Che Liu

Xiaohang Tang, Keyue Jiang, Che Liu et al.

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well aligned with pre-training, these approaches introduce bias through training--inference mismatch by using the ELBO as a likelihood surrogate, which can degrade performance. In this work, we propose Guided Denoiser Self-Distillation (GDSD) to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to $+19.6\%$. These results suggest that direct denoiser self-distillation, without relying on an ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs. Code is available at https://github.com/GaryBall/GDSD.

Che Liu, Sibo Cheng, Miaojing Shi et al.

In the field of medical Vision-Language Pre-training (VLP), significant efforts have been devoted to deriving text and image features from both clinical reports and associated medical images. However, most existing methods may have overlooked the opportunity in leveraging the inherent hierarchical structure of clinical reports, which are generally split into `findings' for descriptive content and `impressions' for conclusive observation. Instead of utilizing this rich, structured format, current medical VLP approaches often simplify the report into either a unified entity or fragmented tokens. In this work, we propose a novel clinical prior guided VLP framework named IMITATE to learn the structure information from medical reports with hierarchical vision-language alignment. The framework derives multi-level visual features from the chest X-ray (CXR) images and separately aligns these features with the descriptive and the conclusive text encoded in the hierarchical medical report. Furthermore, a new clinical-informed contrastive loss is introduced for cross-modal learning, which accounts for clinical prior knowledge in formulating sample correlations in contrastive learning. The proposed model, IMITATE, outperforms baseline VLP methods across six different datasets, spanning five medical imaging downstream tasks. Comprehensive experimental results highlight the advantages of integrating the hierarchical structure of medical reports for vision-language alignment. The code related to this paper is available at https://github.com/cheliu-computation/IMITATE-TMI2024.

Che Liu, Rui Wang, Junfeng Jiang et al.

In this paper, we introduce the task of learning unsupervised dialogue embeddings. Trivial approaches such as combining pre-trained word or sentence embeddings and encoding through pre-trained language models (PLMs) have been shown to be feasible for this task. However, these approaches typically ignore the conversational interactions between interlocutors, resulting in poor performance. To address this issue, we proposed a self-guided contrastive learning approach named dial2vec. Dial2vec considers a dialogue as an information exchange process. It captures the conversational interaction patterns between interlocutors and leverages them to guide the learning of the embeddings corresponding to each interlocutor. The dialogue embedding is obtained by an aggregation of the embeddings from all interlocutors. To verify our approach, we establish a comprehensive benchmark consisting of six widely-used dialogue datasets. We consider three evaluation tasks: domain categorization, semantic relatedness, and dialogue retrieval. Dial2vec achieves on average 8.7, 9.0, and 13.8 points absolute improvements in terms of purity, Spearman's correlation, and mean average precision (MAP) over the strongest baseline on the three tasks respectively. Further analysis shows that dial2vec obtains informative and discriminative embeddings for both interlocutors under the guidance of the conversational interactions and achieves the best performance when aggregating them through the interlocutor-level pooling strategy. All codes and data are publicly available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/dial2vec.

Che Liu, Anand Shah, Wenjia Bai et al.

Medical Vision-Language Pre-training (VLP) learns representations jointly from medical images and paired radiology reports. It typically requires large-scale paired image-text datasets to achieve effective pre-training for both the image encoder and text encoder. The advent of text-guided generative models raises a compelling question: Can VLP be implemented solely with synthetic images generated from genuine radiology reports, thereby mitigating the need for extensively pairing and curating image-text datasets? In this work, we scrutinize this very question by examining the feasibility and effectiveness of employing synthetic images for medical VLP. We replace real medical images with their synthetic equivalents, generated from authentic medical reports. Utilizing three state-of-the-art VLP algorithms, we exclusively train on these synthetic samples. Our empirical evaluation across three subsequent tasks, namely image classification, semantic segmentation and object detection, reveals that the performance achieved through synthetic data is on par with or even exceeds that obtained with real images. As a pioneering contribution to this domain, we introduce a large-scale synthetic medical image dataset, paired with anonymized real radiology reports. This alleviates the need of sharing medical images, which are not easy to curate and share in practice. The code and the dataset can be found in \href{https://github.com/cheliu-computation/MedSyn-RepLearn/tree/main}{https://github.com/cheliu-computation/MedSyn-RepLearn/tree/main}.

Jun Li, Che Liu, Sibo Cheng et al.

