Singular Value Decomposition (SVD) is a promising post-training compression technique for large language models (LLMs), as it is training-free, hardware-friendly, and preserves regular dense computation. However, existing SVD-based methods suffer from severe accuracy degradation under aggressive compression. This is mainly because hard rank truncation discards useful singular directions, while the two resulting SVD factors are usually compressed symmetrically despite their different roles, shapes, and quantization sensitivities. In this paper, we propose DELTA, an asymmetric low-rank compression method that decouples the latent heterogeneity between the two SVD factors. Under a fixed memory budget, the proposed method jointly determines the retained rank, factor-wise quantization bit-widths, and quantization directions, allowing more effective singular directions to be preserved with low-bit representations. To efficiently select compression configurations, we use an activation-aware reconstruction error as a layer-wise proxy objective. Experiments on multiple LLMs and datasets show that DELTA consistently outperforms existing post-training SVD-based compression methods, especially under high compression ratios. Compared with the strongest SVD-based baseline at the same compression ratio, DELTA reduces perplexity on WikiText-2 and C4 by up to 68.2% and 53.2%, respectively, and improves average zero-shot accuracy by up to 11 percentage points. These results demonstrate that asymmetric factor-aware compression can effectively mitigate the accuracy degradation of SVD-based LLM compression at high compression ratios. Code is available at https://github.com/BlacksheepJane/DELTA.

Leveraging knowledge from multiple tasks through introducing a small number of task specific parameters into each transformer layer, also known as adapters, receives much attention recently. However, adding an extra fusion layer to implement knowledge composition not only increases the inference time but also is non-scalable for some applications. To avoid these issues, we propose a two-stage knowledge distillation algorithm called AdapterDistillation. In the first stage, we extract task specific knowledge by using local data to train a student adapter. In the second stage, we distill the knowledge from the existing teacher adapters into the student adapter to help its inference. Extensive experiments on frequently asked question retrieval in task-oriented dialog systems validate the efficiency of AdapterDistillation. We show that AdapterDistillation outperforms existing algorithms in terms of accuracy, resource consumption and inference time.

We propose a new paradigm for zero-shot learners that is format agnostic, i.e., it is compatible with any format and applicable to a list of language tasks, such as text classification, commonsense reasoning, coreference resolution, and sentiment analysis. Zero-shot learning aims to train a model on a given task such that it can address new learning tasks without any additional training. Our approach converts zero-shot learning into multiple-choice tasks, avoiding problems in commonly used large-scale generative models such as FLAN. It not only adds generalization ability to models but also significantly reduces the number of parameters. Our method shares the merits of efficient training and deployment. Our approach shows state-of-the-art performance on several benchmarks and produces satisfactory results on tasks such as natural language inference and text classification. Our model achieves this success with only 235M parameters, which is substantially smaller than state-of-the-art models with billions of parameters. The code and pre-trained models are available at https://github.com/IDEA-CCNL/Fengshenbang-LM/tree/main/fengshen/examples/unimc .

In Visual Question Answering (VQA), existing bilinear methods focus on the interaction between images and questions. As a result, the answers are either spliced into the questions or utilized as labels only for classification. On the other hand, trilinear models such as the CTI model efficiently utilize the inter-modality information between answers, questions, and images, while ignoring intra-modality information. Inspired by this observation, we propose a new trilinear interaction framework called MIRTT (Learning Multimodal Interaction Representations from Trilinear Transformers), incorporating the attention mechanisms for capturing inter-modality and intra-modality relationships. Moreover, we design a two-stage workflow where a bilinear model reduces the free-form, open-ended VQA problem into a multiple-choice VQA problem. Furthermore, to obtain accurate and generic multimodal representations, we pre-train MIRTT with masked language prediction. Our method achieves state-of-the-art performance on the Visual7W Telling task and VQA-1.0 Multiple Choice task and outperforms bilinear baselines on the VQA-2.0, TDIUC and GQA datasets.