Yao Liu

Xiangning Chen, Chen Liang, Da Huang et al. · cmu, deepmind

We present a method to formulate algorithm discovery as program search, and apply it to discover optimization algorithms for deep neural network training. We leverage efficient search techniques to explore an infinite and sparse program space. To bridge the large generalization gap between proxy and target tasks, we also introduce program selection and simplification strategies. Our method discovers a simple and effective optimization algorithm, $\textbf{Lion}$ ($\textit{Evo$\textbf{L}$ved S$\textbf{i}$gn M$\textbf{o}$me$\textbf{n}$tum}$). It is more memory-efficient than Adam as it only keeps track of the momentum. Different from adaptive optimizers, its update has the same magnitude for each parameter calculated through the sign operation. We compare Lion with widely used optimizers, such as Adam and Adafactor, for training a variety of models on different tasks. On image classification, Lion boosts the accuracy of ViT by up to 2% on ImageNet and saves up to 5x the pre-training compute on JFT. On vision-language contrastive learning, we achieve 88.3% $\textit{zero-shot}$ and 91.1% $\textit{fine-tuning}$ accuracy on ImageNet, surpassing the previous best results by 2% and 0.1%, respectively. On diffusion models, Lion outperforms Adam by achieving a better FID score and reducing the training compute by up to 2.3x. For autoregressive, masked language modeling, and fine-tuning, Lion exhibits a similar or better performance compared to Adam. Our analysis of Lion reveals that its performance gain grows with the training batch size. It also requires a smaller learning rate than Adam due to the larger norm of the update produced by the sign function. Additionally, we examine the limitations of Lion and identify scenarios where its improvements are small or not statistically significant. Lion is also successfully deployed in production systems such as Google search ads CTR model.

Lishen Qu, Yao Liu, Jie Liang et al.

This paper presents NTIRE 2026, the 3rd Restore Any Image Model (RAIM) challenge on multi-exposure image fusion in dynamic scenes. We introduce a benchmark that targets a practical yet difficult HDR imaging setting, where exposure bracketing must be fused under scene motion, illumination variation, and handheld camera jitter. The challenge data contains 100 training sequences with 7 exposure levels and 100 test sequences with 5 exposure levels, reflecting real-world scenarios that frequently cause misalignment and ghosting artefacts. We evaluate submissions with a leaderboard score derived from PSNR, SSIM, and LPIPS, while also considering perceptual quality, efficiency, and reproducibility during the final review. This track attracted 114 participating teams and received 987 submissions. The winning methods significantly improved the ability to remove artifacts from multi-exposure fusion and recover fine details. The dataset and the code of each team can be found at the repository: https://github.com/qulishen/RAIM-HDR.

Sherry Ruan, Allen Nie, William Steenbergen et al. · stanford

Resource limitations make it hard to provide all students with one of the most effective educational interventions: personalized instruction. Reinforcement learning could be a key tool to reduce the development cost and improve the effectiveness of intelligent tutoring software that aims to provide the right support, at the right time, to a student. Here we illustrate that deep reinforcement learning can be used to provide adaptive pedagogical support to students learning about the concept of volume in a narrative storyline software. Using explainable artificial intelligence tools, we extracted interpretable insights about the pedagogical policy learned and demonstrated that the resulting policy had similar performance in a different student population. Most importantly, in both studies, the reinforcement-learning narrative system had the largest benefit for those students with the lowest initial pretest scores, suggesting the opportunity for AI to adapt and provide support for those most in need.

Yao Liu, Dipendra Misra, Miro Dudík et al. · stanford

We study reinforcement learning (RL) in settings where observations are high-dimensional, but where an RL agent has access to abstract knowledge about the structure of the state space, as is the case, for example, when a robot is tasked to go to a specific room in a building using observations from its own camera, while having access to the floor plan. We formalize this setting as transfer reinforcement learning from an abstract simulator, which we assume is deterministic (such as a simple model of moving around the floor plan), but which is only required to capture the target domain's latent-state dynamics approximately up to unknown (bounded) perturbations (to account for environment stochasticity). Crucially, we assume no prior knowledge about the structure of observations in the target domain except that they can be used to identify the latent states (but the decoding map is unknown). Under these assumptions, we present an algorithm, called TASID, that learns a robust policy in the target domain, with sample complexity that is polynomial in the horizon, and independent of the number of states, which is not possible without access to some prior knowledge. In synthetic experiments, we verify various properties of our algorithm and show that it empirically outperforms transfer RL algorithms that require access to "full simulators" (i.e., those that also simulate observations).

Zuxin Liu, Jesse Zhang, Kavosh Asadi et al. · cmu

The full potential of large pretrained models remains largely untapped in control domains like robotics. This is mainly because of the scarcity of data and the computational challenges associated with training or fine-tuning these large models for such applications. Prior work mainly emphasizes either effective pretraining of large models for decision-making or single-task adaptation. But real-world problems will require data-efficient, continual adaptation for new control tasks. Recognizing these constraints, we introduce TAIL (Task-specific Adapters for Imitation Learning), a framework for efficient adaptation to new control tasks. Inspired by recent advancements in parameter-efficient fine-tuning in language domains, we explore efficient fine-tuning techniques -- e.g., Bottleneck Adapters, P-Tuning, and Low-Rank Adaptation (LoRA) -- in TAIL to adapt large pretrained models for new tasks with limited demonstration data. Our extensive experiments in large-scale language-conditioned manipulation tasks comparing prevalent parameter-efficient fine-tuning techniques and adaptation baselines suggest that TAIL with LoRA can achieve the best post-adaptation performance with only 1\% of the trainable parameters of full fine-tuning, while avoiding catastrophic forgetting and preserving adaptation plasticity in continual learning settings.

Yao Liu, Yannis Flet-Berliac, Emma Brunskill · stanford

Offline policy optimization could have a large impact on many real-world decision-making problems, as online learning may be infeasible in many applications. Importance sampling and its variants are a commonly used type of estimator in offline policy evaluation, and such estimators typically do not require assumptions on the properties and representational capabilities of value function or decision process model function classes. In this paper, we identify an important overfitting phenomenon in optimizing the importance weighted return, in which it may be possible for the learned policy to essentially avoid making aligned decisions for part of the initial state space. We propose an algorithm to avoid this overfitting through a new per-state-neighborhood normalization constraint, and provide a theoretical justification of the proposed algorithm. We also show the limitations of previous attempts to this approach. We test our algorithm in a healthcare-inspired simulator, a logged dataset collected from real hospitals and continuous control tasks. These experiments show the proposed method yields less overfitting and better test performance compared to state-of-the-art batch reinforcement learning algorithms.

Kavosh Asadi, Shoham Sabach, Yao Liu et al. · stanford

We study the convergence behavior of the celebrated temporal-difference (TD) learning algorithm. By looking at the algorithm through the lens of optimization, we first argue that TD can be viewed as an iterative optimization algorithm where the function to be minimized changes per iteration. By carefully investigating the divergence displayed by TD on a classical counter example, we identify two forces that determine the convergent or divergent behavior of the algorithm. We next formalize our discovery in the linear TD setting with quadratic loss and prove that convergence of TD hinges on the interplay between these two forces. We extend this optimization perspective to prove convergence of TD in a much broader setting than just linear approximation and squared loss. Our results provide a theoretical explanation for the successful application of TD in reinforcement learning.