The electrocardiogram (ECG) is one of the most commonly used non-invasive, convenient medical monitoring tools that assist in the clinical diagnosis of heart diseases. Recently, deep learning (DL) techniques, particularly self-supervised learning (SSL), have demonstrated great potential in the classification of ECG. SSL pre-training has achieved competitive performance with only a small amount of annotated data after fine-tuning. However, current SSL methods rely on the availability of annotated data and are unable to predict labels not existing in fine-tuning datasets. To address this challenge, we propose Multimodal ECG-Text Self-supervised pre-training (METS), the first work to utilize the auto-generated clinical reports to guide ECG SSL pre-training. We use a trainable ECG encoder and a frozen language model to embed paired ECG and automatically machine-generated clinical reports separately. The SSL aims to maximize the similarity between paired ECG and auto-generated report while minimize the similarity between ECG and other reports. In downstream classification tasks, METS achieves around 10% improvement in performance without using any annotated data via zero-shot classification, compared to other supervised and SSL baselines that rely on annotated data. Furthermore, METS achieves the highest recall and F1 scores on the MIT-BIH dataset, despite MIT-BIH containing different classes of ECG compared to the pre-trained dataset. The extensive experiments have demonstrated the advantages of using ECG-Text multimodal self-supervised learning in terms of generalizability, effectiveness, and efficiency.

Che Liu, Sibo Cheng, Weiping Ding et al.

Electrocardiography (ECG) signals can be considered as multi-variable time-series. The state-of-the-art ECG data classification approaches, based on either feature engineering or deep learning techniques, treat separately spectral and time domains in machine learning systems. No spectral-time domain communication mechanism inside the classifier model can be found in current approaches, leading to difficulties in identifying complex ECG forms. In this paper, we proposed a novel deep learning model named Spectral Cross-domain neural network (SCDNN) with a new block called Soft-adaptive threshold spectral enhancement (SATSE), to simultaneously reveal the key information embedded in spectral and time domains inside the neural network. More precisely, the domain-cross information is captured by a general Convolutional neural network (CNN) backbone, and different information sources are merged by a self-adaptive mechanism to mine the connection between time and spectral domains. In SATSE, the knowledge from time and spectral domains is extracted via the Fast Fourier Transformation (FFT) with soft trainable thresholds in modified Sigmoid functions. The proposed SCDNN is tested with several classification tasks implemented on the public ECG databases \textit{PTB-XL} and \textit{MIT-BIH}. SCDNN outperforms the state-of-the-art approaches with a low computational cost regarding a variety of metrics in all classification tasks on both databases, by finding appropriate domains from the infinite spectral mapping. The convergence of the trainable thresholds in the spectral domain is also numerically investigated in this paper. The robust performance of SCDNN provides a new perspective to exploit knowledge across deep learning models from time and spectral domains. The repository can be found: https://github.com/DL-WG/SCDNN-TS

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

Sibo Cheng, Cesar Quilodran-Casas, Said Ouala et al.

Data Assimilation (DA) and Uncertainty quantification (UQ) are extensively used in analysing and reducing error propagation in high-dimensional spatial-temporal dynamics. Typical applications span from computational fluid dynamics (CFD) to geoscience and climate systems. Recently, much effort has been given in combining DA, UQ and machine learning (ML) techniques. These research efforts seek to address some critical challenges in high-dimensional dynamical systems, including but not limited to dynamical system identification, reduced order surrogate modelling, error covariance specification and model error correction. A large number of developed techniques and methodologies exhibit a broad applicability across numerous domains, resulting in the necessity for a comprehensive guide. This paper provides the first overview of the state-of-the-art researches in this interdisciplinary field, covering a wide range of applications. This review aims at ML scientists who attempt to apply DA and UQ techniques to improve the accuracy and the interpretability of their models, but also at DA and UQ experts who intend to integrate cutting-edge ML approaches to their systems. Therefore, this article has a special focus on how ML methods can overcome the existing limits of DA and UQ, and vice versa. Some exciting perspectives of this rapidly developing research field are also discussed.

Yuchong Sun, Che Liu, Kun Zhou et al.

Humans often interact with large language models (LLMs) in multi-turn interaction to obtain desired answers or more information. However, most existing studies overlook the multi-turn instruction following ability of LLMs, in terms of training dataset, training method, and evaluation benchmark. In this paper, we introduce Parrot, a solution aiming to enhance multi-turn instruction following for LLMs. First, we introduce an efficient but effective method for collecting multi-turn instructions that feature human-like queries, such as anaphora and ellipsis. Second, we propose a context-aware preference optimization strategy to further enhance LLMs for complex queries in multi-turn interaction. Moreover, to quantitatively evaluate LLMs in multi-turn instruction following, we manually build a multi-turn benchmark derived from existing ones. Extensive experiments show that Parrot improves current LLMs by up to 7.2% in multi-turn instruction following. Our dataset and codes will be open-sourced to facilitate future research.

Che Liu, Sibo Cheng, Chen Chen et al.

Medical vision-language models enable co-learning and integrating features from medical imaging and clinical text. However, these models are not easy to train and the latent representation space can be complex. Here we propose a novel way for pre-training and regularising medical vision-language models. The proposed method, named Medical vision-language pre-training with Frozen language models and Latent spAce Geometry optimization (M-FLAG), leverages a frozen language model for training stability and efficiency and introduces a novel orthogonality loss to harmonize the latent space geometry. We demonstrate the potential of the pre-trained model on three downstream tasks: medical image classification, segmentation, and object detection. Extensive experiments across five public datasets demonstrate that M-FLAG significantly outperforms existing medical vision-language pre-training approaches and reduces the number of parameters by 78\%. Notably, M-FLAG achieves outstanding performance on the segmentation task while using only 1\% of the RSNA dataset, even outperforming ImageNet pre-trained models that have been fine-tuned using 100\% of the data.

