The evolution paradigm of Large Language Models (LLMs) is shifting from scaling training compute to scaling inference-time compute.While Reinforcement Learning with Verifiable Rewards (RLVR) has become a key engine for this transition, standard approaches often fail to equip models with the autonomous improvement capabilities required for test-time scaling.Existing critique-guided methods attempt to mitigate this by leveraging external feedback or ground-truth signals; however, these dependencies are unavailable at test time, fundamentally limiting the model's capacity for continuous self-improvement.To bridge this gap, we propose CURE (Critique-driven Unified REinforcement Learning), a framework that jointly optimizes a single policy for standard solving, critiquing, and guided re-exploration.Uniquely, CURE facilitates re-exploration by generating strategic hints while discarding initial incorrect solutions to mitigate anchoring bias.Empirical results across diverse mathematical reasoning and code generation benchmarks demonstrate that CURE not only maintains competitive single-turn performance but, more importantly, unlocks effective inferencetime scaling, enabling the model to significantly boost accuracy through iterative selfimprovement.

Superalignment, where humans act as weak supervisors for superhuman models, has become a crucial problem with the rapid development of Large Language Models (LLMs). Recent work has preliminarily studied this problem by using weak models to supervise strong models, and discovered that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception, where strong models deceive weak models by exhibiting well-aligned in areas known to weak models but producing misaligned behaviors in cases weak models do not know. We take an initial step towards exploring this security issue in a specific but realistic multi-objective alignment case, where there may be some alignment targets conflicting with each other (e.g., helpfulness v.s. harmlessness). We aim to explore whether, in such cases, strong models might deliberately make mistakes in areas known to them but unknown to weak models within one alignment dimension, in exchange for a higher reward in another dimension. Through extensive experiments in both the reward modeling and preference optimization scenarios, we find: (1) The weak-to-strong deception phenomenon exists across all settings. (2) The deception intensifies as the capability gap between weak and strong models increases. (3) Bootstrapping with an intermediate model can mitigate the deception to some extent, though its effectiveness remains limited. Our work highlights the urgent need to pay more attention to the true reliability of superalignment.

Recently, tool use with LLMs has become one of the primary research topics as it can help LLM generate truthful and helpful responses.Existing studies on tool use with LLMs primarily focus on enhancing the tool-calling ability of LLMs.In practice, like chat assistants, LLMs are also required to align with human values in the context of tool use.Specifically, LLMs should refuse to answer unsafe tool use relevant instructions and insecure tool responses to ensure their reliability and harmlessness.At the same time, LLMs should demonstrate autonomy in tool use to reduce the costs associated with tool calling.To tackle this issue, we first introduce the principle that LLMs should follow in tool use scenarios: H2A.The goal of H2A is to align LLMs with helpfulness, harmlessness, and autonomy.In addition, we propose ToolAlign, a dataset comprising instruction-tuning data and preference data to align LLMs with the H2A principle for tool use.Based on ToolAlign, we develop LLMs by supervised fine-tuning and preference learning, and experimental results demonstrate that the LLMs exhibit remarkable toolcalling capabilities, while also refusing to engage with harmful content, and displaying a high degree of autonomy in tool utilization.

The existing graph neural architecture search (GNAS) methods heavily rely on supervised labels during the search process, failing to handle ubiquitous scenarios where supervisions are not available. In this paper, we study the problem of unsupervised graph neural architecture search, which remains unexplored in the literature. The key problem is to discover the latent graph factors that drive the formation of graph data as well as the underlying relations between the factors and the optimal neural architectures. Handling this problem is challenging given that the latent graph factors together with architectures are highly entangled due to the nature of the graph and the complexity of the neural architecture search process. To address the challenge, we propose a novel Disentangled Self-supervised Graph Neural Architecture Search (DSGAS) model, which is able to discover the optimal architectures capturing various latent graph factors in a self-supervised fashion based on unlabeled graph data. Specifically, we first design a disentangled graph super-network capable of incorporating multiple architectures with factor-wise disentanglement, which are optimized simultaneously. Then, we estimate the performance of architectures under different factors by our proposed self-supervised training with joint architecture-graph disentanglement. Finally, we propose a contrastive search with architecture augmentations to discover architectures with factor-specific expertise. Extensive experiments on 11 real-world datasets demonstrate that the proposed model is able to achieve state-of-the-art performance against several baseline methods in an unsupervised manner.