Yao Liu, Pratik Chaudhari, Rasool Fakoor · stanford

The main challenge of offline reinforcement learning, where data is limited, arises from a sequence of counterfactual reasoning dilemmas within the realm of potential actions: What if we were to choose a different course of action? These circumstances frequently give rise to extrapolation errors, which tend to accumulate exponentially with the problem horizon. Hence, it becomes crucial to acknowledge that not all decision steps are equally important to the final outcome, and to budget the number of counterfactual decisions a policy make in order to control the extrapolation. Contrary to existing approaches that use regularization on either the policy or value function, we propose an approach to explicitly bound the amount of out-of-distribution actions during training. Specifically, our method utilizes dynamic programming to decide where to extrapolate and where not to, with an upper bound on the decisions different from behavior policy. It balances between the potential for improvement from taking out-of-distribution actions and the risk of making errors due to extrapolation. Theoretically, we justify our method by the constrained optimality of the fixed point solution to our $Q$ updating rules. Empirically, we show that the overall performance of our method is better than the state-of-the-art offline RL methods on tasks in the widely-used D4RL benchmarks.

Guanyu Zhou, Yao Liu, Yanglei Gan et al.

Spatio-temporal point processes (STPPs) provide a principled framework for modeling asynchronous events in continuous time and space. Recent diffusion-based approaches offer a flexible alternative to deterministic prediction by modeling complex conditional distributions, but their application to STPPs remains challenging: reverse sampling from pure noise is costly, and weak structural constraints in sparse spatial domains can lead to poorly localized probability mass. We propose \textbf{GLIDE} (Graph-guided Leap Inference for Diffusion Estimation), a conditional diffusion framework for next-event modeling in STPPs. GLIDE organizes historical events into a multi-scale historical graph and encodes temporal evolution and spatial topology through a dual-stream architecture, yielding a structured conditioning context for a dual-branch diffusion denoiser. It further introduces a prior-guided leap inference mechanism, in which a lightweight mean predictor provides a deterministic anchor and the reverse process starts from an intermediate diffusion step instead of from pure Gaussian noise. Experiments on multiple real-world datasets show that GLIDE improves both distribution fitting and next-event prediction, with the largest gains appearing on the spatial side. The results also indicate that prior-guided leap inference substantially reduces reverse-sampling cost while preserving the stochastic generation capability of diffusion models.

Peng He, Yao Liu, Yanglei Gan et al.

While Transformer-based architectures have established themselves as a dominant paradigm in Multivariate Time Series Forecasting (MTSF), their core self-attention mechanism inherently functions as a low-pass filter, systematically smoothing out high-frequency signals vital for sharp local changes. Recent advancements have increasingly incorporated frequency-domain operations to address this bias, however, most existing designs rely on fixed spectral bases and apply sequence-wise (uniform) modulation, implicitly assuming a time-invariant frequency response. This overlooks a key property of real-world series that their spectral characteristics often evolve over time, making uniform modulation insufficient for capturing fine-grained temporal dynamics. To tackle these limitations, we propose FAiT, a Frequency-Aware inverted Transformer. Specifically, FAiT rectifies the spectral bias internally through Inverted Attention, which interprets the attention map as a learnable low-pass operator and constructs a dedicated complementary high-pass branch by inverting the attention matrix to recover attenuated transient signals. Furthermore, FAiT introduces Dynamic Temporal-Frequency Modulation (DTFM), which synthesizes instance-conditioned weights to adaptively re-calibrate the energy of spectral sub-bands, enabling fine-grained control over evolving multi-scale patterns. Extensive experiments on widely used benchmarks demonstrate that FAiT consistently outperforms state-of-the-art Transformer-based and frequency-enhanced baselines, while maintaining computational efficiency.

Yao Liu, Zesheng Ye, Rui Wang et al.

Tremendous efforts have been put forth on predicting pedestrian trajectory with generative models to accommodate uncertainty and multi-modality in human behaviors. An individual's inherent uncertainty, e.g., change of destination, can be masked by complex patterns resulting from the movements of interacting pedestrians. However, latent variable-based generative models often entangle such uncertainty with complexity, leading to limited either latent expressivity or predictive diversity. In this work, we propose to separately model these two factors by implicitly deriving a flexible latent representation to capture intricate pedestrian movements, while integrating predictive uncertainty of individuals with explicit bivariate Gaussian mixture densities over their future locations. More specifically, we present a model-agnostic uncertainty-aware pedestrian trajectory prediction framework, parameterizing sufficient statistics for the mixture of Gaussians that jointly comprise the multi-modal trajectories. We further estimate these parameters of interest by approximating a denoising process that progressively recovers pedestrian movements from noise. Unlike previous studies, we translate the predictive stochasticity to explicit distributions, allowing it to readily generate plausible future trajectories indicating individuals' self-uncertainty. Moreover, our framework is compatible with different neural net architectures. We empirically show the performance gains over state-of-the-art even with lighter backbones, across most scenes on two public benchmarks.

Zhe Qu, Xingyu Li, Rui Duan et al.

Federated Learning (FL) is a promising framework for performing privacy-preserving, distributed learning with a set of clients. However, the data distribution among clients often exhibits non-IID, i.e., distribution shift, which makes efficient optimization difficult. To tackle this problem, many FL algorithms focus on mitigating the effects of data heterogeneity across clients by increasing the performance of the global model. However, almost all algorithms leverage Empirical Risk Minimization (ERM) to be the local optimizer, which is easy to make the global model fall into a sharp valley and increase a large deviation of parts of local clients. Therefore, in this paper, we revisit the solutions to the distribution shift problem in FL with a focus on local learning generality. To this end, we propose a general, effective algorithm, \texttt{FedSAM}, based on Sharpness Aware Minimization (SAM) local optimizer, and develop a momentum FL algorithm to bridge local and global models, \texttt{MoFedSAM}. Theoretically, we show the convergence analysis of these two algorithms and demonstrate the generalization bound of \texttt{FedSAM}. Empirically, our proposed algorithms substantially outperform existing FL studies and significantly decrease the learning deviation.

Rui Duan, Zhe Qu, Leah Ding et al.

Audio adversarial examples (AEs) have posed significant security challenges to real-world speaker recognition systems. Most black-box attacks still require certain information from the speaker recognition model to be effective (e.g., keeping probing and requiring the knowledge of similarity scores). This work aims to push the practicality of the black-box attacks by minimizing the attacker's knowledge about a target speaker recognition model. Although it is not feasible for an attacker to succeed with completely zero knowledge, we assume that the attacker only knows a short (or a few seconds) speech sample of a target speaker. Without any probing to gain further knowledge about the target model, we propose a new mechanism, called parrot training, to generate AEs against the target model. Motivated by recent advancements in voice conversion (VC), we propose to use the one short sentence knowledge to generate more synthetic speech samples that sound like the target speaker, called parrot speech. Then, we use these parrot speech samples to train a parrot-trained(PT) surrogate model for the attacker. Under a joint transferability and perception framework, we investigate different ways to generate AEs on the PT model (called PT-AEs) to ensure the PT-AEs can be generated with high transferability to a black-box target model with good human perceptual quality. Real-world experiments show that the resultant PT-AEs achieve the attack success rates of 45.8% - 80.8% against the open-source models in the digital-line scenario and 47.9% - 58.3% against smart devices, including Apple HomePod (Siri), Amazon Echo, and Google Home, in the over-the-air scenario.