Yinda Chen, Che Liu, Wei Huang et al.

Vision-Language Pretraining (VLP) has demonstrated remarkable capabilities in learning visual representations from textual descriptions of images without annotations. Yet, effective VLP demands large-scale image-text pairs, a resource that suffers scarcity in the medical domain. Moreover, conventional VLP is limited to 2D images while medical images encompass diverse modalities, often in 3D, making the learning process more challenging. To address these challenges, we present Generative Text-Guided 3D Vision-Language Pretraining for Unified Medical Image Segmentation (GTGM), a framework that extends of VLP to 3D medical images without relying on paired textual descriptions. Specifically, GTGM utilizes large language models (LLM) to generate medical-style text from 3D medical images. This synthetic text is then used to supervise 3D visual representation learning. Furthermore, a negative-free contrastive learning objective strategy is introduced to cultivate consistent visual representations between augmented 3D medical image patches, which effectively mitigates the biases associated with strict positive-negative sample pairings. We evaluate GTGM on three imaging modalities - Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and electron microscopy (EM) over 13 datasets. GTGM's superior performance across various medical image segmentation tasks underscores its effectiveness and versatility, by enabling VLP extension into 3D medical imagery while bypassing the need for paired text.

Tianbao Xie, Fan Zhou, Zhoujun Cheng et al.

Language agents show potential in being capable of utilizing natural language for varied and intricate tasks in diverse environments, particularly when built upon large language models (LLMs). Current language agent frameworks aim to facilitate the construction of proof-of-concept language agents while neglecting the non-expert user access to agents and paying little attention to application-level designs. We present OpenAgents, an open platform for using and hosting language agents in the wild of everyday life. OpenAgents includes three agents: (1) Data Agent for data analysis with Python/SQL and data tools; (2) Plugins Agent with 200+ daily API tools; (3) Web Agent for autonomous web browsing. OpenAgents enables general users to interact with agent functionalities through a web user interface optimized for swift responses and common failures while offering developers and researchers a seamless deployment experience on local setups, providing a foundation for crafting innovative language agents and facilitating real-world evaluations. We elucidate the challenges and opportunities, aspiring to set a foundation for future research and development of real-world language agents.

Jun Li, Mingxuan Liu, Jiazhen Pan et al.

Clinical abnormality grounding for rare diseases is often hindered by data scarcity, making supervised fine-tuning impractical and single-pass inference highly unstable. We propose Dynamic Decision Learning (DDL), a framework that enables frozen large vision-language models (LVLMs) to refine their decisions across both language and visual spaces by optimizing instructions and consolidating predictions under visual perturbations. This process improves localization quality and produces a consensus-based reliability score that quantifies model confidence. Results on brain imaging benchmarks, including a rare-disease dataset with 281 pathology types across models ranging from 3B to 72B parameters, show that DDL improves mAP@75 by up to 105% on rare-disease cases and outperforms adaptation baselines and supervised fine-tuning. Furthermore, DDL demonstrates stronger calibration between reliability scores and localization accuracy under severe distribution shifts and increasing task difficulty. Code is available at: https://lijunrio.github.io/DDL/

Che Liu, Zhongwei Wan, Sibo Cheng et al.

In the domain of cardiovascular healthcare, the Electrocardiogram (ECG) serves as a critical, non-invasive diagnostic tool. Although recent strides in self-supervised learning (SSL) have been promising for ECG representation learning, these techniques often require annotated samples and struggle with classes not present in the fine-tuning stages. To address these limitations, we introduce ECG-Text Pre-training (ETP), an innovative framework designed to learn cross-modal representations that link ECG signals with textual reports. For the first time, this framework leverages the zero-shot classification task in the ECG domain. ETP employs an ECG encoder along with a pre-trained language model to align ECG signals with their corresponding textual reports. The proposed framework excels in both linear evaluation and zero-shot classification tasks, as demonstrated on the PTB-XL and CPSC2018 datasets, showcasing its ability for robust and generalizable cross-modal ECG feature learning.

Sibo Cheng, Che Liu, Yike Guo et al.

Variational Data Assimilation (DA) has been broadly used in engineering problems for field reconstruction and prediction by performing a weighted combination of multiple sources of noisy data. In recent years, the integration of deep learning (DL) techniques in DA has shown promise in improving the efficiency and accuracy in high-dimensional dynamical systems. Nevertheless, existing deep DA approaches face difficulties in dealing with unstructured observation data, especially when the placement and number of sensors are dynamic over time. We introduce a novel variational DA scheme, named Voronoi-tessellation Inverse operator for VariatIonal Data assimilation (VIVID), that incorporates a DL inverse operator into the assimilation objective function. By leveraging the capabilities of the Voronoi-tessellation and convolutional neural networks, VIVID is adept at handling sparse, unstructured, and time-varying sensor data. Furthermore, the incorporation of the DL inverse operator establishes a direct link between observation and state space, leading to a reduction in the number of minimization steps required for DA. Additionally, VIVID can be seamlessly integrated with Proper Orthogonal Decomposition (POD) to develop an end-to-end reduced-order DA scheme, which can further expedite field reconstruction. Numerical experiments in a fluid dynamics system demonstrate that VIVID can significantly outperform existing DA and DL algorithms. The robustness of VIVID is also accessed through the application of various levels of prior error, the utilization of varying numbers of sensors, and the misspecification of error covariance in DA.

