Learning with Less: Knowledge Distillation from Large Language Models via Unlabeled Data

In real-world NLP applications, Large Language Models (LLMs) offer promising solutions due to their extensive training on vast datasets. However, the large size and high computation demands of LLMs limit their practicality in many applications, especially when further fine-tuning is required. To address these limitations, smaller models are typically preferred for deployment. However, their training is hindered by the scarcity of labeled data. In contrast, unlabeled data is often readily which can be leveraged by using LLMs to generate pseudo-labels for training smaller models. This enables the smaller models (student) to acquire knowledge from LLMs(teacher) while reducing computational costs. This process introduces challenges, such as potential noisy pseudo-labels. Selecting high-quality and informative data is therefore critical to enhance model performance while improving the efficiency of data utilization. To address this, we propose LLKD that enables Learning with Less computational resources and less data for Knowledge Distillation from LLMs. LLKD is an adaptive sample selection method that incorporates signals from both the teacher and student. Specifically, it prioritizes samples where the teacher demonstrates high confidence in its labeling, indicating reliable labels, and where the student exhibits a high information need, identifying challenging samples that require further learning. Our comprehensive experiments show that LLKD achieves superior performance across various datasets with higher data efficiency.

HC$^3$L-Diff: Hybrid conditional latent diffusion with high frequency enhancement for CBCT-to-CT synthesis

Background: Cone-beam computed tomography (CBCT) plays a crucial role in image-guided radiotherapy, but artifacts and noise make them unsuitable for accurate dose calculation. Artificial intelligence methods have shown promise in enhancing CBCT quality to produce synthetic CT (sCT) images. However, existing methods either produce images of suboptimal quality or incur excessive time costs, failing to satisfy clinical practice standards. Methods and materials: We propose a novel hybrid conditional latent diffusion model for efficient and accurate CBCT-to-CT synthesis, named HC$^3$L-Diff. We employ the Unified Feature Encoder (UFE) to compress images into a low-dimensional latent space, thereby optimizing computational efficiency. Beyond the use of CBCT images, we propose integrating its high-frequency knowledge as a hybrid condition to guide the diffusion model in generating sCT images with preserved structural details. This high-frequency information is captured using our designed High-Frequency Extractor (HFE). During inference, we utilize denoising diffusion implicit model to facilitate rapid sampling. We construct a new in-house prostate dataset with paired CBCT and CT to validate the effectiveness of our method. Result: Extensive experimental results demonstrate that our approach outperforms state-of-the-art methods in terms of sCT quality and generation efficiency. Moreover, our medical physicist conducts the dosimetric evaluations to validate the benefit of our method in practical dose calculation, achieving a remarkable 93.8% gamma passing rate with a 2%/2mm criterion, superior to other methods. Conclusion: The proposed HC$^3$L-Diff can efficiently achieve high-quality CBCT-to-CT synthesis in only over 2 mins per patient. Its promising performance in dose calculation shows great potential for enhancing real-world adaptive radiotherapy.

PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference

In this paper, we make the first attempt to align diffusion models for image inpainting with human aesthetic standards via a reinforcement learning framework, significantly improving the quality and visual appeal of inpainted images. Specifically, instead of directly measuring the divergence with paired images, we train a reward model with the dataset we construct, consisting of nearly 51,000 images annotated with human preferences. Then, we adopt a reinforcement learning process to fine-tune the distribution of a pre-trained diffusion model for image inpainting in the direction of higher reward. Moreover, we theoretically deduce the upper bound on the error of the reward model, which illustrates the potential confidence of reward estimation throughout the reinforcement alignment process, thereby facilitating accurate regularization. Extensive experiments on inpainting comparison and downstream tasks, such as image extension and 3D reconstruction, demonstrate the effectiveness of our approach, showing significant improvements in the alignment of inpainted images with human preference compared with state-of-the-art methods. This research not only advances the field of image inpainting but also provides a framework for incorporating human preference into the iterative refinement of generative models based on modeling reward accuracy, with broad implications for the design of visually driven AI applications. Our code and dataset are publicly available at https://prefpaint.github.io.

Irregular Tensor Low-Rank Representation for Hyperspectral Image Representation

Spectral variation is a common problem for hyperspectral image (HSI) representation. Low-rank tensor representation is an important approach to alleviate spectral variations. However, the spatial distribution of the HSI is always irregular, while the previous tensor low-rank representation methods can only be applied to the regular data cubes, which limits the performance. To remedy this issue, in this paper we propose a novel irregular tensor low-rank representation model. We first segment the HSI data into several irregular homogeneous regions. Then, we propose a novel irregular tensor low-rank representation method that can efficiently model the irregular 3D cubes. We further use a non-convex nuclear norm to pursue the low-rankness and introduce a negative global low-rank term that improves global consistency. This proposed model is finally formulated as a convex-concave optimization problem and solved by alternative augmented Lagrangian method. Through experiments on four public datasets, the proposed method outperforms the existing low-rank based HSI methods significantly. Code is available at: https://github.com/hb-studying/ITLRR.

