Outcome-based reinforcement learning has made notable advances in training language models (LMs) for reasoning. However, without explicit incentives and controls, this paradigm has limitations and instability in eliciting high-quality reasoning trajectories with diverse actions—particularly for models whose pretraining lacked extensive reasoning-related data. To this end, we introduce MetaAct-RL, a new RL framework that frames LMs’ thinking as sequential decision making over meta-actions. In this framework, the model chooses and executes a high-level action at each step—such as forward reasoning, critique, or refinement—to gradually reach the correct answer. To encourage deeper exploration, richer action diversity, and to improve sampling efficiency in the RL optimization process, MetaAct-RL incorporates appropriate length-based reward and regularization, and a key-state restart mechanism. Extensive experiments across six benchmarks show that MetaAct-RL improves reasoning performance by 7.99 on Llama3.2-1B and 7.17 on Llama3.1-8B relative to vanilla RL method. Moreover, on the challenging AIME-2024, our method outperforms the vanilla RL by 7.5 with Qwen2.5-1.5B.

Xin Guo, Zhiheng Xi, Yiwen Ding, Yitao Zhai, Xiaowei Shi, Xunliang Cai, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Xin Guo, Dingwen Yang, Chenyang Liao, Wei He, Songyang Gao, Lu Chen, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We present MANTA, a visual-text anomaly detection dataset for tiny objects. The visual component comprises over 137.3K images across 38 object categories spanning five typical domains, of which 8.6K images are labeled as anomalous with pixel-level annotations. Each image is captured from five distinct viewpoints to ensure comprehensive object coverage. The text component consists of two subsets: Declarative Knowledge, including 875 words that describe common anomalies across various domains and specific categories, with detailed explanations for ⟨what, why, how⟩, including causes and visual characteristics; and Constructivist Learning, providing 2K multiple-choice questions with varying levels of difficulty, each paired with images and corresponded answer explanations. We also propose a baseline for visual-text tasks and conduct extensive benchmarking experiments to evaluate advanced methods across different settings, highlighting the challenges and efficacy of our dataset.

Yiwen Ding, Zhiheng Xi, Wei He, Lizhuoyuan Lizhuoyuan, Yitao Zhai, Shi Xiaowei, Xunliang Cai, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Grain appearance inspection is crucial for evaluating grain quality and determining seed stratification. Typically, trained inspectors manually examine each grain kernel to identify and remove defective ones, which is time-consuming and error-prone. In this article, we present GrainBrain, a robotic vision-based system comprising a hardware prototype (A100) and a deep learning model (GrainAD). A100 is equipped with five cameras to capture high-quality, multiview images of each kernel. The identification of defective kernels is treated as an unsupervised anomaly detection task. GrainAD trains a classifier to distinguish between healthy and pseudoanomaly samples generated at both image and feature levels, and a supervised contrastive learning loss is employed to obtain compact feature representations of healthy kernels. In addition, we release a large-scale dataset containing over 100K annotated images of four types of cereal grains. Extensive experiments were conducted to verify the superiority of our system, achieving an average AUROC of 94.4/90.4% at the image/pixel level. Our system excelled in both efficiency and consistency, as demonstrated by experiments comparing human experts to the system.

This paper presents a comprehensive evaluation of the performance of different numerical methods in cardiac electrophysiological simulations, including the Explicit Euler (EE) Method, Implicit Euler (IE) Method, Trapezoidal Rule (TR), and the Fourth-Order Runge-Kutta (RK4) Method within the framework of one-dimensional (1D), two-dimensional (2D), and three-dimensional (3D) models of the human ventricles. Simulation accuracy was investigated for varying the time step (h) for different tissue scales and complexities. It was shown that, as the dimension of the model increased, the upper limit of h for acceptable simulation accuracy (hul) gradually decreased. However, the difference in hulamong the methods was small. As the h or the tissue dimension increased, the EE method showed greater computational error compared to other methods. The experiment results also showed that the amplitude of the external stimulation pulse current, the size of the spatial range of tissue receiving stimuli, the spatial dimension of the tissue models, as well as the diffusion coefficient of the model, also affected simulation accuracy. The present study discusses the influence of numerical methods on the accuracy of cardiac simulations, providing insights for choosing optimal numerical methods for cardiac simulations.