Yao Liu, Gangfeng Cui, Jiahui Luo et al.

As a fundamental aspect of human life, two-person interactions contain meaningful information about people's activities, relationships, and social settings. Human action recognition serves as the foundation for many smart applications, with a strong focus on personal privacy. However, recognizing two-person interactions poses more challenges due to increased body occlusion and overlap compared to single-person actions. In this paper, we propose a point cloud-based network named Two-stream Multi-level Dynamic Point Transformer for two-person interaction recognition. Our model addresses the challenge of recognizing two-person interactions by incorporating local-region spatial information, appearance information, and motion information. To achieve this, we introduce a designed frame selection method named Interval Frame Sampling (IFS), which efficiently samples frames from videos, capturing more discriminative information in a relatively short processing time. Subsequently, a frame features learning module and a two-stream multi-level feature aggregation module extract global and partial features from the sampled frames, effectively representing the local-region spatial information, appearance information, and motion information related to the interactions. Finally, we apply a transformer to perform self-attention on the learned features for the final classification. Extensive experiments are conducted on two large-scale datasets, the interaction subsets of NTU RGB+D 60 and NTU RGB+D 120. The results show that our network outperforms state-of-the-art approaches in most standard evaluation settings.

Xuanpu Zhao, Zhentao Tan, Dianmo Sheng et al.

To enhance the perception and reasoning capabilities of multimodal large language models in complex visual scenes, recent research has introduced agent-based workflows. In these works, MLLMs autonomously utilize image cropping tool to analyze regions of interest for question answering. While existing training strategies, such as those employing supervised fine-tuning and reinforcement learning, have made significant progress, our empirical analysis reveals a key limitation. We demonstrate the model's strong reliance on global input and its weak dependence on the details within the cropped region. To address this issue, we propose a novel two-stage reinforcement learning framework that does not require trajectory supervision. In the first stage, we introduce the ``Information Gap" mechanism by adjusting the granularity of the global image. This mechanism trains the model to answer questions by focusing on cropped key regions, driven by the information gain these regions provide. The second stage further enhances cropping precision by incorporating a grounding loss, using a small number of bounding box annotations. Experiments show that our method significantly enhances the model's attention to cropped regions, enabling it to achieve state-of-the-art performance on high-resolution visual question-answering benchmarks. Our method provides a more efficient approach for perceiving and reasoning fine-grained details in MLLMs. Code is available at: https://github.com/XuanPu-Z/LFPC.

Yuming, Huang, Yao Liu et al.

Subjective evaluation of LLM behavior -- empathy, restraint, calibrated emotional tone -- is hard. Human inter-rater agreement on such qualities saturates near rho ~ 0.45, and an LLM-as-judge proxy alone risks circularity: a judge sharing the target's training cohort cannot independently verify it. Anchoring validity to a single human-rater consensus does not extend to capabilities where humans themselves disagree. We propose a replication-first paradigm: instead of anchoring on one rater group, we certify the instrument via four orthogonal properties -- reliability across K runs, cross-instrument replication across architecturally distinct judges, historical-footprint calibration via judges from earlier training cohorts, and pre-registered prediction. We test it on emotional accompaniment by letting the rubric self-evolve data-driven across iterations: the dimensions are not pre-stipulated and the procedure stabilizes to a 9-dimension set. Pre-registration applies to 10 falsifiable hypotheses and 11 forward predictions, committed before any test data was collected. Applied to 49 models across 8 families, the paradigm surfaces what aggregate scores hide. On advice-restraint -- whether a model refrains from giving unsolicited solutions in empathic contexts -- gpt-5 falls 1.87 points from gpt-4.1 and Opus-4.7 falls 0.629 from Opus-4.6, while aggregate scores stay flat. The regression survives three user-proxy swaps (95% of magnitude), replicates across a 5-family judge stack and a 17-month cohort gap, and persists on 74 held-out real ESConv conversations (rho in [0.749, 0.850]); the instrument reaches ordinal Krippendorff alpha = 0.91. As a by-product, the paradigm acts as a saturation-source diagnostic, separating instrumental ceilings (breakable by rubric refinement) from structural ceilings (needing scenario or roster intervention).

Xinzhe Li, Sun Rui, Yiming Niu et al.

Precipitation prediction plays a crucial role in modern agriculture and industry. However, it poses significant challenges due to the diverse patterns and dynamics in time and space, as well as the scarcity of high precipitation events. To address this challenge, we propose an ensemble learning framework that leverages multiple learners to capture the diverse patterns of precipitation distribution. Specifically, the framework consists of a precipitation predictor with multiple lightweight heads (learners) and a controller that combines the outputs from these heads. The learners and the controller are separately optimized with a proposed 3-stage training scheme. By utilizing provided satellite images, the proposed approach can effectively model the intricate rainfall patterns, especially for high precipitation events. It achieved 1st place on the core test as well as the nowcasting leaderboards of the Weather4Cast 2023 competition. For detailed implementation, please refer to our GitHub repository at: https://github.com/lxz1217/weather4cast-2023-lxz.

Yunyun Wang, Yao Liu, Songcan Chen

Universal domain adaptation (UniDA) is a general unsupervised domain adaptation setting, which addresses both domain and label shifts in adaptation. Its main challenge lies in how to identify target samples in unshared or unknown classes. Previous methods commonly strive to depict sample "confidence" along with a threshold for rejecting unknowns, and align feature distributions of shared classes across domains. However, it is still hard to pre-specify a "confidence" criterion and threshold which are adaptive to various real tasks, and a mis-prediction of unknowns further incurs misalignment of features in shared classes. In this paper, we propose a new UniDA method with adaptive Unknown Authentication by Classifier Paradox (UACP), considering that samples with paradoxical predictions are probably unknowns belonging to none of the source classes. In UACP, a composite classifier is jointly designed with two types of predictors. That is, a multi-class (MC) predictor classifies samples to one of the multiple source classes, while a binary one-vs-all (OVA) predictor further verifies the prediction by MC predictor. Samples with verification failure or paradox are identified as unknowns. Further, instead of feature alignment for shared classes, implicit domain alignment is conducted in output space such that samples across domains share the same decision boundary, though with feature discrepancy. Empirical results validate UACP under both open-set and universal UDA settings.