Neural-symbolic learning, aiming to combine the perceiving power of neural perception and the reasoning power of symbolic logic together, has drawn increasing research attention. However, existing works simply cascade the two components together and optimize them isolatedly, failing to utilize the mutual enhancing information between them. To address this problem, we propose DeepLogic, a framework with joint learning of neural perception and logical reasoning, such that these two components are jointly optimized through mutual supervision signals. In particular, the proposed DeepLogic framework contains a deep-logic module that is capable of representing complex first-order-logic formulas in a tree structure with basic logic operators. We then theoretically quantify the mutual supervision signals and propose the deep&logic optimization algorithm for joint optimization. We further prove the convergence of DeepLogic and conduct extensive experiments on model performance, convergence, and generalization, as well as its extension to the continuous domain. The experimental results show that through jointly learning both perceptual ability and logic formulas in a weakly supervised manner, our proposed DeepLogic framework can significantly outperform DNN-based baselines by a great margin and beat other strong baselines without out-of-box tools.

Humans can perceive subtle emotions from various cues and contexts, even without hearing or seeing others. However, existing video datasets mainly focus on recognizing the emotions of the speakers from complete modalities. In this work, we present the task of multimodal emotion reasoning in videos. Beyond directly recognizing emotions from multimodal signals of target persons, this task requires a machine capable of reasoning about human emotions from the contexts and surrounding world. To facilitate the study towards this task, we introduce a new dataset, MEmoR, that provides fine-grained emotion annotations for both speakers and non-speakers. The videos in MEmoR are collected from TV shows closely in real-life scenarios. In these videos, while speakers may be non-visually described, non-speakers always deliver no audio-textual signals and are often visually inconspicuous. This modality-missing characteristic makes MEmoR a more practical yet challenging testbed for multimodal emotion reasoning. In support of various reasoning behaviors, the proposed MEmoR dataset provides both short-term contexts and external knowledge. We further propose an attention-based reasoning approach to model the intra-personal emotion contexts, inter-personal emotion propagation, and the personalities of different individuals. Experimental results demonstrate that our proposed approach outperforms related baselines significantly. We isolate and analyze the validity of different reasoning modules across various emotions of speakers and non-speakers. Finally, we draw forth several future research directions for multimodal emotion reasoning with MEmoR, aiming to empower high Emotional Quotient (EQ) in modern artificial intelligence systems. The code and dataset released on https://github.com/sunlightsgy/MEmoR.

Temporal action localization aims to accurately localize and recognize all possible action instances from an untrimmed video automatically. Most existing methods perform this task by first generating a set of proposals and then recognizing each independently. However, due to the complex structures and large content variations in action instances, recognizing them individually can be difficult. Fortunately, some proposals often share information regarding one specific action. Such information, which is ignored in existing methods, can be used to boost recognition performance. In this paper, we propose a novel mechanism, called relation attention, to exploit informative relations among proposals based on their appearance or optical flow features. Specifically, we propose a relation attention module to enhance representation power by capturing useful information from other proposals. This module does not change the dimensions of the original input and output and does not rely on any specific proposal generation methods or feature extraction backbone networks. Experimental results show that the proposed relation attention mechanism improves performance significantly on both Thumos14 and ActivityNet1.3 datasets compared to existing architectures. For example, relying on Structured Segment Networks (SSN), the proposed relation attention module helps to increase the mAP from 41.4 to 43.7 on the Thumos14 dataset and outperforms the state-of-the-art results.

There has been a prominent emergence of work on video prediction, aiming to extrapolate the future video frames from the past. Existing temporal-based methods are limited to certain numbers of frames. In this paper, we study video prediction from a single still image in the facial expression domain, a.k.a, facial image-to-video translation. Our main approach, dubbed AffineGAN, associates each facial image with an expression intensity and leverages an affine transformation in the latent space. AffineGAN allows users to control the number of frames to predict as well as the expression intensity for each of them. Unlike previous intensity-based methods, We derive an inverse formulation to the affine transformation, enabling automatic inference of the facial expression intensities from videos --- manual annotation is not only tedious but also ambiguous as people express in various ways and have different opinions about the intensity of a facial image. Both quantitative and qualitative results verify the superiority of AffineGAN over the state of the arts. Notably, in a Turing test with web faces, more than 50% of the facial expression videos generated by AffineGAN are considered real by the Amazon Mechanical Turk workers. This work could improve users' communication experience by enabling them to conveniently and creatively produce expression GIFs, which are popular art forms in online messaging and social networks.