Sibo Cheng, Jinyang Min, Che Liu et al.

Data assimilation techniques are often confronted with challenges handling complex high dimensional physical systems, because high precision simulation in complex high dimensional physical systems is computationally expensive and the exact observation functions that can be applied in these systems are difficult to obtain. It prompts growing interest in integrating deep learning models within data assimilation workflows, but current software packages for data assimilation cannot handle deep learning models inside. This study presents a novel Python package seamlessly combining data assimilation with deep neural networks to serve as models for state transition and observation functions. The package, named TorchDA, implements Kalman Filter, Ensemble Kalman Filter (EnKF), 3D Variational (3DVar), and 4D Variational (4DVar) algorithms, allowing flexible algorithm selection based on application requirements. Comprehensive experiments conducted on the Lorenz 63 and a two-dimensional shallow water system demonstrate significantly enhanced performance over standalone model predictions without assimilation. The shallow water analysis validates data assimilation capabilities mapping between different physical quantity spaces in either full space or reduced order space. Overall, this innovative software package enables flexible integration of deep learning representations within data assimilation, conferring a versatile tool to tackle complex high dimensional dynamical systems across scientific domains.

Jiao Ou, Junda Lu, Che Liu et al.

Large language models (LLMs) have achieved remarkable breakthroughs in new dialogue capabilities by leveraging instruction tuning, which refreshes human impressions of dialogue systems. The long-standing goal of dialogue systems is to be human-like enough to establish long-term connections with users. Therefore, there has been an urgent need to evaluate LLMs as human-like dialogue systems. In this paper, we propose DialogBench, a dialogue evaluation benchmark that contains 12 dialogue tasks to probe the capabilities of LLMs as human-like dialogue systems should have. Specifically, we prompt GPT-4 to generate evaluation instances for each task. We first design the basic prompt based on widely used design principles and further mitigate the existing biases to generate higher-quality evaluation instances. Our extensive tests on English and Chinese DialogBench of 26 LLMs show that instruction tuning improves the human likeness of LLMs to a certain extent, but most LLMs still have much room for improvement as human-like dialogue systems. Interestingly, results also show that the positioning of assistant AI can make instruction tuning weaken the human emotional perception of LLMs and their mastery of information about human daily life.

Weixiang Shen, Yanzhu Hu, Che Liu et al.

Currently, evaluating vision-language models (VLMs) in medical imaging tasks oversimplifies clinical reality by relying on pre-selected 2D images that demand significant manual labor to curate. This setup misses the core challenge of realworld diagnostics: a true clinical agent must actively navigate full 3D volumes across multiple sequences or modalities to gather evidence and ultimately support a final decision. To address this, we propose MEDOPENCLAW, an auditable runtime designed to let VLMs operate dynamically within standard medical tools or viewers (e.g., 3D Slicer). On top of this runtime, we introduce MEDFLOWBENCH, a full-study medical imaging benchmark covering multi-sequence brain MRI and lung CT/PET. It systematically evaluates medical agentic capabilities across viewer-only, tool-use, and open-method tracks. Initial results reveal a critical insight: while state-of-the-art LLMs/VLMs (e.g., Gemini 3.1 Pro and GPT-5.4) can successfully navigate the viewer to solve basic study-level tasks, their performance paradoxically degrades when given access to professional support tools due to a lack of precise spatial grounding. By bridging the gap between static-image perception and interactive clinical workflows, MEDOPENCLAW and MEDFLOWBENCH establish a reproducible foundation for developing auditable, full-study medical imaging agents.

Xiaoyuan Cheng, Wenxuan Yuan, Zhancun Mu et al.

Model-based reinforcement learning (RL) can be effectively supported at scale through the use of world models. However, in practice, scaling such approaches remains fundamentally limited. A commonly recognized challenge is model bias and error compounding, which degrade long-horizon predictions. Beyond these issues, we identify a more critical yet underexplored bottleneck: a structural misalignment between search and value learning in existing world model approaches. In particular, policy improvement often relies on value functions induced by a separate, non-search policy, resulting in training inconsistency and ultimately suboptimal learning. To address this limitation, we propose Model-Based Diffusion Policy Optimization (MBDPO) in world models, a framework that unifies search and policy optimization through diffusion policy representations, thereby unlocking the potential of world models for scalable policy learning. Instead of constructing an explicit planner over a learned world model, we reformulate policy optimization as a diffusion process over searched trajectories in latent world models. In this view, we extract an implicit energy function from the collected dataset that anchors the policy, enabling MBDPO to refine the score field for policy optimization while mitigating misalignment. We evaluate MBDPO across a wide range of settings, including multi-task offline pretraining, online learning, and offline-to-online fine-tuning. In the offline regime, we further investigate its scaling behavior by pretraining on large-scale datasets, observing consistent and monotonic performance gains with increasing model capacity.

