As applications of large language models (LLMs) become increasingly complex, the demand for robust complex instruction following capabilities is growing accordingly.We argue that a thorough understanding of the instruction itself, especially the latent reasoning structure embedded between the lines, is crucial for improving instruction following.Therefore we target complex instructions that involve implicit reasoning, intricate logical relations, and multiconstraint dependencies.We propose ImpRIF, a method to enhance LLMs' understanding of implicit reasoning instructions, thereby improving its ability to follow complex instructions.We formalize such instructions as verifiable reasoning graphs, enabling programmatic verification and graph-driven chain-of-thought reasoning.Based on this formulation, we synthesize large-scale single-and multi-turn data, propose fine-tuning with graph reasoning, and apply reinforcement learning to explicitly train models to reason along the graph.On five complex instruction following benchmarks, our models substantially outperform their base models.These results demonstrate that enhancing implicit reasoning capabilities can significantly improve complex instruction following.

Traffic sign detection is crucial for both autonomous and manual driving, but high-resolution images from in-vehicle cameras pose challenges due to the small size of distant objects and the substantial computational cost. Due to the limited computational and parallel processing capabilities of in-vehicle computing platforms, achieving real-time, high-precision traffic sign detection in practical scenarios based on high-resolution images becomes impractical. Therefore, we propose a lightweight subgraph-based deep learning traffic sign detection framework tailored for in-vehicle embedded computing devices. The framework comprises two sub-models: 1) a Subgraph Filtering Model based on lightweight Convolutional Neural Networks, and 2) a Lightweight Deep Learning Traffic Sign Detection Model. The proposed framework enables to input high-resolution images while significantly reducing computational costs and produces high-precision traffic sign detection results. Experimental results on the TT100K dataset show an [email protected] of 0.954 with only 0.93M parameters, demonstrating the framework’s ability to perform real-time, high-precision traffic sign detection on in-vehicle embedded devices.

Few-shot class-incremental learning (FSCIL) presents significant challenges due to the constraints of the sequential learning framework, which limits the availability of training samples for new classes and prohibits revisiting samples from previously learned classes. To address these challenges, we propose a novel approach that leverages dynamic classifier weights and gradient backpropagation from pseudo-features of all encountered classes to classify incrementally appearing content. By directly synthesizing features for previously learned classes, we address the issue of catastrophic forgetting. Moreover, generating features for current classes helps mitigate overfitting, especially given the scarcity of original features. Our method utilizes pseudo-features to form deformable weights for each class, recognizing that optimal weights can vary across sessions, even for the same class. Unlike prior approaches that rely on fixed classification weights per class or generative models with pre-calculated parameters (e.g., GAN), which may produce sub-optimal pseudo-features for new classes, we introduce an updated sparse reconstruction method. This method generates pseudo-features through Sampling and Reconstruction (S&R). Extensive experiments on benchmark datasets, including mini-ImageNet, CUB-200, and CIFAR-100, demonstrate that our method outperforms state-of-the-art approaches under comparable experimental settings, highlighting its effectiveness in addressing the challenges of FSCIL.