Depression detection is a significant issue for human well-being. In previous studies, online detection has proven effective in Twitter, enabling proactive care for depressed users. Owing to cultural differences, replicating the method to other social media platforms, such as Chinese Weibo, however, might lead to poor performance because of insufficient available labeled (self-reported depression) data for model training. In this paper, we study an interesting but challenging problem of enhancing detection in a certain target domain (e.g. Weibo) with ample Twitter data as the source domain. We first systematically analyze the depression-related feature patterns across domains and summarize two major detection challenges, namely isomerism and divergency. We further propose a cross-domain Deep Neural Network model with Feature Adaptive Transformation & Combination strategy (DNN-FATC) that transfers the relevant information across heterogeneous domains. Experiments demonstrate improved performance compared to existing heterogeneous transfer methods or training directly in the target domain (over 3.4% improvement in F1), indicating the potential of our model to enable depression detection via social media for more countries with different cultural settings.

Depression is a major contributor to the overall global burden of diseases. Traditionally, doctors diagnose depressed people face to face via referring to clinical depression criteria. However, more than 70% of the patients would not consult doctors at early stages of depression, which leads to further deterioration of their conditions. Meanwhile, people are increasingly relying on social media to disclose emotions and sharing their daily lives, thus social media have successfully been leveraged for helping detect physical and mental diseases. Inspired by these, our work aims to make timely depression detection via harvesting social media data. We construct well-labeled depression and non-depression dataset on Twitter, and extract six depression-related feature groups covering not only the clinical depression criteria, but also online behaviors on social media. With these feature groups, we propose a multimodal depressive dictionary learning model to detect the depressed users on Twitter. A series of experiments are conducted to validate this model, which outperforms (+3% to +10%) several baselines. Finally, we analyze a large-scale dataset on Twitter to reveal the underlying online behaviors between depressed and non-depressed users.

Psychological stress is threatening people's health. It is non-trivial to detect stress timely for proactive care. With the popularity of social media, people are used to sharing their daily activities and interacting with friends on social media platforms, making it feasible to leverage online social network data for stress detection. In this paper, we find that users stress state is closely related to that of his/her friends in social media, and we employ a large-scale dataset from real-world social platforms to systematically study the correlation of users' stress states and social interactions. We first define a set of stress-related textual, visual, and social attributes from various aspects, and then propose a novel hybrid model - a factor graph model combined with Convolutional Neural Network to leverage tweet content and social interaction information for stress detection. Experimental results show that the proposed model can improve the detection performance by 6-9 percent in F1-score. By further analyzing the social interaction data, we also discover several intriguing phenomena, i.e., the number of social structures of sparse connections (i.e., with no delta connections) of stressed users is around 14 percent higher than that of non-stressed users, indicating that the social structure of stressed users' friends tend to be less connected and less complicated than that of non-stressed users.

How do people describe clothing? The words like “formal”or "casual" are usually used. However, recent works often focus on recognizing or extracting visual features (e.g., sleeve length, color distribution and clothing pattern) from clothing images accurately. How can we bridge the gap between the visual features and the aesthetic words? In this paper, we formulate this task to a novel three-level framework: visual features(VF) - image-scale space (ISS) - aesthetic words space(AWS). Leveraging the art-field image-scale space served as an intermediate layer, we first propose a Stacked Denoising Autoencoder Guided by CorrelativeLabels (SDAE-GCL) to map the visual features to the image-scale space; and then according to the semantic distances computed byWordNet::Similarity, we map the most often used aesthetic words in online clothing shops to the image-scale space too. Employing upper body menswear images downloaded from several global online clothing shops as experimental data, the results indicate that the proposed three-level framework can help to capture the subtle relationship between visual features and aesthetic words better compared to several baselines. To demonstrate that our three-level framework and its implementation methods are universally applicable, we finally present some interesting analyses on the fashion trend of menswear in the last 10 years.