Zhongwei Wan, Xin Wang, Che Liu et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in important tasks such as natural language understanding and language generation, and thus have the potential to make a substantial impact on our society. Such capabilities, however, come with the considerable resources they demand, highlighting the strong need to develop effective techniques for addressing their efficiency challenges. In this survey, we provide a systematic and comprehensive review of efficient LLMs research. We organize the literature in a taxonomy consisting of three main categories, covering distinct yet interconnected efficient LLMs topics from model-centric, data-centric, and framework-centric perspective, respectively. We have also created a GitHub repository where we organize the papers featured in this survey at https://github.com/AIoT-MLSys-Lab/Efficient-LLMs-Survey. We will actively maintain the repository and incorporate new research as it emerges. We hope our survey can serve as a valuable resource to help researchers and practitioners gain a systematic understanding of efficient LLMs research and inspire them to contribute to this important and exciting field.

Sicheng Wang, Che Liu, Rossella Arcucci

Recent advancements in medical vision-language pre-training (MedVLP) have significantly enhanced zero-shot medical vision tasks such as image classification by leveraging large-scale medical image-text pair pre-training. However, the performance of these tasks can be heavily influenced by the variability in textual prompts describing the categories, necessitating robustness in MedVLP models to diverse prompt styles. Yet, this sensitivity remains underexplored. In this work, we are the first to systematically assess the sensitivity of three widely-used MedVLP methods to a variety of prompts across 15 different diseases. To achieve this, we designed six unique prompt styles to mirror real clinical scenarios, which were subsequently ranked by interpretability. Our findings indicate that all MedVLP models evaluated show unstable performance across different prompt styles, suggesting a lack of robustness. Additionally, the models' performance varied with increasing prompt interpretability, revealing difficulties in comprehending complex medical concepts. This study underscores the need for further development in MedVLP methodologies to enhance their robustness to diverse zero-shot prompts.

Jiazhen Pan, Che Liu, Junde Wu et al.

Reasoning is a critical frontier for advancing medical image analysis, where transparency and trustworthiness play a central role in both clinician trust and regulatory approval. Although Medical Visual Language Models (VLMs) show promise for radiological tasks, most existing VLMs merely produce final answers without revealing the underlying reasoning. To address this gap, we introduce MedVLM-R1, a medical VLM that explicitly generates natural language reasoning to enhance transparency and trustworthiness. Instead of relying on supervised fine-tuning (SFT), which often suffers from overfitting to training distributions and fails to foster genuine reasoning, MedVLM-R1 employs a reinforcement learning framework that incentivizes the model to discover human-interpretable reasoning paths without using any reasoning references. Despite limited training data (600 visual question answering samples) and model parameters (2B), MedVLM-R1 boosts accuracy from 55.11% to 78.22% across MRI, CT, and X-ray benchmarks, outperforming larger models trained on over a million samples. It also demonstrates robust domain generalization under out-of-distribution tasks. By unifying medical image analysis with explicit reasoning, MedVLM-R1 marks a pivotal step toward trustworthy and interpretable AI in clinical practice. Inference model is available at: https://huggingface.co/JZPeterPan/MedVLM-R1.

Yihong Tang, Jiao Ou, Che Liu et al.

Role-playing is an emerging application in the field of Human-Computer Interaction (HCI), primarily implemented through the alignment training of a large language model (LLM) with assigned characters. Despite significant progress, role-playing agents (RPLAs) still struggle with maintaining role-consistency across conversations, particularly when confronted with boundary queries subtly related to character attributes. In this paper, we present ERABAL, a framework aimed at enhancing RPLAs' role-playing capabilities through boundary-aware learning. ERABAL encompasses a generation pipeline for role-specific dialogues and a concomitant methodology for alignment training. Through comprehensive evaluations, we demonstrate that ERABAL is both efficient and effective. By training with significantly fewer dialogues than those used in leading approaches, ERABAL achieves notable improvements across WikiRoleEval, CharacterEval, and the role-playing subset of MT-Bench compared to the generalist baseline models. Our code and datasets will be made publicly available to support further research.

Yi Zhang, Che Liu, Xiancong Ren et al.