Yao Liu

TypedCSIP is a typed counterfactual pretraining method for the conflict-classification task of the LCR-CN benchmark (Zhao et al., 2026): given a (superior, subordinate) provision pair, predict whether the pair conflicts and which of four legal-doctrine types (Responsibility, Condition, Sanction, Definition) describes the inconsistency. We exploit LCR-CN's expert-written minimal revisions as training-time counterfactual supervision; at test time the classifier reads only the original pair. Stage 1 pretrains a shared encoder with a typed Counterfactual Selective Intervention Pretraining objective on (superior, subordinate, expert-revised) triplets, treating the expert revision as a counterfactual that the typed factor head must classify as carrying no conflict evidence. Stage 2 transfers the encoder to a five-way classification head. The confirmatory test was registered on the Open Science Framework before observing v6 measurements: 18 seeds, locked rule requiring mean per-seed difference at least 0.8 pp with both seed-bootstrap and Student-t 95% lower bounds above zero. On the 696-record test split, the v2 variant improves macro-F1 over the strongest single-model baseline by +0.916 pp on chinese-roberta-wwm-ext and +1.288 pp on the SAILER cross-backbone replication; both cells pass the rule. A cold-start stratified result on the 244 Unseen-gB records keeps the gain positive on both backbones. A cross-task diagnostic shows the Stage-2 encoder is classification-specialized and does not transfer to LCR-CN's superior-law retrieval task, so we scope the contribution to conflict classification. We release code, 72 pre-registered prediction files, matched-seed and MLM-control auxiliaries, and the OSF pre-registration record.

Yuxiang Liu, Qiao Liu, Tong Luo et al.

Predicting irregularly spaced event sequences with discrete marks poses significant challenges due to the complex, asynchronous dependencies embedded within continuous-time data streams.Existing sequential approaches capture dependencies among event tokens but ignore the continuous evolution between events, while Neural Ordinary Differential Equation (Neural ODE) methods model smooth dynamics yet fail to account for how event types influence future timing.To overcome these limitations, we propose NEXTPP, a dual-channel framework that unifies discrete and continuous representations via Event-granular Neural Evolution with Cross-Interaction for Marked Temporal Point Processes. Specifically, NEXTPP encodes discrete event marks via a self-attention mechanism, simultaneously evolving a latent continuous-time state using a Neural ODE. These parallel streams are then fused through a crossattention module to enable explicit bidirectional interaction between continuous and discrete representations. The fused representations drive the conditional intensity function of the neural Hawkes process, while an iterative thinning sampler is employed to generate future events. Extensive evaluations on five real-world datasets demonstrate that NEXTPP consistently outperforms state-of-the-art models. The source code can be found at https://github.com/AONE-NLP/NEXTPP.

Rui Duan, Zhe Qu, Shangqing Zhao et al.

Recently, adversarial machine learning attacks have posed serious security threats against practical audio signal classification systems, including speech recognition, speaker recognition, and music copyright detection. Previous studies have mainly focused on ensuring the effectiveness of attacking an audio signal classifier via creating a small noise-like perturbation on the original signal. It is still unclear if an attacker is able to create audio signal perturbations that can be well perceived by human beings in addition to its attack effectiveness. This is particularly important for music signals as they are carefully crafted with human-enjoyable audio characteristics. In this work, we formulate the adversarial attack against music signals as a new perception-aware attack framework, which integrates human study into adversarial attack design. Specifically, we conduct a human study to quantify the human perception with respect to a change of a music signal. We invite human participants to rate their perceived deviation based on pairs of original and perturbed music signals, and reverse-engineer the human perception process by regression analysis to predict the human-perceived deviation given a perturbed signal. The perception-aware attack is then formulated as an optimization problem that finds an optimal perturbation signal to minimize the prediction of perceived deviation from the regressed human perception model. We use the perception-aware framework to design a realistic adversarial music attack against YouTube's copyright detector. Experiments show that the perception-aware attack produces adversarial music with significantly better perceptual quality than prior work.

Zexue He, Yu Wang, An Yan et al.

Curated datasets for healthcare are often limited due to the need of human annotations from experts. In this paper, we present MedEval, a multi-level, multi-task, and multi-domain medical benchmark to facilitate the development of language models for healthcare. MedEval is comprehensive and consists of data from several healthcare systems and spans 35 human body regions from 8 examination modalities. With 22,779 collected sentences and 21,228 reports, we provide expert annotations at multiple levels, offering a granular potential usage of the data and supporting a wide range of tasks. Moreover, we systematically evaluated 10 generic and domain-specific language models under zero-shot and finetuning settings, from domain-adapted baselines in healthcare to general-purposed state-of-the-art large language models (e.g., ChatGPT). Our evaluations reveal varying effectiveness of the two categories of language models across different tasks, from which we notice the importance of instruction tuning for few-shot usage of large language models. Our investigation paves the way toward benchmarking language models for healthcare and provides valuable insights into the strengths and limitations of adopting large language models in medical domains, informing their practical applications and future advancements.

Lin Qiu, Hanqing Zeng, Yao Liu et al.

Reinforcement learning (RL) has emerged as an effective paradigm for improving the reasoning capability of vision-language models (VLMs). However, RL-based optimization typically depends on costly high-quality annotations that are difficult to scale. Existing unsupervised alternatives may drift toward biased solutions due to weak visual grounding and the lack of reliable verification signals. We propose a self-evolving post-training framework, DUEL, where supervision emerges from adversarial interactions between two policies initialized from the same pretrained VLM. A Challenger generates an image-grounded true claim together with a minimally perturbed hard-negative counterpart, while a Solver verifies both claims against the image, encouraging fine-grained visual discrimination under near-neighbor semantics. To stabilize optimization, we introduce a length-normalized log-likelihood reward that preserves informative optimization signals beyond binary outcome supervision and improves learning stability under sparse feedback. Experiments show that DUEL consistently improves visual reasoning and robust discrimination without additional human annotations, external reward models, or image editing tools.

Can Xu, Yao Cheng, Jianxiang Yu et al.

Existing methods for graph out-of-distribution (OOD) generalization primarily rely on empirical studies on synthetic datasets. Such approaches tend to overemphasize the causal relationships between invariant sub-graphs and labels, thereby neglecting the non-negligible role of environment in real-world scenarios. In contrast to previous studies that impose rigid independence assumptions on environments and invariant sub-graphs, this paper presents the theorems of environment-label dependency and mutable rationale invariance, where the former characterizes the usefulness of environments in determining graph labels while the latter refers to the mutable importance of graph rationales. Based on analytic investigations, a novel variational inference based method named ``Probability Dependency on Environments and Rationales for OOD Graphs on Real-world Data'' (DEROG) is introduced. To alleviate the adverse effect of unknown prior knowledge on environments and rationales, DEROG utilizes generalized Bayesian inference. Further, DEROG employs an EM-based algorithm for optimization. Finally, extensive experiments on real-world datasets under different distribution shifts are conducted to show the superiority of DEROG. Our code is publicly available at https://github.com/LEOXC1571/DEROG.

Zongshun Zhang, Yao Liu, Qiao Liu et al.

