Current Multimodal Chain-of-Thought (MCoT) methods suffer from low-quality multimodal reasoning, characterized by overthinking on simple queries and inefficient utilization of visual information, resulting in vast inefficient and ineffective computations. In this paper, we discover that Multimodal Large Language Models (MLLMs) possess inherent capabilities to distinguish between simple and difficult queries and enhance task-related visual information, which remain underutilized by existing approaches. Based on this insight, we propose Self-Driven Refined Multimodal CoT (SDR-MCoT), a training-free framework that mitigates these issues through two self-driven modules. First, our selective thinking module employs entropy-based confidence estimation to determine whether queries require detailed reasoning, preventing overthinking on simple questions. Second, our step-wise visual enhancement module strengthens attention to relevant visual regions at each reasoning step without inserting additional tokens, achieving fine-grained visual grounding and enhancement with minimal overhead. Moreover, SDR-MCoT can be seamlessly integrated into various MLLMs, offering a practical solution for improving multimodal reasoning. Comprehensive experiments across eight benchmarks from diverse domains (multimodal reasoning, visual understanding, hallucination, and mathematical reasoning) demonstrate that SDR-MCoT consistently outperforms existing MCoT methods on four different base models with reduced overhead. For instance, on Qwen2-VL-7B, our method improves average accuracy by over 6% while reducing token consumption by approximately 60% compared to zero-shot CoT.

While Large Language Models (LLMs) demonstrate strong performance across domains, their long-context capabilities are limited by transient neural activations causing information decay and unstructured feed-forward network (FFN) weights leading to semantic fragmentation. Inspired by the brain’s working memory and cortical modularity, we propose PaceLLM, featuring two innovations: (1) a Persistent Activity (PA) Mechanism that mimics prefrontal cortex (PFC) neurons’ persistent firing by introducing an activation-level memory bank to dynamically retrieve, reuse, and update critical FFN states, addressing contextual decay; and (2) Cortical Expert (CE) Clustering that emulates task-adaptive neural specialization to reorganize FFN weights into semantic modules, establishing cross-token dependencies and mitigating fragmentation. Extensive evaluations show that PaceLLM achieves 6% improvement on LongBench’s Multi-document QA and 12.5–17.5% performance gains on $\infty$-Bench tasks, while extending measurable context length to 200K tokens in Needle-In-A-Haystack (NIAH) tests. This work pioneers brain-inspired LLM optimization and is complementary to other works. Besides, it can be generalized to any model and enhance their long-context performance and interpretability without structural overhauls.

Multimodal Large Language Models (MLLMs) have shown impressive video content understanding capabilities but struggle with fine-grained motion comprehension. To comprehensively assess the motion understanding ability of existing MLLMs, we introduce FAVOR-Bench, which comprises 1,776 videos from both ego-centric and third-person perspectives and enables assessment through both close-ended and open-ended tasks. For close-ended evaluation, we carefully design 8,184 multiple-choice question-answer pairs spanning six distinct sub-tasks. For open-ended evaluation, we employ the GPT-assisted evaluation and develop a novel cost-efficient LLM-free assessment method, where the latter can enhance benchmarking interpretability and accessibility. Comprehensive experiments with21 state-of-the-art MLLMs reveal significant limitations in their ability to comprehend and describe detailed temporal dynamics in video motions. To alleviate this limitation, we further build FAVOR-Train, a dataset of 17,152 videos with fine-grained motion annotations. Finetuning Qwen2.5-VL on FAVOR-Train yields consistent improvements on motion-related tasks across TVBench, MotionBenchand our FAVOR-Bench. Our assessment results demonstrate that the proposed FAVOR-Bench and FAVOR-Train provide valuable tools for the community to develop more powerful video understanding models.

As Vision Transformers (ViTs) become increasingly popular in various vision tasks, one may question:if a new training scheme for ViTs exists that can improve performance without increasing training and inference computation cost?In this paper, we affirmatively answer this question with a novel Sparse-to-Dense (S2D) training scheme. Specifically, we decouple the training and inference phases of ViTs. During training, we replace some Feed-Forward Network (FFN) layers of ViTs with computationally efficient RUP-Mixture-of-FFN (RUP-MoF) layers, each comprising multiple FFN experts and allocating tokens to experts via Random Uniform Partition (RUP). Furthermore, an additional Experts Weights Averaging (EWA) update is performed specifically on these RUP-MoF layers after each gradient update. After training, we convert each RUP-MoF layer back to a single FFN layer by averaging the experts, transforming the training-time sparse model back to the original dense ViT model for inference. We further provide theoretical analysis to illustrate why and how it works. Comprehensive experiments across various 2D and 3D vision tasks, ViT architectures and datasets validate the effectiveness and generalization ability of the proposed S2D training scheme. Besides, we show that, S2D training scheme can also be applied to improve the performance of Transformer-based language models, and EWA update technique can also significantly improve the effectiveness of classic Mixture-of-Experts on various 2D vision small-scale datasets and 3D vision tasks.

Improving the efficiency of Neural Architecture Search (NAS) is a challenging but significant task that has received much attention. Previous studies mainly adopt the Differentiable Architecture Search (DARTS) and improve its search strategies or modules to enhance search efficiency. Recently, some methods have started considering data reduction for speedup, but they are not tightly coupled with the architecture search process and cannot capture the training dynamics of DARTS well, resulting in sub-optimal performances. To this end, this work pioneers an exploration into the critical role of dataset characteristics in the bi-level optimization of DARTS, and then proposes a novel Bi-level Data Pruning (BDP) paradigm that targets the weights and architecture levels of DARTS to enhance efficiency from a data perspective. Specifically, we introduce a progressive bi-level data pruning strategy that utilizes supernet prediction dynamics as the metric to gradually prune unsuitable samples for DARTS during the search. An effective automatic class balance constraint is also integrated into BDP, to suppress potential class imbalances resulting from data-efficient algorithms. Comprehensive evaluations on the NAS-Bench-201 search space, DARTS search space, and MobileNet-like search space validate that BDP reduces search costs by over 50% while achieving superior performance when applied to the baseline DARTS. Besides, we demonstrate that BDP can harmoniously integrate with advanced DARTS variants, like P-DARTS, PC-DARTS, EG-NAS, and$\beta $-DARTS, offering an approximately$2\times $speedup with minimal performance compromise.