This report presents Pelican-VL 1.0, a new family of open-source embodied brain models with parameter scales ranging from 7 billion to 72 billion. Our explicit mission is clearly stated as: To embed powerful intelligence into various embodiments. Pelican-VL 1.0 is currently the largest-scale open-source embodied multimodal brain model. Its core advantage lies in the in-depth integration of data power and intelligent adaptive learning mechanisms. Specifically, metaloop distilled a high-quality dataset from a raw dataset containing 4+ billion tokens. Pelican-VL 1.0 is trained on a large-scale cluster of 1000+ A800 GPUs, consuming over 50k+ A800 GPU-hours per checkpoint. This translates to a 20.3% performance uplift from its base model and outperforms 100B-level open-source counterparts by 10.6%, placing it on par with leading proprietary systems on well-known embodied benchmarks. We establish a novel framework, DPPO (Deliberate Practice Policy Optimization), inspired by human metacognition to train Pelican-VL 1.0. We operationalize this as a metaloop that teaches the AI to practice deliberately, which is a RL-Refine-Diagnose-SFT loop.

Che Liu, Zhongwei Wan, Cheng Ouyang et al.

Electrocardiograms (ECGs) are non-invasive diagnostic tools crucial for detecting cardiac arrhythmic diseases in clinical practice. While ECG Self-supervised Learning (eSSL) methods show promise in representation learning from unannotated ECG data, they often overlook the clinical knowledge that can be found in reports. This oversight and the requirement for annotated samples for downstream tasks limit eSSL's versatility. In this work, we address these issues with the Multimodal ECG Representation Learning (MERL}) framework. Through multimodal learning on ECG records and associated reports, MERL is capable of performing zero-shot ECG classification with text prompts, eliminating the need for training data in downstream tasks. At test time, we propose the Clinical Knowledge Enhanced Prompt Engineering (CKEPE) approach, which uses Large Language Models (LLMs) to exploit external expert-verified clinical knowledge databases, generating more descriptive prompts and reducing hallucinations in LLM-generated content to boost zero-shot classification. Based on MERL, we perform the first benchmark across six public ECG datasets, showing the superior performance of MERL compared against eSSL methods. Notably, MERL achieves an average AUC score of 75.2% in zero-shot classification (without training data), 3.2% higher than linear probed eSSL methods with 10\% annotated training data, averaged across all six datasets. Code and models are available at https://github.com/cheliu-computation/MERL

Yinda Chen, Che Liu, Xiaoyu Liu et al.

The burgeoning integration of 3D medical imaging into healthcare has led to a substantial increase in the workload of medical professionals. To assist clinicians in their diagnostic processes and alleviate their workload, the development of a robust system for retrieving similar case studies presents a viable solution. While the concept holds great promise, the field of 3D medical text-image retrieval is currently limited by the absence of robust evaluation benchmarks and curated datasets. To remedy this, our study presents a groundbreaking dataset, {BIMCV-R}, which includes an extensive collection of 8,069 3D CT volumes, encompassing over 2 million slices, paired with their respective radiological reports. Expanding upon the foundational work of our dataset, we craft a retrieval strategy, MedFinder. This approach employs a dual-stream network architecture, harnessing the potential of large language models to advance the field of medical image retrieval beyond existing text-image retrieval solutions. It marks our preliminary step towards developing a system capable of facilitating text-to-image, image-to-text, and keyword-based retrieval tasks. Our project is available at \url{https://huggingface.co/datasets/cyd0806/BIMCV-R}.

Boyong Wu, Chao Yan, Chen Hu et al.

This paper presents Step-Audio 2, an end-to-end multi-modal large language model designed for industry-strength audio understanding and speech conversation. By integrating a latent audio encoder and reasoning-centric reinforcement learning (RL), Step-Audio 2 achieves promising performance in automatic speech recognition (ASR) and audio understanding. To facilitate genuine end-to-end speech conversation, Step-Audio 2 incorporates the generation of discrete audio tokens into language modeling, significantly enhancing its responsiveness to paralinguistic information such as speaking styles and emotions. To effectively leverage the rich textual and acoustic knowledge in real-world data, Step-Audio 2 integrates retrieval-augmented generation (RAG) and is able to call external tools such as web search to mitigate hallucination and audio search to switch timbres. Trained on millions of hours of speech and audio data, Step-Audio 2 delivers intelligence and expressiveness across diverse conversational scenarios. Evaluation results demonstrate that Step-Audio 2 achieves state-of-the-art performance on various audio understanding and conversational benchmarks compared to other open-source and commercial solutions. Please visit https://github.com/stepfun-ai/Step-Audio2 for more information.

Ziliang Gan, Yu Lu, Dong Zhang et al.

In recent years, multimodal benchmarks for general domains have guided the rapid development of multimodal models on general tasks. However, the financial field has its peculiarities. It features unique graphical images (e.g., candlestick charts, technical indicator charts) and possesses a wealth of specialized financial knowledge (e.g., futures, turnover rate). Therefore, benchmarks from general fields often fail to measure the performance of multimodal models in the financial domain, and thus cannot effectively guide the rapid development of large financial models. To promote the development of large financial multimodal models, we propose MME-Finance, an bilingual open-ended and practical usage-oriented Visual Question Answering (VQA) benchmark. The characteristics of our benchmark are finance and expertise, which include constructing charts that reflect the actual usage needs of users (e.g., computer screenshots and mobile photography), creating questions according to the preferences in financial domain inquiries, and annotating questions by experts with 10+ years of experience in the financial industry. Additionally, we have developed a custom-designed financial evaluation system in which visual information is first introduced in the multi-modal evaluation process. Extensive experimental evaluations of 19 mainstream MLLMs are conducted to test their perception, reasoning, and cognition capabilities. The results indicate that models performing well on general benchmarks cannot do well on MME-Finance; for instance, the top-performing open-source and closed-source models obtain 65.69 (Qwen2VL-72B) and 63.18 (GPT-4o), respectively. Their performance is particularly poor in categories most relevant to finance, such as candlestick charts and technical indicator charts. In addition, we propose a Chinese version, which helps compare performance of MLLMs under a Chinese context.

