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observe no significant difference in performance between phased (sequentially training on data divided into phases) and stacked (training on the entire dataset at once) strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets as well as model families and sizes, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive research environment for LLM development.

We model the evolution of tokens within a deep stack of Transformer layers as a continuous-time flow on the unit sphere, governed by a mean-field interacting particle system, building on the framework introduced in Geshkovski et al. (2023). Studying the corresponding mean-field Partial Differential Equation (PDE), which can be interpreted as a Wasserstein gradient flow, in this paper we provide a mathematical investigation of the long-term behavior of this system, with a particular focus on the emergence and persistence of meta-stable phases and clustering phenomena, key elements in applications like next-token prediction. More specifically, we perform a perturbative analysis of the mean-field PDE around the iid uniform initialization and prove that, in the limit of large number of tokens, the model remains close to a meta-stable manifold of solutions with a given structure (e.g., periodicity). Further, the structure characterizing the meta-stable manifold is explicitly identified, as a function of the inverse temperature parameter of the model, by the index maximizing a certain rescaling of Gegenbauer polynomials.

Low-Rank Adaptation (LoRA) presents an effective solution for federated fine-tuning of Large Language Models (LLMs), as it substantially reduces communication overhead. However, a straightforward combination of FedAvg and LoRA results in suboptimal performance, especially under data heterogeneity. We noted this stems from both intrinsic (i.e., constrained parameter space) and extrinsic (i.e., client drift) limitations, which hinder it effectively learn global knowledge. In this work, we proposed a novel Federated Residual Low-Rank Adaption method, namely FRLoRA, to tackle above two limitations. It directly sums the weight of the global model parameters with a residual low-rank matrix product (\ie, weight change) during the global update step, and synchronizes this update for all local models. By this, FRLoRA performs global updates in a higher-rank parameter space, enabling a better representation of complex knowledge structure. Furthermore, FRLoRA reinitializes the local low-rank matrices with the principal singular values and vectors of the pre-trained weights in each round, to calibrate their inconsistent convergence, thereby mitigating client drift. Our extensive experiments demonstrate that FRLoRA consistently outperforms various state-of-the-art FL methods across nine different benchmarks in natural language understanding and generation under different FL scenarios.

The Mixture of Experts (MoE) architecture has emerged as a promising solution to reduce computational overhead by selectively activating subsets of model parameters. The effectiveness of MoE models depends primarily on their routing mechanisms, with the widely adopted Top-K routing scheme used for activating experts. However, the Top-K scheme has notable limitations, including unnecessary activations and underutilization of experts. In this work, rather than modifying the routing mechanism as done in previous studies, we propose the Ternary Choice MoE (TC-MoE), a novel approach that expands the expert space by applying the ternary set {-1, 0, 1} to each expert. This expansion allows more efficient and effective expert activations without incurring significant computational costs. Additionally, given the unique characteristics of the expanded expert space, we introduce a new load balance loss and reward loss to ensure workload balance and achieve a flexible trade-off between effectiveness and efficiency. Extensive experiments demonstrate that TC-MoE achieves an average improvement of over 1.1% compared with traditional approaches, while reducing the average number of activated experts by up to 9%. These results confirm that TC-MoE effectively addresses the inefficiencies of conventional routing schemes, offering a more efficient and scalable solution for MoE-based large language models. Code and models are available at https://github.com/stiger1000/TC-MoE.

In recent years, *task arithmetic* has garnered increasing attention. This approach edits pre-trained models directly in weight space by combining the fine-tuned weights of various tasks into a *unified model*. Its efficiency and cost-effectiveness stem from its training-free combination, contrasting with traditional methods that require model training on large datasets for multiple tasks. However, applying such a unified model to individual tasks can lead to interference from other tasks (lack of *weight disentanglement*). To address this issue, Neural Tangent Kernel (NTK) linearization has been employed to leverage a ''kernel behavior'', facilitating weight disentanglement and mitigating adverse effects from unrelated tasks. Despite its benefits, NTK linearization presents drawbacks, including doubled training costs, as well as reduced performance of individual models. To tackle this problem, we propose a simple yet effective and efficient method that is to finetune the attention modules only in the Transformer. Our study reveals that the attention modules exhibit kernel behavior, and fine-tuning the attention modules only significantly improves weight disentanglement. To further understand how our method improves the weight disentanglement of task arithmetic, we present a comprehensive study of task arithmetic by differentiating the role of the representation module and task-specific module. In particular, we find that the representation module plays an important role in improving weight disentanglement whereas the task-specific modules such as the classification heads can degenerate the weight disentanglement performance.

Self-adaptive large language models (LLMs) aim to solve the challenges posed by traditional fine-tuning methods, which are often computationally intensive and static in their ability to handle diverse tasks. We introduce Transformer-Squared, a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference, Transformer-Squared employs a two-pass mechanism: first, a dispatch system identifies the task properties, and then task-specific 'expert' vectors, trained using reinforcement learning, are dynamically mixed to obtain targeted behavior for the incoming prompt. Our method consistently outperforms ubiquitous approaches such as LoRA, with fewer parameters and greater efficiency. Furthermore, Transformer-Squared demonstrates versatility across different LLM architectures and modalities, including vision-language tasks. Transformer-Squared represents a significant leap forward, offering a scalable, efficient solution for enhancing the adaptability and task-specific performance of LLMs, paving the way for truly dynamic, self-organizing AI systems.

