ROSE: A Reward-Oriented Data Selection Framework for LLM Task-Specific Instruction Tuning

Instruction tuning has underscored the significant potential of large language models (LLMs) in producing more human-controllable and effective outputs in various domains. In this work, we focus on the data selection problem for task-specific instruction tuning of LLMs. Prevailing methods primarily rely on the crafted similarity metrics to select training data that aligns with the test data distribution. The goal is to minimize instruction tuning loss on the test data, ultimately improving performance on the target task. However, it has been widely observed that instruction tuning loss (i.e., cross-entropy loss for next token prediction) in LLMs often fails to exhibit a monotonic relationship with actual task performance. This misalignment undermines the effectiveness of current data selection methods for task-specific instruction tuning. To address this issue, we introduce ROSE, a novel Reward-Oriented inStruction data sElection method which leverages pairwise preference loss as a reward signal to optimize data selection for task-specific instruction tuning. Specifically, ROSE adapts an influence formulation to approximate the influence of training data points relative to a few-shot preference validation set to select the most task-related training data points. Experimental results show that by selecting just 5% of the training data using ROSE, our approach can achieve competitive results compared to fine-tuning with the full training dataset, and it surpasses other state-of-the-art data selection methods for task-specific instruction tuning. Our qualitative analysis further confirms the robust generalizability of our method across multiple benchmark datasets and diverse model architectures.

Correct after Answer: Enhancing Multi-Span Question Answering with Post-Processing Method

Multi-Span Question Answering (MSQA) requires models to extract one or multiple answer spans from a given context to answer a question. Prior work mainly focuses on designing specific methods or applying heuristic strategies to encourage models to predict more correct predictions. However, these models are trained on gold answers and fail to consider the incorrect predictions. Through a statistical analysis, we observe that models with stronger abilities do not predict less incorrect predictions compared with other models. In this work, we propose Answering-Classifying-Correcting (ACC) framework, which employs a post-processing strategy to handle incorrect predictions. Specifically, the ACC framework first introduces a classifier to classify the predictions into three types and exclude "wrong predictions", then introduces a corrector to modify "partially correct predictions". Experiments on several MSQA datasets show that ACC framework significantly improves the Exact Match (EM) scores, and further analysis demostrates that ACC framework efficiently reduces the number of incorrect predictions, improving the quality of predictions.

A Lightweight Multi Aspect Controlled Text Generation Solution For Large Language Models

Large language models (LLMs) show remarkable abilities with instruction tuning. However, they fail to achieve ideal tasks when lacking high-quality instruction tuning data on target tasks. Multi-Aspect Controllable Text Generation (MCTG) is a representative task for this dilemma, where aspect datasets are usually biased and correlated. Existing work exploits additional model structures and strategies for solutions, limiting adaptability to LLMs. To activate MCTG ability of LLMs, we propose a lightweight MCTG pipeline based on data augmentation. We analyze bias and correlations in traditional datasets, and address these concerns with augmented control attributes and sentences. Augmented datasets are feasible for instruction tuning. In our experiments, LLMs perform better in MCTG after data augmentation, with a 20% accuracy rise and less aspect correlations.

Towards Comprehensive Detection of Chinese Harmful Memes

This paper has been accepted in the NeurIPS 2024 D & B Track. Harmful memes have proliferated on the Chinese Internet, while research on detecting Chinese harmful memes significantly lags behind due to the absence of reliable datasets and effective detectors. To this end, we focus on the comprehensive detection of Chinese harmful memes. We construct ToxiCN MM, the first Chinese harmful meme dataset, which consists of 12,000 samples with fine-grained annotations for various meme types. Additionally, we propose a baseline detector, Multimodal Knowledge Enhancement (MKE), incorporating contextual information of meme content generated by the LLM to enhance the understanding of Chinese memes. During the evaluation phase, we conduct extensive quantitative experiments and qualitative analyses on multiple baselines, including LLMs and our MKE. The experimental results indicate that detecting Chinese harmful memes is challenging for existing models while demonstrating the effectiveness of MKE. The resources for this paper are available at https://github.com/DUT-lujunyu/ToxiCN_MM.

Aggregation of Multi Diffusion Models for Enhancing Learned Representations

Diffusion models have achieved remarkable success in image generation, particularly with the various applications of classifier-free guidance conditional diffusion models. While many diffusion models perform well when controlling for particular aspect among style, character, and interaction, they struggle with fine-grained control due to dataset limitations and intricate model architecture design. This paper introduces a novel algorithm, Aggregation of Multi Diffusion Models (AMDM), which synthesizes features from multiple diffusion models into a specified model, enhancing its learned representations to activate specific features for fine-grained control. AMDM consists of two key components: spherical aggregation and manifold optimization. Spherical aggregation merges intermediate variables from different diffusion models with minimal manifold deviation, while manifold optimization refines these variables to align with the intermediate data manifold, enhancing sampling quality. Experimental results demonstrate that AMDM significantly improves fine-grained control without additional training or inference time, proving its effectiveness. Additionally, it reveals that diffusion models initially focus on features such as position, attributes, and style, with later stages improving generation quality and consistency. AMDM offers a new perspective for tackling the challenges of fine-grained conditional control generation in diffusion models: We can fully utilize existing conditional diffusion models that control specific aspects, or develop new ones, and then aggregate them using the AMDM algorithm. This eliminates the need for constructing complex datasets, designing intricate model architectures, and incurring high training costs. Code is available at: https://github.com/Hammour-steak/AMDM