Weixiang Shen, Bailiang Jian, Jun Li et al.

Tool-augmented large language model (LLM) agents can orchestrate specialist classifiers, segmentation models, and visual question-answering modules to interpret chest X-rays. However, these agents still solve each case in isolation: they fail to accumulate experience across cases, correct recurrent reasoning mistakes, or adapt their tool-use behavior without expensive reinforcement learning. While a radiologist naturally improves with every case, current agents remain static. In this work, we propose Evo-MedAgent, a self-evolving memory module that equips a medical agent with the capacity for inter-case learning at test time. Our memory comprises three complementary stores: (1)~\emph{Retrospective Clinical Episodes} that retrieve problem-solving experiences from similar past cases, (2)~an \emph{Adaptive Procedural Heuristics} bank curating priority-tagged diagnostic rules that evolves via reflection, much like a physician refining their internal criteria, and (3)~a \emph{Tool Reliability Controller} that tracks per-tool trustworthiness. On ChestAgentBench, Evo-MedAgent raises multiple-choice question (MCQ) accuracy from 0.68 to 0.79 on GPT-5-mini, and from 0.76 to 0.87 on Gemini-3 Flash. With a strong base model, evolving memory improves performance more effectively than orchestrating external tools on qualitative diagnostic tasks. Because Evo-MedAgent requires no training, its per-case overhead is bounded by one additional retrieval pass and a single reflection call, making it deployable on top of any frozen model.

Zhongwei Wan, Che Liu, Xin Wang et al.

Electrocardiogram (ECG) is the primary non-invasive diagnostic tool for monitoring cardiac conditions and is crucial in assisting clinicians. Recent studies have concentrated on classifying cardiac conditions using ECG data but have overlooked ECG report generation, which is time-consuming and requires clinical expertise. To automate ECG report generation and ensure its versatility, we propose the Multimodal ECG Instruction Tuning (MEIT) framework, the first attempt to tackle ECG report generation with LLMs and multimodal instructions. To facilitate future research, we establish a benchmark to evaluate MEIT with various LLMs backbones across two large-scale ECG datasets. Our approach uniquely aligns the representations of the ECG signal and the report, and we conduct extensive experiments to benchmark MEIT with nine open-source LLMs using more than 800,000 ECG reports. MEIT's results underscore the superior performance of instruction-tuned LLMs, showcasing their proficiency in quality report generation, zero-shot capabilities, resilience to signal perturbation, and alignment with human expert evaluation. These findings emphasize the efficacy of MEIT and its potential for real-world clinical application.

Zhongwei Wan, Hui Shen, Xin Wang et al.

Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial resources as their multimodal Key-Value (KV) caches grow with increasing input lengths, challenging inference efficiency. Existing methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, thus often adopting uniform or progressive reduction strategies for layer-wise cache allocation. In this work, we propose MEDA, a dynamic layer-wise KV cache allocation method for efficient multimodal long-context inference. As its core, MEDA utilizes cross-modal attention entropy to determine the KV cache size at each MLLMs layer. Given the dynamically allocated KV cache size at each layer, MEDA also employs a KV pair selection scheme to identify which KV pairs to select and a KV pair merging strategy that merges the selected and non-selected ones to preserve information from the entire context. MEDA achieves up to 72% KV cache memory reduction and 2.82 times faster decoding speed, while maintaining or enhancing performance on various multimodal tasks in long-context settings, including multi-images and long-video scenarios. Our code is released at https://github.com/AIoT-MLSys-Lab/MEDA.

Yi Zhang, Yinda Chen, Che Liu et al.

We present Pelican-Unified 1.0, the first embodied foundation model trained according to the principle of unification. Pelican-Unified 1.0 uses a single VLM as a unified understanding module, mapping scenes, instructions, visual contexts, and action histories into a shared semantic space. The same VLM also serves as a unified reasoning module, autoregressively producing task-, action-, and future-oriented chains of thought in a single forward pass and projecting the final hidden state into a dense latent variable. A Unified Future Generator (UFG) then conditions on this latent variable and jointly generates future videos and future actions through two modality-specific output heads within the same denoising process. The language, video, and action losses are all backpropagated into the shared representation, enabling the model to jointly optimize understanding, reasoning, imagination, and action during training, rather than training three isolated expert systems. Experiments demonstrate that unification does not imply compromise. With a single checkpoint, Pelican-Unified 1.0 achieves strong performance across all three capabilities: 64.7 on eight VLM benchmarks, the best among comparable-scale models; 66.03 on WorldArena, ranking first; and 93.5 on RoboTwin, the second-best average among compared action methods. These results show that the unified paradigm succeeds in preserving specialist strength while bringing understanding, reasoning, imagination, and action into one model.