This paper introduces Quantum-PEFT that leverages quantum computations for parameter-efficient fine-tuning (PEFT). Unlike other additive PEFT methods, such as low-rank adaptation (LoRA), Quantum-PEFT exploits an underlying full-rank yet surprisingly parameter efficient _quantum unitary parameterization_. With the use of Pauli parameterization, the number of trainable parameters grows only logarithmically with the ambient dimension, as opposed to linearly as in LoRA-based PEFT methods. Quantum-PEFT achieves vanishingly smaller number of trainable parameters than the lowest-rank LoRA as dimensions grow, enhancing parameter efficiency while maintaining a competitive performance. We apply Quantum-PEFT to several transfer learning benchmarks in language and vision, demonstrating significant advantages in parameter efficiency.

Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL (**T**raining-Fre**e** **A**ctivation Sparsity in **L**LMs), a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50\% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53× and 1.8× at 40\% and 50\% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

Direct Preference Optimization (DPO) has recently emerged as a popular approach to improve reinforcement learning from human feedback (RLHF), leading to better techniques to fine-tune large language models (LLM). A weakness of DPO, however, lies in its lack of capability to characterize the diversity of human preferences. Inspired by Mallows' theory of preference ranking, we develop in this paper a new approach, the *MallowsPO*. A distinct feature of this approach is a *dispersion index*, which reflects the dispersion of human preference to prompts. We show that existing DPO models can be reduced to special cases of this dispersion index, thus unified with MallowsPO. More importantly, we demonstrate empirically how to use this dispersion index to enhance the performance of DPO in a broad array of benchmark tasks, from synthetic bandit selection to controllable generation and dialogues, while maintaining great generalization capabilities. MallowsPO is also compatible with other SOTA offline preference optimization methods, boosting nearly 2\% extra LC win rate when used as a plugin for fine-tuning Llama3-Instruct.

Low-Rank Adaptation (LoRA) has become a widely used method for parameter-efficient fine-tuning of large-scale, pre-trained neural networks. However, LoRA and its extensions face several challenges, including the need for rank adaptivity, robustness, and computational efficiency during the fine-tuning process. We introduce GeoLoRA, a novel approach that addresses these limitations by leveraging dynamical low-rank approximation theory. GeoLoRA requires only a single backpropagation pass over the small-rank adapters, significantly reducing computational cost as compared to similar dynamical low-rank training methods and making it faster than popular baselines such as AdaLoRA. This allows GeoLoRA to efficiently adapt the allocated parameter budget across the model, achieving smaller low-rank adapters compared to heuristic methods like AdaLoRA and LoRA, while maintaining critical convergence, descent, and error-bound theoretical guarantees. The resulting method is not only more efficient but also more robust to varying hyperparameter settings. We demonstrate the effectiveness of GeoLoRA on several state-of-the-art benchmarks, showing that it outperforms existing methods in both accuracy and computational efficiency

Large language models (LLMs) acquire substantial knowledge during pretraining but often need adaptation to new contexts, tasks, or domains, typically achieved through fine-tuning or prompting. However, fine-tuning incurs significant training costs, while prompting increases inference overhead. Inspired by fast weight memory, we introduce GenerativeAdapter, an effective and efficient adaptation method that encode test-time context into language model parameters with a single forward pass. GenerativeAdapter augments a frozen pretrained LM with a lightweight adapter generator, trained via self-supervised learning, to produce parameter-efficient adapters. Notably, our generator is general-purpose, i.e., one generator can adapt the corresponding base model for all langauge processing scenarios. We apply GenerativeAdapter to two pretrained LMs (Mistral-7B-Instruct and Llama2-7B-Chat) and evaluate the adapted models across knowledge acquisition from documents, learning from demonstrations, and personalization for users. In StreamingQA, our approach is effective in injecting knowledge into the LM's parameters, achieving a 63.5\% improvement in F1 score over the model with supervised fine-tuning (from $19.5$ to $31.5$) for contexts as long as 32K tokens. In the MetaICL in-context learning evaluation, our method achieves an average accuracy of $44.9$ across 26 tasks, outperforming the base model. On MSC, our method proves to be highly competitive in memorizing user information from conversations with a 4x reduction in computation and memory costs compared to prompting with full conversation history. Overall, GenerativeAdapter provides a viable solution for adapting large LMs to evolving information and providing tailored user experience, while reducing training and inference costs relative to traditional fine-tuning and prompting techniques.