Video-based lie detection aims to identify deceptive behaviors from visual cues. Despite recent progress, its core challenge lies in learning sparse yet discriminative representations. Deceptive signals are typically subtle and short-lived, easily overwhelmed by redundant information, while individual and contextual variations introduce strong identity-related noise. To address this issue, we propose GenLie, a Global-Enhanced Lie Detection Network that performs local feature modeling under global supervision. Specifically, sparse and subtle deceptive cues are captured at the local level, while global supervision and optimization ensure robust and discriminative representations by suppressing identity-related noise. Experiments on three public datasets, covering both high- and low-stakes scenarios, show that GenLie consistently outperforms state-of-the-art methods. Source code is available at https://github.com/AliasDictusZ1/GenLie.

Jessica Echterhoff, Yao Liu, Abeer Alessa et al.

Large language models (LLMs) offer significant potential as tools to support an expanding range of decision-making tasks. Given their training on human (created) data, LLMs have been shown to inherit societal biases against protected groups, as well as be subject to bias functionally resembling cognitive bias. Human-like bias can impede fair and explainable decisions made with LLM assistance. Our work introduces BiasBuster, a framework designed to uncover, evaluate, and mitigate cognitive bias in LLMs, particularly in high-stakes decision-making tasks. Inspired by prior research in psychology and cognitive science, we develop a dataset containing 13,465 prompts to evaluate LLM decisions on different cognitive biases (e.g., prompt-induced, sequential, inherent). We test various bias mitigation strategies, while proposing a novel method utilizing LLMs to debias their own human-like cognitive bias within prompts. Our analysis provides a comprehensive picture of the presence and effects of cognitive bias across commercial and open-source models. We demonstrate that our selfhelp debiasing effectively mitigates model answers that display patterns akin to human cognitive bias without having to manually craft examples for each bias.

Yibo Zhao, Jiapeng Zhu, Can Xu et al.

The rapid growth of social media platforms has raised significant concerns regarding online content toxicity. When Large Language Models (LLMs) are used for toxicity detection, two key challenges emerge: 1) the absence of domain-specific toxic knowledge leads to false negatives; 2) the excessive sensitivity of LLMs to toxic speech results in false positives, limiting freedom of speech. To address these issues, we propose a novel method called MetaTox, leveraging graph search on a meta-toxic knowledge graph to enhance hatred and toxicity detection. First, we construct a comprehensive meta-toxic knowledge graph by utilizing LLMs to extract toxic information through a three-step pipeline, with toxic benchmark datasets serving as corpora. Second, we query the graph via retrieval and ranking processes to supplement accurate, relevant toxic knowledge. Extensive experiments and in-depth case studies across multiple datasets demonstrate that our MetaTox significantly decreases the false positive rate while boosting overall toxicity detection performance. Our code is available at https://github.com/YiboZhao624/MetaTox.

Peiyuan Jiang, Yao Liu, Yanglei Gan et al.

Non-contact automatic deception detection remains challenging because visual and auditory deception cues often lack stable cross-subject patterns. In contrast, galvanic skin response (GSR) provides more reliable physiological cues and has been widely used in contact-based deception detection. In this work, we leverage stable deception-related knowledge in GSR to guide representation learning in non-contact modalities through cross-modal knowledge distillation. A key obstacle, however, is the lack of a suitable dataset for this setting. To address this, we introduce MuDD, a large-scale Multimodal Deception Detection dataset containing recordings from 130 participants over 690 minutes. In addition to video, audio, and GSR, MuDD also provides Photoplethysmography, heart rate, and personality traits, supporting broader scientific studies of deception. Based on this dataset, we propose GSR-guided Progressive Distillation (GPD), a cross-modal distillation framework for mitigating the negative transfer caused by the large modality mismatch between GSR and non-contact signals. The core innovation of GPD is the integration of progressive feature-level and digit-level distillation with dynamic routing, which allows the model to adaptively determine how teacher knowledge should be transferred during training, leading to more stable cross-modal knowledge transfer. Extensive experiments and visualizations show that GPD outperforms existing methods and achieves state-of-the-art performance on both deception detection and concealed-digit identification.

Ruoyu Wang, Yao Liu, Yuanjiang Cao et al.

Current research in Visual Navigation reveals opportunities for improvement. First, the direct adoption of RNNs and Transformers often overlooks the specific differences between Embodied AI and traditional sequential data modelling, potentially limiting its performance in Embodied AI tasks. Second, the reliance on task-specific configurations, such as pre-trained modules and dataset-specific logic, compromises the generalizability of these methods. We address these constraints by initially exploring the unique differences between Navigation tasks and other sequential data tasks through the lens of Causality, presenting a causal framework to elucidate the inadequacies of conventional sequential methods for Navigation. By leveraging this causal perspective, we propose Causality-Aware Transformer (CAT) Networks for Navigation, featuring a Causal Understanding Module to enhance the models's Environmental Understanding capability. Meanwhile, our method is devoid of task-specific inductive biases and can be trained in an End-to-End manner, which enhances the method's generalizability across various contexts. Empirical evaluations demonstrate that our methodology consistently surpasses benchmark performances across a spectrum of settings, tasks and simulation environments. Extensive ablation studies reveal that the performance gains can be attributed to the Causal Understanding Module, which demonstrates effectiveness and efficiency in both Reinforcement Learning and Supervised Learning settings.

Hanwen Zhang, Yao Liu, Die Dai et al.

Fine-grained Vision-Language Pre-training (FVLP) demonstrates significant potential in 3D medical image understanding by aligning anatomy-level visual representations with corresponding textual descriptions. However, existing FVLP paradigms often suffer from severe representation collapse in the textual embedding space, where text embeddings of distinct anatomical structures become highly clustered and indistinguishable. This distributional degeneracy renders the model hypersensitive to prompt variations, hindering reliable clinical deployment. To address these challenges, we propose a novel Cross-Anatomy Global-Local Contrastive Learning framework (CA-GCL). CA-GCL introduces a global contrastive objective that enforces separation between anatomical categories in the latent space, effectively counteracting the aggregation tendency induced by local alignment. Furthermore, we incorporate a clinical-aware text augmentation strategy based on permutation invariance and partial completeness to enhance robustness against descriptive incompleteness. Extensive evaluations on the CT-RATE and Rad-ChestCT datasets demonstrate that CA-GCL consistently outperforms existing VLP paradigms in zero-shot abnormality detection, achieving superior performance while exhibiting strong cross-dataset generalization. Crucially, CA-GCL reduces performance variance across diverse prompt templates, transforming the collapsed textual similarity distribution into a bell-shaped distribution. These results validate CA-GCL as an effective framework for robust 3D medical image understanding.

Peng He, Yao Liu, Yanglei Gan et al.

Sequential recommendation (SR) aims to predict a user's next item preference by modeling historical interaction sequences. Recent advances often integrate frequency-domain modules to compensate for self-attention's low-pass nature by restoring the high-frequency signals critical for personalized recommendations. Nevertheless, existing frequency-aware solutions process each session in isolation and optimize exclusively with time-domain objectives. Consequently, they overlook cross-session spectral dependencies and fail to enforce alignment between predicted and actual spectral signatures, leaving valuable frequency information under-exploited. To this end, we propose FreqRec, a Frequency-Enhanced Dual-Path Network for sequential Recommendation that jointly captures inter-session and intra-session behaviors via a learnable Frequency-domain Multi-layer Perceptrons. Moreover, FreqRec is optimized under a composite objective that combines cross entropy with a frequency-domain consistency loss, explicitly aligning predicted and true spectral signatures. Extensive experiments on three benchmarks show that FreqRec surpasses strong baselines and remains robust under data sparsity and noisy-log conditions.

