Empowering LLMs with the ability to precisely understand long contexts is crucial for many downstream applications. However, handling long contexts with conventional transformer architecture requires substantial training and inference resources. Existing context condensing methods cannot accurately understand the full context, as there is a considerable amount of information loss in the condensing process. To address these issues, we present **FocusLLM**, a framework designed to extend the fixed context length of any decoder-only LLM, allowing the model to focus on relevant information from very long sequences. FocusLLM first divides long text input into chunks based on the model’s original context length. It then employs the **_dynamic condensing_** process to distill crucial information from each chunk. Ultimately, through the novel **_parallel decoding_** mechanism, FocusLLM can integrate the extracted information into its local context. FocusLLM stands out for great training efficiency and versatility: trained with an 8K input length and with much less training cost than previous methods, FocusLLM exhibits superior performance across downstream tasks and maintains strong language modeling ability when handling extensive long texts, even up to 400K tokens. Our code is available at https://github.com/leezythu/FocusLLM.

Effective content moderation is essential for video platforms to safeguard user experience and uphold community standards. While traditional video classification models effectively handle well-defined moderation tasks, they struggle with complicated scenarios such as implicit harmful content and contextual ambiguity. Multimodal large language models (MLLMs) offer a promising solution to these limitations with their superior cross-modal reasoning and contextual understanding. However, two key challenges hinder their industrial adoption. First, the high computational cost of MLLMs makes full-scale deployment impractical. Second, adapting generative models for discriminative classification remains an open research problem. In this paper, we first introduce an efficient method to transform a generative MLLM into a multimodal classifier using minimal discriminative training data. To enable industry-scale deployment, we then propose a router-ranking cascade system that integrates MLLMs with a lightweight router model. Offline experiments demonstrate that our MLLM-based approach improves F1 score by 66.50% over traditional classifiers while requiring only 2% of the fine-tuning data. Online evaluations show that our system increases automatic content moderation volume by 41%, while the cascading deployment reduces computational cost to only 1.5% of direct full-scale deployment.

This paper presents Tencent’s submission to the WMT20 Quality Estimation (QE) Shared Task: Sentence-Level Post-editing Effort for English-Chinese in Task 2. Our system ensembles two architectures, XLM-based and Transformer-based Predictor-Estimator models. For the XLM-based Predictor-Estimator architecture, the predictor produces two types of contextualized token representations, i.e., masked XLM and non-masked XLM; the LSTM-estimator and Transformer-estimator employ two effective strategies, top-K and multi-head attention, to enhance the sentence feature representation. For Transformer-based Predictor-Estimator architecture, we improve a top-performing model by conducting three modifications: using multi-decoding in machine translation module, creating a new model by replacing the transformer-based predictor with XLM-based predictor, and finally integrating two models by a weighted average. Our submission achieves a Pearson correlation of 0.664, ranking first (tied) on English-Chinese.