Fine-tuning large language models (LLMs) is necessary for specific downstream tasks, but classic first-order optimizer entails prohibitive GPU memory because of the back propagation. Recent works such as MeZO have turned to zeroth-order optimizers for fine-tuning, which reduce substantial memory by using two forward passes. However, heterogeneous curvatures across different parameter dimensions in LLMs often cause model convergence instability or even failure. In this work, we propose HiZOO, a diagonal Hessian informed Zeroth-Order Optimizer , which is the first work to leverage the diagonal Hessian to enhance ZOO for fine-tuning LLMs. We provide theoretical proof for HiZOO and visualize the optimization trajectories on test functions to illustrate how it improves convergence in handling heterogeneous curvatures. Extensive experiments on various models (RoBERTa, OPT, Phi-2 and LLama3, with 350M$\sim$66B parameters) indicate that HiZOO significantly reduces training steps and enhances model accuracy, while keeping the memory advantage of ZOO. For example, on SST2 task HiZOO achieves $8\times$ speedup and better accuracy over MeZO across different models. We also propose HiZOO-L, which reduces the Hessian memory cost to 10\% of the MeZO, while maintaining almost same performance. Compared with ZO-Adam, HiZOO-L achieves a 4.3\% improvement, just using 50\% of the GPU memory. Code is available at https://anonymous.4open.science/r/HiZOO-27F8.

Quantum-centric supercomputing presents a compelling framework for large-scale hybrid quantum-classical tasks. Although quantum machine learning (QML) offers theoretical benefits in various applications, challenges such as large-size data encoding in the input stage and the reliance on quantum resources in the inference stage limit its practicality for tasks like fine-tuning large language models (LLMs). Quantum parameter generation, a novel approach of QML, addresses these limitations by using quantum neural networks (QNNs) to generate classical model weights (parameters) exclusively during training, thereby decoupling inference from quantum hardware. In this work, we introduce Quantum Parameter Adaptation (QPA) in the framework of quantum parameter generation, which integrates QNNs with a classical multi-layer perceptron mapping model to generate parameters for fine-tuning methods. Using Gemma-2 and GPT-2 as case studies, QPA demonstrates significant parameter reduction for parameter-efficient fine-tuning methods, such as Low-Rank Adaptation (LoRA), while maintaining comparable or improved performance in text generation tasks. Specifically, QPA reduces the number of parameters to $52.06\%$ of the original LoRA for GPT-2 with a slight performance gain of $0.75\%$, and to $16.84\%$ for Gemma-2, with a marginal performance improvement of $0.07\%$. These results highlight QPA’s ability to achieve efficient parameter reduction without sacrificing performance in the quantum parameter generation framework. This work showcases the potential of quantum-enhanced parameter reduction, offering a scalable quantum-classical solution for fine-tuning LLMs while preserving the feasibility of inference on classical hardware.

Human sensing, which employs various sensors and advanced deep learning technologies to accurately capture and interpret human body information, has significantly impacted fields like public security and robotics. However, current human sensing primarily depends on modalities such as cameras and LiDAR, each of which has its own strengths and limitations. Furthermore, existing multimodal fusion solutions are typically designed for fixed modality combinations, requiring extensive retraining when modalities are added or removed for diverse scenarios. In this paper, we propose a modality-invariant foundation model for all modalities, X-Fi, to address these issues. X-Fi enables the independent or combinatory use of sensor modalities without additional training by utilizing a transformer structure to accommodate variable input sizes and incorporating a novel "X-fusion" mechanism to preserve modality-specific features during multimodal integration. This approach not only enhances adaptability but also facilitates the learning of complementary features across modalities. Extensive experiments conducted on the MM-Fi and XRF55 datasets, employing six distinct modalities, demonstrate that X-Fi achieves state-of-the-art performance in human pose estimation (HPE) and human activity recognition (HAR) tasks. The findings indicate that our proposed model can efficiently support a wide range of human sensing applications, ultimately contributing to the evolution of scalable, multimodal sensing technologies.

Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries---states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities. All our code is open-sourced at: https://github.com/conglu1997/intelligent-go-explore.

Parameter-efficient fine-tuning (PEFT) has revolutionized supervised fine-tuning, where LoRA and its variants gain the most popularity due to their low training costs and zero inference latency. However, LoRA tuning not only injects knowledgeable features but also noisy hallucination during fine-tuning, which hinders the utilization of tunable parameters with the increasing LoRA rank. In this work, we first investigate in-depth the redundancies among LoRA parameters with substantial empirical studies. Aiming to resemble the learning capacity of high ranks from the findings, we set up a new fine-tuning framework, \textbf{P}arameter-\textbf{Re}dundant \textbf{F}ine-\textbf{T}uning (\preft), which follows the vanilla LoRA tuning process but is required to reduce redundancies before merging LoRA parameters back to pre-trained models. Based on this framework, we propose \textbf{No}ise reduction with \textbf{R}eserved \textbf{M}ajority~(\norm), which decomposes the LoRA parameters into majority parts and redundant parts with random singular value decomposition. The major components are determined by the proposed \search method, specifically employing subspace similarity to confirm the parameter groups that share the highest similarity with the base weight. By employing \norm, we enhance both the learning capacity and benefits from larger ranks, which consistently outperforms both LoRA and other \preft-based methods on various downstream tasks, such as general instruction tuning, math reasoning and code generation. Code is available at \url{https://github.com/pixas/NoRM}.