Zihao Ding, Cheng-Tse Lee, Mufeng Zhu et al.

3D Gaussian Splatting (3DGS) is an emerging media representation that reconstructs real-world 3D scenes in high fidelity, enabling 6-degrees-of-freedom (6-DoF) navigation in virtual reality (VR). However, developing and evaluating 3DGS-enabled applications and optimizing their rendering performance, require realistic user navigation data. Such data is currently unavailable for photorealistic 3DGS reconstructions of real-world scenes. This paper introduces EyeNavGS (EyeNavGS), the first publicly available 6-DoF navigation dataset featuring traces from 46 participants exploring twelve diverse, real-world 3DGS scenes. The dataset was collected at two sites, using the Meta Quest Pro headsets, recording the head pose and eye gaze data for each rendered frame during free world standing 6-DoF navigation. For each of the twelve scenes, we performed careful scene initialization to correct for scene tilt and scale, ensuring a perceptually-comfortable VR experience. We also release our open-source SIBR viewer software fork with record-and-replay functionalities and a suite of utility tools for data processing, conversion, and visualization. The EyeNavGS dataset and its accompanying software tools provide valuable resources for advancing research in 6-DoF viewport prediction, adaptive streaming, 3D saliency, and foveated rendering for 3DGS scenes. The EyeNavGS dataset is available at: https://symmru.github.io/EyeNavGS/.

Yao Liu

Online travel agents (Booking, Trip.com, Expedia) have replaced ranked-list interfaces with conversational LLM agents that compress many options into one sentence of advice. Each booking earns the OTA commission and different suppliers pay different rates: the agent has a structural incentive to favor higher-margin recommendations. Whether any deployed agent does this, and by how much, no one can currently measure. Disclosure banners, conversion A/B testing, UI dark-pattern taxonomies, and generic LLM safety scores were built for older interfaces and miss the prose-recommendation surface where the steering happens. We propose TourMart, an applied intelligent-system audit instrument for LLM-OTA commission governance. Two governance levers -- lambda (gain on message-induced perception in the traveler's accept/reject decision) and kappa (budget-normalized cap on how far the message can shift perceived welfare) -- drive a paired counterfactual: holding the traveler and bundle fixed, the steering delta is read off between a commission-aware prompt and a minimum-disclosure factual template. A symmetric six-gate producer audit separates LLM-engineering failures (template collapse, refusal, internal-ID leakage) from genuine commercial steering. At deployed (lambda=1, kappa=0.05), a Qwen-14B reader shows +7.69pp steering (exact McNemar p=0.003); a Llama-3.1-8B reader shows +3.50pp in the same direction at n=143, with an extended-n supplement (n=270) confirming significance (+2.96pp, p=0.008). Across the (lambda, kappa) grid both arms pass family-wise scenario-clustered correction (p<0.001 / p=0.008). TourMart outputs a sentence a compliance report can quote: "at this deployment, 7.7 extra commission-steered recommendations per 100 paired traveler sessions."

Zihao Ding, Mufeng Zhu, Zhongze Tang et al.

Nowadays, visual intelligence tools have become ubiquitous, offering all kinds of convenience and possibilities. However, these tools have high computational requirements that exceed the capabilities of resource-constrained mobile and wearable devices. While offloading visual data to the cloud is a common solution, it introduces significant privacy vulnerabilities during transmission and server-side computation. To address this, we propose a novel distributed, hierarchical offloading framework for Vision Transformers (ViTs) that addresses these privacy challenges by design. Our approach uses a local trusted edge device, such as a mobile phone or an Nvidia Jetson, as the edge orchestrator. This orchestrator partitions the user's visual data into smaller portions and distributes them across multiple independent cloud servers. By design, no single external server possesses the complete image, preventing comprehensive data reconstruction. The final data merging and aggregation computation occurs exclusively on the user's trusted edge device. We apply our framework to the Segment Anything Model (SAM) as a practical case study, which demonstrates that our method substantially enhances content privacy over traditional cloud-based approaches. Evaluations show our framework maintains near-baseline segmentation performance while substantially reducing the risk of content reconstruction and user data exposure. Our framework provides a scalable, privacy-preserving solution for vision tasks in the edge-cloud continuum.

Yao Liu, Hang Shao, Bing Bai

This paper introduces a new Convolutional Neural Network (ConvNet) architecture inspired by a class of partial differential equations (PDEs) called quasi-linear hyperbolic systems. With comparable performance on the image classification task, it allows for the modification of the weights via a continuous group of symmetry. This is a significant shift from traditional models where the architecture and weights are essentially fixed. We wish to promote the (internal) symmetry as a new desirable property for a neural network, and to draw attention to the PDE perspective in analyzing and interpreting ConvNets in the broader Deep Learning community.

Peiyuan Jiang, Yao Liu, Qiao Liu et al.

Multimodal emotion recognition (MER) aims to identify emotional states by integrating and analyzing information from multiple modalities. However, inherent modality heterogeneity and inconsistencies in emotional cues remain key challenges that hinder performance. To address these issues, we propose a Decoupled Representations with Knowledge Fusion (DRKF) method for MER. DRKF consists of two main modules: an Optimized Representation Learning (ORL) Module and a Knowledge Fusion (KF) Module. ORL employs a contrastive mutual information estimation method with progressive modality augmentation to decouple task-relevant shared representations and modality-specific features while mitigating modality heterogeneity. KF includes a lightweight self-attention-based Fusion Encoder (FE) that identifies the dominant modality and integrates emotional information from other modalities to enhance the fused representation. To handle potential errors from incorrect dominant modality selection under emotionally inconsistent conditions, we introduce an Emotion Discrimination Submodule (ED), which enforces the fused representation to retain discriminative cues of emotional inconsistency. This ensures that even if the FE selects an inappropriate dominant modality, the Emotion Classification Submodule (EC) can still make accurate predictions by leveraging preserved inconsistency information. Experiments show that DRKF achieves state-of-the-art (SOTA) performance on IEMOCAP, MELD, and M3ED. The source code is publicly available at https://github.com/PANPANKK/DRKF.

Hanqi Duan, Yao Cheng, Jianxiang Yu et al.

Graph Neural Networks (GNNs) have emerged as powerful models for learning from graph-structured data. However, GNNs lack the inherent semantic understanding capability of rich textual node attributes, limiting their effectiveness in applications. On the other hand, we empirically observe that for existing GNN models, no one can consistently outperforms others across diverse datasets. In this paper, we study whether LLMs can act as an ensembler for multi-GNNs and propose the LensGNN model. The model first aligns multiple GNNs, mapping the representations of different GNNs into the same space. Then, through LoRA fine-tuning, it aligns the space between the GNN and the LLM, injecting graph tokens and textual information into LLMs. This allows LensGNN to ensemble multiple GNNs and take advantage of the strengths of LLM, leading to a deeper understanding of both textual semantic information and graph structural information. The experimental results show that LensGNN outperforms existing models. This research advances text-attributed graph ensemble learning by providing a robust and superior solution for integrating semantic and structural information. We provide our code and data here: https://github.com/AquariusAQ/LensGNN.

