Large language models (LLMs) empowered by chain-of-thought reasoning have achieved impressive accuracy on complex tasks but suffer from excessive inference costs and latency when applied uniformly to all problems. We propose SABER (Switchable and Balanced Training for Efficient LLM Reasoning), a reinforcement learning framework that endows LLMs with user‑controllable, token‑budgeted reasoning. SABER first profiles each training example’s base‑model thinking token usage and assigns it to one of the predefined budget tiers. During fine‑tuning, the model is guided by system prompts and length‑aware rewards to respect its assigned budget. In parallel, we incorporate no‑think examples to ensure the model remains reliable even when explicit reasoning is turned off. SABER further supports four discrete inference modes—NoThink, FastThink, CoreThink, and DeepThink, enabling flexible trade‑offs between latency and reasoning depth. Extensive evaluations on math reasoning (MATH, GSM8K), code generation (MBPP), and logical reasoning (LiveBench-Reasoning) demonstrate that SABER achieves high accuracy under tight budgets, graceful degradation, and effective cross-scale and cross‑domain generalization. In particular, SABER‑FastThink cuts reasoning length by 65.4% and yields a 3.6% accuracy gain compared with the base model on the MATH benchmark.

Video Question-Answering (VideoQA) enables machines to interpret and respond to complex video content, advancing human-computer interaction. However, existing multimodal large language models (MLLMs) often provide incomplete or opaque explanations and existing benchmarks mainly focus on the correction of final answers, limiting insight into their reasoning processes and hindering both transparency and verifiability. To address this gap, we propose the Question Parsing, Video Alignment and Answer Aggregation framework (QPVA$^{3}$3), which leverages a compositional graph to drive visual and logical reasoning in VideoQA. Specifically, QPVA$^{3}$3 consists of three core components, the planner, executor, and reasoner to generate the compositional graph and conduct graph-driven reasoning. For the original question, the planner parses it into the compositional graph, capturing the underlying reasoning logic and structuring it into a series of interconnected questions. For each question in compositional graph, the executor aligns the video by selecting relevant video clips and generates answers, ensuring accurate, context-specific responses. For each question with its first-order descents, the reasoner aggregates answers by integrating reasoning logic with visual evidence, resolving conflicts to produce a coherent and accurate response. Moreover, to assess the performance of existing MLLMs in the reasoning processes of VideoQA, we introduce novel compositional consistency metrics and construct a VideoQA benchmark (QPVA$^{3}$3 Bench) with 3,492 question-video tuples, each annotated with detailed compositional graphs and fine-grained answers. We evaluate the QPVA$^{3}$3 framework on QPVA$^{3}$3 Bench and 5 other VideoQA benchmarks. Experimental results demonstrate that our framework improves both consistency and accuracy compared to baselines, leading to a more transparent and verifiable VideoQA system. This approach has the potential to advance the field, as supported by our comprehensive evaluation and benchmarking efforts.

Video Question-Answering (VideoQA) remains challenging in achieving advanced cognitive reasoning due to the uncontrollable and opaque reasoning processes in existing Multimodal Large Language Models (MLLMs). To address this issue, we propose a novel Language-centric Tree Reasoning (LTR) framework that targets on enhancing the reasoning ability of models. In detail, it recursively divides the original question into logically manageable parts and conquers them piece by piece, enhancing the reasoning capabilities and interpretability of existing MLLMs. Specifically, in the first stage, the LTR focuses on language to recursively generate a language-centric logical tree, which gradually breaks down the complex cognitive question into simple perceptual ones and plans the reasoning path through a RAG-based few-shot approach. In the second stage, with the aid of video content, the LTR performs bottom-up logical reasoning within the tree to derive the final answer along with the traceable reasoning path. Experiments across 11 VideoQA benchmarks demonstrate that our LTR framework significantly improves both accuracy and interpretability compared to state-of-the-art MLLMs. To our knowledge, this is the first work to implement a language-centric logical tree to guide MLLM reasoning in VideoQA, paving the way for language-centric video understanding from perception to cognition.

While large-scale language models (LLMs) have demonstrated remarkable capabilities in specific natural language processing (NLP) tasks, they may still lack proficiency compared to specialized models in certain domains, such as grammatical error correction (GEC). Drawing inspiration from the concept of curriculum learning, we have delved into refining LLMs into proficient GEC experts by devising effective curriculum learning (CL) strategies. In this paper, we introduce a novel approach, termed LLM-based curriculum learning, which capitalizes on the robust semantic comprehension and discriminative prowess inherent in LLMs to gauge the complexity of GEC training data. Unlike traditional curriculum learning techniques, our method closely mirrors human expert-designed curriculums. Leveraging the proposed LLM-based CL method, we sequentially select varying levels of curriculums ranging from easy to hard, and iteratively train and refine using the pretrianed T5 and LLaMA series models. Through rigorous testing and analysis across diverse benchmark assessments in English GEC, including the CoNLL14 test, BEA19 test, and BEA19 development sets, our approach showcases a significant performance boost over baseline models and conventional curriculum learning methodologies. Specifically, our method achieves a new state-of-the-art (SOTA) result on the CoNLL14 test set, with an F 0 . 5 score of 69.6. Additionally, on the BEA19 test set and BEA19 development set, our approach outperforms conventional curriculum learning methodologies by 1.0 and 0.3 F 0 . 5 points, respectively.

In industrial recommendation systems on websites and apps, it is essential to recall and predict top-n results relevant to user interests from a content pool of billions within milliseconds. To cope with continuous data growth and improve real-time recommendation performance, we have designed and implemented a high-performance batch query architecture for real-time recommendation systems. Our contributions include optimizing hash structures with a cacheline-aware probing method to enhance coalesced hashing, as well as the implementation of a hybrid storage key-value service built upon it. Our experiments indicate this approach significantly surpasses conventional hash tables in batch query throughput, achieving up to 90% of the query throughput of random memory access when incorporating parallel optimization. The support for NVMe, integrating two-tier storage for hot and cold data, notably reduces resource consumption. Additionally, the system facilitates dynamic updates, automated sharding of attributes and feature embedding tables, and introduces innovative protocols for consistency in batch queries, thereby enhancing the effectiveness of real-time incremental learning updates. This architecture has been deployed and in use in the bilibili recommendation system for over a year, a video content community with hundreds of millions of users, supporting 10x increase in model computation with minimal resource growth, improving outcomes while preserving the system's real-time performance.