This paper proposes shortcut decoding, an efficient framework for accelerating Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs).Existing methods that prune or employ early stopping to reduce latency often compromise reasoning reliability.Motivated by the observation that LLMs frequently converge to correct solutions internally before completing explicit textual reasoning, we propose a dual-signal adaptive controller that integrates lightweight probes over internal hidden states with step-level entropy.This controller detects convergence of reasoning during generation and adaptively selects between a fastexit path and a stability-verified path to remove redundant steps while preserving answer correctness.Experiments across multiple mathematical reasoning benchmarks demonstrate that shortcut decoding reduces token usage by approximately 35%, maintains accuracy comparable to full CoT decoding, and achieves finalanswer accuracy comparable to the full CoT baseline, outperforming existing early-stopping methods without updating the base model.Our code is available at https://github.com/ kuromi9527/shortcut_decoding.

Zihan Zheng, Tianle Cui, Taoran Wang, Fengtao Wang, Jiahui Pan, Lewei He, Qianglong Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Model (LLM)-based agentic systems have shown strong capabilities across various tasks. However, existing multi-agent frameworks often rely on static or task-level workflows, which either over-process simple queries or underperform on complex ones, while also neglecting the efficiency-performance trade-offs across heterogeneous LLMs. To address these limitations, we propose Difficulty-Aware Agentic Orchestration (DAAO), which can dynamically generate query-specific multi-agent workflows guided by predicted query difficulty. DAAO comprises three interdependent modules: a variational autoencoder (VAE) for difficulty estimation, a modular operator allocator, and a cost- and performance-aware LLM router. A self-adjusting policy updates difficulty estimates based on workflow success, enabling simpler workflows for easy queries and more complex strategies for harder ones. Experiments on six benchmarks demonstrate that DAAO surpasses prior multi-agent systems in both accuracy and inference efficiency, validating its effectiveness for adaptive, difficulty-aware reasoning. Our code is open-sourced at https://github.com/AutoAgents-ai/DAAO

One of the research focuses of large language models (LLMs) is the ability to generate action plans.Recent studies have revealed that the performance of LLMs can be significantly improved by integrating external tools.Based on this, we propose a benchmark framework called PlanningArena, which aims to simulate real application scenarios and provide a series of apps and API tools that may be involved in the actual planning process.This framework adopts a modular task structure and combines user portrait analysis to evaluate the ability of LLMs in correctly selecting tools, logical reasoning in complex scenarios, and parsing user information.In addition, we deeply diagnose the task execution effect of LLMs from both macro and micro levels.The experimental results show that even the most outstanding GPT-4o and DeepSeekV3 models only achieved a total score of 56.5% and 41.9% in PlanningArena, respectively, indicating that current LLMs still face challenges in logical reasoning, context memory, and tool calling when dealing with different structures, scenarios, and their complexity.Through this benchmark, we further explore the path to optimize LLMs to perform planning tasks.

Analog machine learning hardware platforms, such as those using wave physics, present potential for edge artificial intelligence (AI) applications due to in-sensor computing architecture, offering superior energy efficiency compared to digital circuits. While the diffractive neural network has been implemented in optical systems, its deployment on integrated acoustic systems has not been achieved due to the challenges associated with hardware optimization. In this paper, we propose the Diffractive Acoustic Neural Network (DANN), a novel approach that applies diffractive neural network algorithms to surface acoustic wave (SAW) systems for in-sensor multibiomarker diagnosis. To address optimization challenges, we introduce a novel training methodology that combines Finite Element Analysis (FEA) with gradient descent. We validate our method on Major Depressive Disorder (MDD) and prostate cancer, achieving accuracies of $74.07 \%$ and $86.0 \%$, respectively, nearly reaching the accuracy levels of clinical diagnoses. By comparing the co-training method with the traditional gradient descent training method and direct training on the FEA model, the co-training method demonstrates its advantages in balancing training efficiency and accuracy. Furthermore, a comparison of power consumption is conducted between the traditional method and the in-sensor computing system, indicating $66 \%$ energy savings attributed to its high level of integration.

Recently, large language models (LLMs) have achieved significant progress in automated code generation. Despite their strong instruction-following capabilities, these models frequently struggled to align with user intent in the coding scenario. In particular, they were hampered by datasets that lacked diversity and failed to address specialized tasks or edge cases. Furthermore, challenges in supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) led to failures in generating precise, human-intent-aligned code. To tackle these challenges and improve the code generation performance for automated programming systems, we propose Feedback-driven Adaptive Long/short-term memory reinforced Coding OptimizatioN (i.e., FALCON). FALCON leverages long-term memory to retain and apply learned knowledge, short-term memory to incorporate immediate feedback, and meta-reinforcement learning with feedback rewards to address global-local bi-level optimization and enhance adaptability across diverse code generation tasks. Extensive experiments show that FALCON achieves state-of-the-art performance, outperforming other reinforcement learning methods by over 4.5% on MBPP and 6.1% on Humaneval, with the code publicly available. https://anonymous.4open.science/r/FALCON-3B64/README.md.