Yabin Zhu, Chenglong Li, Yao Liu et al.

Tiny objects, frequently appearing in practical applications, have weak appearance and features, and receive increasing interests in meany vision tasks, such as object detection and segmentation. To promote the research and development of tiny object tracking, we create a large-scale video dataset, which contains 434 sequences with a total of more than 217K frames. Each frame is carefully annotated with a high-quality bounding box. In data creation, we take 12 challenge attributes into account to cover a broad range of viewpoints and scene complexities, and annotate these attributes for facilitating the attribute-based performance analysis. To provide a strong baseline in tiny object tracking, we propose a novel Multilevel Knowledge Distillation Network (MKDNet), which pursues three-level knowledge distillations in a unified framework to effectively enhance the feature representation, discrimination and localization abilities in tracking tiny objects. Extensive experiments are performed on the proposed dataset, and the results prove the superiority and effectiveness of MKDNet compared with state-of-the-art methods. The dataset, the algorithm code, and the evaluation code are available at https://github.com/mmic-lcl/Datasets-and-benchmark-code.

Haiyun Liu, Shangqing Zhao, Yao Liu et al.

The security of wireless challenge-response Physical Layer Authentication (PLA) based on Orthogonal Frequency Division Multiplexing (OFDM) relies on a sufficiently random fading channel condition, which is commonly assumed in existing studies. However, in practical scenarios, such a condition is not always guaranteed and the responses of OFDM subchannels may exhibit correlation.} Consequently, ensuring the security of such PLA systems remains an unsolved problem. In this paper, we propose a novel adversary model, called Maximum Differential Likelihood Generator (MDLG), which exploits the weak correlation property in practical wireless channel to launch effective attacks against PLA. Based on this model, we create a measurable guideline using randomness testing to decide when we can in fact use PLA in a practical wireless channel condition. Extensive real-world experiments validate the effectiveness of the MDLG attack and demonstrate how the proposed guideline can help protect the security of PLA.

Haiyun Liu, Jiahao Xue, Jie Xu et al.

Today's 5G and NextG wireless networks are moving toward using the coordinated multi-point (CoMP) transmission and reception technique, where a client can be simultaneously served by multiple base stations (BSs) for better communication performance. However, traditional hierarchical federated learning (HFL) architectures impose the constraint that each client can be associated with only one edge server (ES) at a time. If we keep using the traditional HFL architectures in modern hierarchical networks for model training, the benefits of the CoMP technique would remain unexploited and leave room for further improvements in training efficiency. To address this issue, we propose hybrid hierarchical federated learning (HHFL), which allows clients in overlapping regions to simultaneously communicate with multiple edge servers (ESs) for model aggregation. HHFL is able to enhance inter-ES knowledge sharing, thereby mitigating model divergence and improving training efficiency. We provide a rigorous theoretical convergence analysis with a convergence upper bound to validate its effectiveness. Experimental results show that HHFL outperforms traditional HFL, particularly when the data across different ESs is not independent and identically distributed (non-IID). For example, when each ES is dominated by only two of the ten classes and 15 out of the 57 clients can connect to multiple ESs, HHFL achieves up to 2x faster convergence under the same configuration. These results demonstrate that HHFL provides a scalable and efficient solution for FL model training in today's and NextG wireless networks.

Ke Yang, Yao Liu, Sapana Chaudhary et al.

Autonomy via agents using large language models (LLMs) for personalized, standardized tasks boosts human efficiency. Automating web tasks (like booking hotels within a budget) is increasingly sought after. Fulfilling practical needs, the web agent also serves as an important proof-of-concept example for various agent grounding scenarios, with its success promising advancements in many future applications. Prior research often handcrafts web agent strategies (e.g., prompting templates, multi-agent systems, search methods, etc.) and the corresponding in-context examples, which may not generalize well across all real-world scenarios. On the other hand, there has been limited study on the misalignment between a web agent's observation/action representation and the pre-training data of the LLM it's based on. This discrepancy is especially notable when LLMs are primarily trained for language completion rather than tasks involving embodied navigation actions and symbolic web elements. Our study enhances an LLM-based web agent by simply refining its observation and action space to better align with the LLM's capabilities. This approach enables our base agent to significantly outperform previous methods on a wide variety of web tasks. Specifically, on WebArena, a benchmark featuring general-purpose web interaction tasks, our agent AgentOccam surpasses the previous state-of-the-art and concurrent work by 9.8 (+29.4%) and 5.9 (+15.8%) absolute points respectively, and boosts the success rate by 26.6 points (+161%) over similar plain web agents with its observation and action space alignment. We achieve this without using in-context examples, new agent roles, online feedback or search strategies. AgentOccam's simple design highlights LLMs' impressive zero-shot performance on web tasks, and underlines the critical role of carefully tuning observation and action spaces for LLM-based agents.

Jianing Wang, Junda Wu, Yupeng Hou et al.

Do current large language models (LLMs) better solve graph reasoning and generation tasks with parameter updates? In this paper, we propose InstructGraph, a framework that empowers LLMs with the abilities of graph reasoning and generation by instruction tuning and preference alignment. Specifically, we first propose a structured format verbalizer to unify all graph data into a universal code-like format, which can simply represent the graph without any external graph-specific encoders. Furthermore, a graph instruction tuning stage is introduced to guide LLMs in solving graph reasoning and generation tasks. Finally, we identify potential hallucination problems in graph tasks and sample negative instances for preference alignment, the target of which is to enhance the output's reliability of the model. Extensive experiments across multiple graph-centric tasks exhibit that InstructGraph can achieve the best performance and outperform GPT-4 and LLaMA2 by more than 13\% and 38\%, respectively.

Zhepei Wei, Wenlin Yao, Yao Liu et al.

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

Wenwei Zhao, Xiaowen Li, Yao Liu et al.

Federated learning (FL) is vulnerable to poisoning attacks, where malicious clients upload manipulated updates to degrade the performance of the global model. Although detection methods can identify and remove malicious clients, the model remains affected. Retraining from scratch is effective but costly, and existing unlearning methods remain unsatisfactory in both effectiveness and efficiency. We propose Federated Adversarial Unlearning (FAUN), a lightweight framework that retains only a short window of malicious clients' updates and employs adversarial optimization on a proxy dataset to derive updates that eliminate malicious directions. Applying these updates for a few unlearning rounds, followed by benign fine-tuning, enables fast removal of malicious effects and stable recovery. Experiments on three canonical datasets show that FAUN achieves recovery comparable to retraining while requiring far fewer rounds and reduces attack success rates to near zero, confirming FAUN successfully eliminates the contributions of unlearned clients.

Chi Zhang, Changjia Zhu, Xiaowen Li et al.