Large language models (LLMs) have achieved tremendous success in understanding language and processing text.However, questionanswering (QA) on lengthy documents faces challenges of resource constraints and a high propensity for errors, even for the most advanced models such as GPT-4 and Claude2.In this paper, we introduce LONGAGENT, a multi-agent collaboration method that enables efficient and effective QA over 128k-tokenlong documents.LONGAGENT adopts a divideand-conquer strategy, breaking down lengthy documents into shorter, more manageable text chunks.A leader agent comprehends the user's query and organizes the member agents to read their assigned chunks, reasoning a final answer through multiple rounds of discussion.Due to members' hallucinations, it's difficult to guarantee that every response provided by each member is accurate.To address this, we develop an inter-member communication mechanism that facilitates information sharing, allowing for the detection and mitigation of hallucinatory responses.Experimental results show that a LLaMA-2 7B driven by LONGAGENT can effectively support QA over 128k-token documents, achieving 16.42% and 1.63% accuracy gains over GPT-4 on singlehop and multi-hop QA settings, respectively.

In this paper, we propose R$^3$: Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models. The core challenge in applying RL to complex reasoning is to identify a sequence of actions that result in positive rewards and provide appropriate supervision for optimization. Outcome supervision provides sparse rewards for final results without identifying error locations, whereas process supervision offers step-wise rewards but requires extensive manual annotation. R$^3$ overcomes these limitations by learning from correct demonstrations. Specifically, R$^3$ progressively slides the start state of reasoning from a demonstration’s end to its beginning, facilitating easier model exploration at all stages. Thus, R$^3$ establishes a step-wise curriculum, allowing outcome supervision to offer step-level signals and precisely pinpoint errors. Using Llama2-7B, our method surpasses RL baseline on eight reasoning tasks by $4.1$ points on average. Notably, in program-based reasoning, 7B-scale models perform comparably to larger models or closed-source models with our R$^3$.

In this article, we present our work in social emotion prediction to online trending topics. While most prior works focus on emotion from writers or the readers’ emotions evoked by news articles, we investigate discussions from massive social media users and explore the public feelings to the online trending topic. We employ user-generated “#hashtags” to indicate online trending topics and construct a large-scale Chinese dataset for social emotion prediction (LCSEP) to trending topics collected from the Chinese microblog Sina Weibo. It contains more than 20 000 trending topics, each with social emotions voted in 24 fine-grained types, and gathers hashtags, posts, comments, and related metadata to give each trending topic a thorough context. We also propose aHashtag- and Topic-Enhanced Attention Model(HTEAM) that combines a pretrained BERT model, a neural topic model, and an attention mechanism via joint training to understand social emotion. Experiments show that HTEAM outperforms baselines and achieves the state-of-the-art result.

Cereal grain plays a crucial role in the human diet as a major source of essential nutrients. Grain Appearance Inspection (GAI) serves as an essential process to determine grain quality and facilitate grain circulation and processing. However, GAI is routinely performed manually by inspectors with cumbersome procedures, which poses a significant bottleneck in smart agriculture. In this paper, we endeavor to develop an automated GAI system: AI4GrainInsp. By analyzing the distinctive characteristics of grain kernels, we formulate GAI as a ubiquitous problem: Anomaly Detection (AD), in which healthy and edible kernels are considered normal samples while damaged grains or unknown objects are regarded as anomalies. We further propose an AD model, called AD-GAI, which is trained using only normal samples yet can identify anomalies during inference. Moreover, we customize a prototype device for data acquisition and create a large-scale dataset including 220K high-quality images of wheat and maize kernels. Through extensive experiments, AD-GAI achieves considerable performance in comparison with advanced AD methods, and AI4GrainInsp has highly consistent performance compared to human experts and excels at inspection efficiency over 20× speedup. The dataset, code and models will be released at https://github.com/hellodfan/AI4GrainInsp.

Cereal grains are a vital part of human diets and are important commodities for people's livelihood and international trade. Grain Appearance Inspection (GAI) serves as one of the crucial steps for the determination of grain quality and grain stratification for proper circulation, storage and food processing, etc. GAI is routinely performed manually by qualified inspectors with the aid of some hand tools. Automated GAI has the benefit of greatly assisting inspectors with their jobs but has been limited due to the lack of datasets and clear definitions of the tasks. In this paper we formulate GAI as three ubiquitous computer vision tasks: fine-grained recognition, domain adaptation and out-of-distribution recognition. We present a large-scale and publicly available cereal grains dataset called GrainSpace. Specifically, we construct three types of device prototypes for data acquisition, and a total of 5.25 million images determined by professional inspectors. The grain samples including wheat, maize and rice are collected from five countries and more than 30 regions. We also develop a comprehensive benchmark based on semi-supervised learning and self-supervised learning techniques. To the best of our knowledge, GrainSpace is the first publicly released dataset for cereal grain inspection, https://github.com/hellodfan/GrainSpace.