Instruction tuning (IT) improves large language models (LLMs) by aligning their outputs with human instructions, but its success depends critically on training data quality, and datasets such as Alpaca often contain noisy or suboptimal examples that undermine fine-tuning.Prior selection strategies score samples using general-purpose LLMs (e.g., GPT), leveraging their strong language understanding yet introducing inherent biases that misalign with the target model's behavior and yield unstable downstream performance.Influence-based methods address this by estimating each example's marginal contribution to overall performance, but they typically assume additive contributions and therefore overlook higher-order interactions among samples.To overcome these limitations, we propose JI 2 S, a novel framework that jointly models both marginal and combinatorial influences within sample groups.Applying JI 2 S to select the top 1,000 most influential examples from Alpaca, we fine-tune LLaMA2-7B, Mistral-7B, and LLaMA2-13B and evaluate them on Open LLM Benchmarks, MT-Bench, and GPT-4-judged pairwise comparisons.Our experiments show that JI 2 S consistently outperforms full-dataset training and strong baselines, highlighting the value of capturing joint influence for high-quality instruction fine-tuning.We provide our code in this GitHub repository.

Unsupervised text style transfer aims to modify the style of a sentence while preserving its content without parallel corpora. Existing approaches attempt to separate content from style, but some words contain both content and style information. It makes them difficult to disentangle, where unsatisfactory disentanglement results in the loss of the content information or the target style. To address this issue, researchers adopted a “cycle reconstruction” mechanism to maintain content information, but it is still hard to achieve satisfactory content preservation due to incomplete disentanglement. In this paper, we propose a new disentanglement-based method, StyleFlow, which effectively avoids the loss of contents through a better cycle reconstruction via a reversible encoder. The reversible encoder is a normalizing flow that can not only produce output given input but also infer the exact input given the output reversely. We design a stack of attention-aware coupling layers, where each layer is reversible and adopts the attention mechanism to improve the content-style disentanglement. Moreover, we propose a data augmentation method based on normalizing flow to enhance the training data. Our experiments on sentiment transfer and formality transfer tasks show that StyleFlow outperforms strong baselines on both content preservation and style transfer.

Multimodal Large Language Models (MLLMs), by expanding the model's capabilities to perceive and interact through multi-modalities, have significantly enhanced performance across various tasks. The perception of vision is an important modality developed into LLM, enabling research in vision-language to continuously lead the cutting-edge advancements in the MLLM community. However, the standard pre-training pipeline on image-text pairs results in a limited model understanding of relationships between multiple images and texts, as well as visual details. Additionally, the setting of fine-tuning with a frozen visual backbone hinders the enhancement of visual representations on new data. These two issues lead to suboptimal performance in models for demonstrative instruction following about multiple images. This work introduces a novel framework called MLoEM, which first converts long multimodal data into an interleaved image-instruction format, and then adopts a fully autoregressive architecture model, allowing for more robust and coherent learning from naturally occurring multimodal documents than pair-based pipeline. Additionally, we incorporate the Low-Rank Adaptation (LoRA) fine-tuning method, enhancing visual representations while maintaining the stability of previously learned knowledge. Finally, we utilize ensemble methods to enhance model performance on tasks. To alleviate the storage overhead issue of parallel ensembles with large models, we design an ensemble approach that shares the MLLM while only switching the LoRA matrices. In the experiments, the proposed MLoEM shows superior performance on the testing set.