Text-to-image (T2I) diffusion models have the ability to build high-quality pictures from text prompts, but they pose safety concerns because they can generate offensive or disturbing imagery when provided with harmful inputs. Existing safety filters typically rely on text-based classifiers or image-based checkers that completely block the output upon detecting a threat, issuing an explicit allow/block feedback signal to the user. This binary strategy leaves models vulnerable to adversarial attacks that alter keywords to bypass detection, and it causes high false-alarm rates that degrade the experience for benign users. To address such vulnerabilities, we propose Disciplined Diffusion (DDiffusion), a novel robust text-to-image diffusion that counters Not Safe For Work (NSFW) generation by uncovering implicit malicious semantics in prompt embeddings. DDiffusion leverages a semantic retrieval mechanism to evaluate prompts against concept distributions rather than relying on brittle pairwise similarity. Furthermore, it employs a localization method during the diffusion process to selectively edit only the harmful regions of the generated image. By returning locally sanitized images instead of applying uniform blocking, DDiffusion suppresses malicious content while preserving generation fidelity for benign prompts and avoiding the binary allow-deny signal on which existing probing attacks rely.

Yao Liu, Binghao Li, Xianzhi Wang et al.

Trajectory prediction is fundamental to various intelligent technologies, such as autonomous driving and robotics. The motion prediction of pedestrians and vehicles helps emergency braking, reduces collisions, and improves traffic safety. Current trajectory prediction research faces problems of complex social interactions, high dynamics and multi-modality. Especially, it still has limitations in long-time prediction. We propose Attention-aware Social Graph Transformer Networks for multi-modal trajectory prediction. We combine Graph Convolutional Networks and Transformer Networks by generating stable resolution pseudo-images from Spatio-temporal graphs through a designed stacking and interception method. Furthermore, we design the attention-aware module to handle social interaction information in scenarios involving mixed pedestrian-vehicle traffic. Thus, we maintain the advantages of the Graph and Transformer, i.e., the ability to aggregate information over an arbitrary number of neighbors and the ability to perform complex time-dependent data processing. We conduct experiments on datasets involving pedestrian, vehicle, and mixed trajectories, respectively. Our results demonstrate that our model minimizes displacement errors across various metrics and significantly reduces the likelihood of collisions. It is worth noting that our model effectively reduces the final displacement error, illustrating the ability of our model to predict for a long time.

Zizhan Tang, Yao Liu, Jessica Liu

We present a safety-oriented framework for autonomous underwater vehicles (AUVs) that improves localization accuracy, enhances trajectory prediction, and supports efficient search operations during communication loss. Acoustic signals emitted by the AUV are detected by a network of fixed buoys, which compute Time-Difference-of-Arrival (TDOA) range-difference measurements serving as position observations. These observations are subsequently fused with a Kalman-based prediction model to obtain continuous, noise-robust state estimates. The combined method achieves significantly better localization precision and trajectory stability than TDOA-only baselines. Beyond real-time tracking, our framework offers targeted search-and-recovery capability by predicting post-disconnection motion and explicitly modeling uncertainty growth. The search module differentiates between continued navigation and propulsion failure, allowing search resources to be deployed toward the most probable recovery region. Our framework fuses multi-buoy acoustic data with Kalman filtering and uncertainty propagation to maintain navigation accuracy and yield robust search-region definitions during communication loss.

Jia Zhang, Chen-Xi Zhang, Yao Liu et al.

Recent advancements in instruction tuning for large language models (LLMs) suggest that a small, high-quality dataset can significantly equip LLMs with instruction-following capabilities, outperforming large datasets often burdened by quality and redundancy issues. However, the challenge lies in automatically identifying valuable subsets from large datasets to boost both the effectiveness and efficiency of instruction tuning. In this paper, we first establish data selection criteria based on three distinct aspects of data value: diversity, difficulty, and dependability, and then propose the D3 method comprising two key steps of scoring and selection. Specifically, in the scoring step, we define the diversity function to measure sample distinctiveness and introduce the uncertainty-based prediction difficulty to evaluate sample difficulty by mitigating the interference of context-oriented generation diversity. Additionally, we integrate an external LLM for dependability assessment. In the selection step, we formulate the D3 weighted coreset objective, which jointly optimizes three aspects of data value to solve for the most valuable subset. The two steps of D3 can iterate multiple rounds, incorporating feedback to refine the selection focus adaptively. Experiments on both public datasets and the real-world Taobao Live application demonstrate the effectiveness of D3 in endowing LLMs with competitive or even superior instruction-following capabilities using less than 10\% of the entire dataset.

Hanwen Zhang, Yao Liu, Peiyuan Jiang et al.

Explainable Multimodal Emotion Recognition plays a crucial role in applications such as human-computer interaction and social media analytics. However, current approaches struggle with cue-level perception and reasoning due to two main challenges: 1) general-purpose modality encoders are pretrained to capture global structures and general semantics rather than fine-grained emotional cues, resulting in limited sensitivity to emotional signals; and 2) available datasets usually involve a trade-off between annotation quality and scale, which leads to insufficient supervision for emotional cues and ultimately limits cue-level reasoning. Moreover, existing evaluation metrics are inadequate for assessing cue-level reasoning performance. To address these challenges, we propose eXplainable Emotion GPT (XEmoGPT), a novel EMER framework capable of both perceiving and reasoning over emotional cues. It incorporates two specialized modules: the Video Emotional Cue Bridge (VECB) and the Audio Emotional Cue Bridge (AECB), which enhance the video and audio encoders through carefully designed tasks for fine-grained emotional cue perception. To further support cue-level reasoning, we construct a large-scale dataset, EmoCue, designed to teach XEmoGPT how to reason over multimodal emotional cues. In addition, we introduce EmoCue-360, an automated metric that extracts and matches emotional cues using semantic similarity, and release EmoCue-Eval, a benchmark of 400 expert-annotated samples covering diverse emotional scenarios. Experimental results show that XEmoGPT achieves strong performance in both emotional cue perception and reasoning.

Zhepeng Cen, Yao Liu, Siliang Zeng et al.

Language models are often trained to maximize the likelihood of the next token given past tokens in the training dataset. However, during inference time, they are utilized differently, generating text sequentially and auto-regressively by using previously generated tokens as input to predict the next one. Marginal differences in predictions at each step can cascade over successive steps, resulting in different distributions from what the models were trained for and potentially leading to unpredictable behavior. This paper proposes two simple approaches based on model own generation to address this discrepancy between the training and inference time. Our first approach is Batch-Scheduled Sampling, where, during training, we stochastically choose between the ground-truth token from the dataset and the model's own generated token as input to predict the next token. This is done in an offline manner, modifying the context window by interleaving ground-truth tokens with those generated by the model. Our second approach is Reference-Answer-based Correction, where we explicitly incorporate a self-correction capability into the model during training. This enables the model to effectively self-correct the gaps between the generated sequences and the ground truth data without relying on an external oracle model. By incorporating our proposed strategies during training, we have observed an overall improvement in performance compared to baseline methods, as demonstrated by our extensive experiments using summarization, general question-answering, and math question-answering tasks.