Enhanced Control for Diffusion Bridge in Image Restoration

Image restoration refers to the process of restoring a damaged low-quality image back to its corresponding high-quality image. Typically, we use convolutional neural networks to directly learn the mapping from low-quality images to high-quality images achieving image restoration. Recently, a special type of diffusion bridge model has achieved more advanced results in image restoration. It can transform the direct mapping from low-quality to high-quality images into a diffusion process, restoring low-quality images through a reverse process. However, the current diffusion bridge restoration models do not emphasize the idea of conditional control, which may affect performance. This paper introduces the ECDB model enhancing the control of the diffusion bridge with low-quality images as conditions. Moreover, in response to the characteristic of diffusion models having low denoising level at larger values of \(\bm t \), we also propose a Conditional Fusion Schedule, which more effectively handles the conditional feature information of various modules. Experimental results prove that the ECDB model has achieved state-of-the-art results in many image restoration tasks, including deraining, inpainting and super-resolution. Code is avaliable at https://github.com/Hammour-steak/ECDB.

Probing Causality Manipulation of Large Language Models

Large language models (LLMs) have shown various ability on natural language processing, including problems about causality. It is not intuitive for LLMs to command causality, since pretrained models usually work on statistical associations, and do not focus on causes and effects in sentences. So that probing internal manipulation of causality is necessary for LLMs. This paper proposes a novel approach to probe causality manipulation hierarchically, by providing different shortcuts to models and observe behaviors. We exploit retrieval augmented generation (RAG) and in-context learning (ICL) for models on a designed causality classification task. We conduct experiments on mainstream LLMs, including GPT-4 and some smaller and domain-specific models. Our results suggest that LLMs can detect entities related to causality and recognize direct causal relationships. However, LLMs lack specialized cognition for causality, merely treating them as part of the global semantic of the sentence.

Graph Transformers: A Survey

Graph transformers are a recent advancement in machine learning, offering a new class of neural network models for graph-structured data. The synergy between transformers and graph learning demonstrates strong performance and versatility across various graph-related tasks. This survey provides an in-depth review of recent progress and challenges in graph transformer research. We begin with foundational concepts of graphs and transformers. We then explore design perspectives of graph transformers, focusing on how they integrate graph inductive biases and graph attention mechanisms into the transformer architecture. Furthermore, we propose a taxonomy classifying graph transformers based on depth, scalability, and pre-training strategies, summarizing key principles for effective development of graph transformer models. Beyond technical analysis, we discuss the applications of graph transformer models for node-level, edge-level, and graph-level tasks, exploring their potential in other application scenarios as well. Finally, we identify remaining challenges in the field, such as scalability and efficiency, generalization and robustness, interpretability and explainability, dynamic and complex graphs, as well as data quality and diversity, charting future directions for graph transformer research.

Take its Essence, Discard its Dross! Debiasing for Toxic Language Detection via Counterfactual Causal Effect

Current methods of toxic language detection (TLD) typically rely on specific tokens to conduct decisions, which makes them suffer from lexical bias, leading to inferior performance and generalization. Lexical bias has both "useful" and "misleading" impacts on understanding toxicity. Unfortunately, instead of distinguishing between these impacts, current debiasing methods typically eliminate them indiscriminately, resulting in a degradation in the detection accuracy of the model. To this end, we propose a Counterfactual Causal Debiasing Framework (CCDF) to mitigate lexical bias in TLD. It preserves the "useful impact" of lexical bias and eliminates the "misleading impact". Specifically, we first represent the total effect of the original sentence and biased tokens on decisions from a causal view. We then conduct counterfactual inference to exclude the direct causal effect of lexical bias from the total effect. Empirical evaluations demonstrate that the debiased TLD model incorporating CCDF achieves state-of-the-art performance in both accuracy and fairness compared to competitive baselines applied on several vanilla models. The generalization capability of our model outperforms current debiased models for out-of-distribution data.

Enhancing Textual Personality Detection toward Social Media: Integrating Long-term and Short-term Perspectives

Textual personality detection aims to identify personality characteristics by analyzing user-generated content toward social media platforms. Numerous psychological literature highlighted that personality encompasses both long-term stable traits and short-term dynamic states. However, existing studies often concentrate only on either long-term or short-term personality representations, without effectively combining both aspects. This limitation hinders a comprehensive understanding of individuals' personalities, as both stable traits and dynamic states are vital. To bridge this gap, we propose a Dual Enhanced Network(DEN) to jointly model users' long-term and short-term personality for textual personality detection. In DEN, a Long-term Personality Encoding is devised to effectively model long-term stable personality traits. Short-term Personality Encoding is presented to capture short-term dynamic personality states. The Bi-directional Interaction component facilitates the integration of both personality aspects, allowing for a comprehensive representation of the user's personality. Experimental results on two personality detection datasets demonstrate the effectiveness of the DEN model and the benefits of considering both the dynamic and stable nature of personality characteristics for textual personality detection.