Choosing the right large language model (LLM) backbone is the most consequential decision when building a vision-language model (VLM), yet it remains fundamentally unprincipled: compute-based scaling laws fail to generalize across model families, and no framework exists for directly predicting VLM performance before training begins. We propose the Capability-Driven Multimodal Scaling Law, the first cross-family framework that predicts VLM benchmark accuracy from directly observable textual capability. Given a low-dimensional capability score S extracted from LLM textual benchmarks via PCA, we model VLM performance as a function of S, with a per-backbone transfer rate A and an absorption rate B that quantifies data-scaling efficiency. To fit and validate the framework, we train over 150 VLMs on 34 LLMs spanning 7 model families under a strictly controlled recipe. Evaluations on more than 200 textual and 50 multimodal benchmarks show that the law accurately extrapolates transfer rate from models up to 8B parameters to 72B-scale backbones, predicts full VLM training trajectories with high fidelity, and generalizes to entirely held-out model families. Beyond the scaling law, our analysis surfaces actionable insights: certain textual benchmarks negatively correlate with multimodal performance, exposing latent benchmark-gaming behavior; base LLMs outperform instruction-tuned counterparts as VLM backbones due to higher absorption rates and lower data-scaling decay; and different model families occupy distinct positions in the transfer–absorption space. The framework turns backbone selection from costly empirical sweeps into a principled, quantitative decision.

Existing approaches in the vision-and-language pre-training (VLP) paradigm mainly deploy either fusion-based encoders or dual-encoders, failing to achieve both effectiveness and efficiency in downstream multimodal tasks. In this paper, we build a flexible VLP model by incorporating cross-modal fusions into a dual-encoder architecture, where the introduced fusion modules can be easily decoupled from the dual encoder so as to switch the model to a fusion-free one. To better absorb cross-modal features from the fusion modules, we design a cross-modal knowledge transfer strategy along with other comprehensive pre-training tasks to guide the training process, which can further strengthen both the fusion-based and fusion-free representation learning. Extensive experiments conducted on various downstream vision-language tasks show that our proposed model is well-equipped with effectiveness as well as efficiency, demonstrating a superior performance compared with other strong VLP models.

Dialogue Act Recognition (DAR) is a challenging problem in Natural Language Understanding, which aims to attach Dialogue Act (DA) labels to each utterance in a conversation. However, previous studies cannot fully recognize the specific expressions given by users due to the informality and diversity of natural language expressions. To solve this problem, we propose a Heterogeneous User History (HUH) graph convolution network, which utilizes the user’s historical answers grouped by DA labels as additional clues to recognize the DA label of utterances. To handle the noise caused by introducing the user’s historical answers, we design sets of denoising mechanisms, including a History Selection process, a Similarity Re-weighting process, and an Edge Re-weighting process. We evaluate the proposed method on two benchmark datasets MSDialog and MRDA. The experimental results verify the effectiveness of integrating user’s historical answers, and show that our proposed model outperforms the state-of-the-art methods.

Chinese relation extraction is conducted using neural networks with either character-based or word-based inputs, and most existing methods typically suffer from segmentation errors and ambiguity of polysemy. To address the issues, we propose a multi-grained lattice framework (MG lattice) for Chinese relation extraction to take advantage of multi-grained language information and external linguistic knowledge. In this framework, (1) we incorporate word-level information into character sequence inputs so that segmentation errors can be avoided. (2) We also model multiple senses of polysemous words with the help of external linguistic knowledge, so as to alleviate polysemy ambiguity. Experiments on three real-world datasets in distinct domains show consistent and significant superiority and robustness of our model, as compared with other baselines. We will release the source code of this paper in the future.

Event detection (ED) aims to locate trigger words in raw text and then classify them into correct event types. In this task, neural net- work based models became mainstream in re- cent years. However, two problems arise when it comes to languages without natural delim- iters, such as Chinese. First, word-based mod- els severely suffer from the problem of word- trigger mismatch, limiting the performance of the methods. In addition, even if trigger words could be accurately located, the ambi- guity of polysemy of triggers could still af- fect the trigger classification stage. To ad- dress the two issues simultaneously, we pro- pose the Trigger-aware Lattice Neural Net- work (TLNN). (1) The framework dynami- cally incorporates word and character informa- tion so that the trigger-word mismatch issue can be avoided. (2) Moreover, for polysemous characters and words, we model all senses of them with the help of an external linguistic knowledge base, so as to alleviate the prob- lem of ambiguous triggers. Experiments on two benchmark datasets show that our model could effectively tackle the two issues and outperforms previous state-of-the-art methods significantly, giving the best results. The source code of this paper can be obtained from https://github.com/thunlp/TLNN.