Gait-based emotion recognition has received considerable attention due to its non-invasive capturing manner. However, most existing works learn the gait representations by treating different classes independently, which ignores the inherent class ambiguity in this field. Therefore, we propose a Multi-soft-label Guided Supervised Contrastive Learning (MSL-SCL) framework, which leverages the class correlation information in soft labels to explicitly guide the SCL, thereby alleviating the ambiguous gait representation. Specifically, a Soft-Label SCL (Sof-SCL) module is designed to select positive and negative samples based on their soft-label similarity to the anchors, and the similarity is further incorporated into a novel contrastive loss function for the refinement. Moreover, a Prior-Guided SCL (Prior-SCL) is introduced to capture the subtle changes in gait and employed as soft labels to provide an adaptive supervision for SCL. Extensive experimental results on the Emotion-Gait dataset demonstrate that our method outperforms SOTAs with a mean average precision of 89.8%.

Despite the progress achieved by existing adversarial distillation (AD) approaches, most mainstream models suffer from inadequate adversarial robustness, due to the challenges of fixed attack strength and unreliable teacher guidance. In this paper, we propose a novel Strength-Dependent Adaptive Regularization (SDAR) paradigm to reinforce the function of adversarial distillation with strength-adaptive adversarial attack (SAA) and multi-dimensional knowledge distillation (MKD). Different from the traditional adversarial training (AT) methods, the proposed SAA scheme dynamically assigns an adaptive and efficient attack strength for each instance, which aims to facilitate smoother classification boundaries. By incorporating dynamic strength coefficients, a comprehensive MKD strategy is designed to fully explore the valuable context information and narrow distribution discrepancies across teacher-student domains. Particularly, our SDAR paradigm can seamlessly integrate with the current AD frameworks, further enhancing the adversarial robustness of deep learning models. Extensive experiments on multiple benchmark datasets consistently demonstrate the superiority of SDAR over state-of-the-art baselines.

Abstract Currently, traffic sign recognition techniques have been brought into the assistive driving of automobiles. However, small traffic sign recognition in real scenes is still a challenging task due to the class imbalance issue and the size limit of the traffic signs. To address the above issues, a feature‐enhanced hybrid attention network is proposed based on YOLOv5s for a small, fast, and accurate traffic sign detector. First, a series of online data augmentation strategies are designed in the preprocessing module for the model training. Second, the hybrid channel and spatial attention module CSAM are integrated into the backbone for a better feature extraction ability. Third, the channel attention module CAM is used in the detection head for a more efficient feature fusion ability. To validate the approach, extensive experiments are conducted based on the Tsinghua‐Tencent 100K dataset. It is found that the novel method achieves state‐of‐the‐art performance with only negligible increases in the model parameter and computational overhead. Specifically, the , parameters, and FLOPs are 85.8%, 7.13 M, and 16.1 G, respectively.

Vision-based gait analysis provides the possibility to automatically and unobtrusively detect walking pattern alterations caused by stoke. Therefore, it can be used to determine the severity of stroke during stroke rehabilitation outside the hospital, which greatly releases the economic and labor burden on patients and their families. However, state-of-the-art deep learning algorithms for gait analysis usually suffer from high computational complexity and can even lead to overfitting problems on small-scale pathological gait datasets. To realize an efficient and effective system, we constructed a specially designed dataset and proposed a novel lightweight network to lean discriminative gait representation to map the input into one of the stroke severity levels. More specifically, a simulated hemiplegia gait dataset with multiple severity levels is first constructed, including sufficient 2D image sequences collected from 14 subjects. Different from the existing pathological datasets used for coarse classification, which only distinguish different pathological gait types, our proposed dataset is specifically designed for fine classification to assess the severity of hemiplegia that is defined according to medical prior. Second, considering that pathological datasets are usually small-scale, an attention-based lightweight network is proposed. In detail, a lightweight hybrid attention module (LHAM) based on the 1D adaptive convolution for channel attention interaction was developed to enhance the network's ability to integrate and focus on meaningful spatial and channel features. To further lighten the networks, a proposed efficient ghost module (EGM) is used in the bottleneck structure instead of the normal convolutional layer. Extensive experiments on both self-constructed and publicly available datasets demonstrate that the proposed efficient hybrid attention-based GhostNet realizes an effective and efficient gait analysis for stroke rehabilitation.