SimRAG: Self-Improving Retrieval-Augmented Generation for Adapting Large Language Models to Specialized Domains

Retrieval-augmented generation (RAG) enhances the question-answering (QA) abilities of large language models (LLMs) by integrating external knowledge. However, adapting general-purpose RAG systems to specialized fields such as science and medicine poses unique challenges due to distribution shifts and limited access to domain-specific data. To tackle this, we propose SimRAG, a self-training approach that equips the LLM with joint capabilities of question answering and question generation for domain adaptation. Our method first fine-tunes the LLM on instruction-following, question-answering, and search-related data. Then, it prompts the same LLM to generate diverse domain-relevant questions from unlabeled corpora, with an additional filtering strategy to retain high-quality synthetic examples. By leveraging these synthetic examples, the LLM can improve their performance on domain-specific RAG tasks. Experiments on 11 datasets, spanning two backbone sizes and three domains, demonstrate that SimRAG outperforms baselines by 1.2\%--8.6\%.

Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. To thoroughly evaluate this phenomenon, we conduct comprehensive experiments using various quantization techniques across multiple precision levels. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. Based on our empirical findings, we provide a theoretical explanation for the observed phenomenon and propose a quantization-robust unlearning strategy to mitigate this intricate issue...

Exploring Social Desirability Response Bias in Large Language Models: Evidence from GPT-4 Simulations

Large language models (LLMs) are employed to simulate human-like responses in social surveys, yet it remains unclear if they develop biases like social desirability response (SDR) bias. To investigate this, GPT-4 was assigned personas from four societies, using data from the 2022 Gallup World Poll. These synthetic samples were then prompted with or without a commitment statement intended to induce SDR. The results were mixed. While the commitment statement increased SDR index scores, suggesting SDR bias, it reduced civic engagement scores, indicating an opposite trend. Additional findings revealed demographic associations with SDR scores and showed that the commitment statement had limited impact on GPT-4's predictive performance. The study underscores potential avenues for using LLMs to investigate biases in both humans and LLMs themselves.

Divide-Verify-Refine: Aligning LLM Responses with Complex Instructions

Recent studies show that LLMs, particularly open-source models, struggle to follow complex instructions with multiple constraints. Despite the importance, methods to improve LLMs' adherence to such constraints remain unexplored, and current research focuses on evaluating this ability rather than developing solutions. While a few studies enhance constraint adherence through model tuning, this approach is computationally expensive and heavily reliant on training data quality. An alternative is to leverage LLMs' self-correction capabilities, allowing them to adjust responses to better meet specified constraints. However, this self-correction ability of LLMs is limited by the feedback quality, as LLMs cannot autonomously generate reliable feedback or detect errors. Moreover, the self-refinement process heavily depends on few-shot examples that illustrate how to modify responses to meet constraints. As constraints in complex instructions are diverse and vary widely, manually crafting few-shot examples for each constraint type can be labor-intensive and sub-optimal. To deal with these two challenges, we propose the Divide-Verify-Refine (DVR) framework with three steps: (1) Divide complex instructions into single constraints and prepare appropriate tools; (2) Verify: To address the feedback quality problem, these tools will rigorously verify responses and provide reliable feedback; (3) Refine: To address the constraint diversity challenge, we design a refinement repository that collects successful refinement processes and uses them as few-shot demonstrations for future cases, allowing LLMs to learn from the past experience during inference. Additionally, we develop a new dataset of complex instructions, each containing 1-6 constraints. Experiments show that the framework significantly improves performance, doubling LLama3.1-8B's constraint adherence on instructions with 6 constraints.

Towards the Effect of Examples on In-Context Learning: A Theoretical Case Study

In-context learning (ICL) has emerged as a powerful capability for large language models (LLMs) to adapt to downstream tasks by leveraging a few (demonstration) examples. Despite its effectiveness, the mechanism behind ICL remains underexplored. To better understand how ICL integrates the examples with the knowledge learned by the LLM during pre-training (i.e., pre-training knowledge) and how the examples impact ICL, this paper conducts a theoretical study in binary classification tasks. In particular, we introduce a probabilistic model extending from the Gaussian mixture model to exactly quantify the impact of pre-training knowledge, label frequency, and label noise on the prediction accuracy. Based on our analysis, when the pre-training knowledge contradicts the knowledge in the examples, whether ICL prediction relies more on the pre-training knowledge or the examples depends on the number of examples. In addition, the label frequency and label noise of the examples both affect the accuracy of the ICL prediction, where the minor class has a lower accuracy, and how the label noise impacts the accuracy is determined by the specific noise level of the two classes. Extensive simulations are conducted to verify the correctness of the theoretical results, and real-data experiments also align with the theoretical insights. Our work reveals the role of pre-training knowledge and examples in ICL, offering a deeper understanding of LLMs' behaviors in classification tasks.

Learning Efficient and Effective Trajectories for Differential Equation-based Image Restoration

The differential equation-based image restoration approach aims to establish learnable trajectories connecting high-quality images to a tractable distribution, e.g., low-quality images or a Gaussian distribution. In this paper, we reformulate the trajectory optimization of this kind of method, focusing on enhancing both reconstruction quality and efficiency. Initially, we navigate effective restoration paths through a reinforcement learning process, gradually steering potential trajectories toward the most precise options. Additionally, to mitigate the considerable computational burden associated with iterative sampling, we propose cost-aware trajectory distillation to streamline complex paths into several manageable steps with adaptable sizes. Moreover, we fine-tune a foundational diffusion model (FLUX) with 12B parameters by using our algorithms, producing a unified framework for handling 7 kinds of image restoration tasks. Extensive experiments showcase the significant superiority of the proposed method, achieving a maximum PSNR improvement of 2.1 dB over state-of-the-art methods, while also greatly enhancing visual perceptual quality. Project page: \url{https://zhu-zhiyu.github.io/FLUX-IR/}.