Jian Yang, Wei Zhang, Shawn Guo et al.

In this report, we introduce the IQuest-Coder-V1 series-(7B/14B/40B/40B-Loop), a new family of code large language models (LLMs). Moving beyond static code representations, we propose the code-flow multi-stage training paradigm, which captures the dynamic evolution of software logic through different phases of the pipeline. Our models are developed through the evolutionary pipeline, starting with the initial pre-training consisting of code facts, repository, and completion data. Following that, we implement a specialized mid-training stage that integrates reasoning and agentic trajectories in 32k-context and repository-scale in 128k-context to forge deep logical foundations. The models are then finalized with post-training of specialized coding capabilities, which is bifurcated into two specialized paths: the thinking path (utilizing reasoning-driven RL) and the instruct path (optimized for general assistance). IQuest-Coder-V1 achieves state-of-the-art performance among competitive models across critical dimensions of code intelligence: agentic software engineering, competitive programming, and complex tool use. To address deployment constraints, the IQuest-Coder-V1-Loop variant introduces a recurrent mechanism designed to optimize the trade-off between model capacity and deployment footprint, offering an architecturally enhanced path for efficacy-efficiency trade-off. We believe the release of the IQuest-Coder-V1 series, including the complete white-box chain of checkpoints from pre-training bases to the final thinking and instruction models, will advance research in autonomous code intelligence and real-world agentic systems.

Minghe Shen, Zhuo Zhi, Chonghan Liu et al.

While Vision-Language Models (VLMs) post-trained with Reinforcement Learning (RL) show impressive general reasoning, their evaluation is often confined to language-dominant tasks (e.g., math). This raises a critical question: can RL post-training truly extend the inherent capability boundary of a base VLM, particularly for visual-centric spatial tasks where it initially fails? To investigate this, we introduce Ariadne, a framework utilizing synthetic mazes for multi-step spatial reasoning where task difficulty (e.g., path length, turns) is precisely controlled. We leverage this controllable environment to train VLMs using Reinforcement Learning with Verified Rewards (RLVR) in a difficulty-aware curriculum. Surprisingly, post-RLVR training, the VLM achieves over 50% accuracy on a problem set where the base model scored 0%, demonstrating that our approach expands the model's initial capability boundary. To assess real-world viability, we evaluate out-of-distribution (OOD) generalization on practical benchmarks. Despite training only on synthetic maze samples, Ariadne achieves significant zero-shot improvements, averaging 16% on MapBench (e.g., museum navigation) and 24% on ReasonMap (subway transfer tasks). These results confirm that our method not only broadens the model's fundamental limits but also enhances its generalization to real-world spatial reasoning. We acknowledge our study is limited to the post-training phase, given the opaqueness of pre-training data, and hope our research motivates further work on specialized, capability-extending alignment.

Xiaoyu Chen, Changde Du, Che Liu et al.

Decoding language information from brain signals represents a vital research area within brain-computer interfaces, particularly in the context of deciphering the semantic information from the fMRI signal. However, many existing efforts concentrate on decoding small vocabulary sets, leaving space for the exploration of open vocabulary continuous text decoding. In this paper, we introduce a novel method, the \textbf{Brain Prompt GPT (BP-GPT)}. By using the brain representation that is extracted from the fMRI as a prompt, our method can utilize GPT-2 to decode fMRI signals into stimulus text. Further, we introduce a text-to-text baseline and align the fMRI prompt to the text prompt. By introducing the text-to-text baseline, our BP-GPT can extract a more robust brain prompt and promote the decoding of pre-trained LLM. We evaluate our BP-GPT on the open-source auditory semantic decoding dataset and achieve a significant improvement up to $4.61\%$ on METEOR and $2.43\%$ on BERTScore across all the subjects compared to the state-of-the-art method. The experimental results demonstrate that using brain representation as a prompt to further drive LLM for auditory neural decoding is feasible and effective.

Jiazhen Pan, Bailiang Jian, Paul Hager et al.

Ensuring the safety and reliability of large language models (LLMs) in clinical practice is critical to prevent patient harm and promote trustworthy healthcare applications of AI. However, LLMs are advancing so rapidly that static safety benchmarks often become obsolete upon publication, yielding only an incomplete and sometimes misleading picture of model trustworthiness. We demonstrate that a Dynamic, Automatic, and Systematic (DAS) red-teaming framework that continuously stress-tests LLMs can reveal significant weaknesses of current LLMs across four safety-critical domains: robustness, privacy, bias/fairness, and hallucination. A suite of adversarial agents is applied to autonomously mutate test cases, identify/evolve unsafe-triggering strategies, and evaluate responses, uncovering vulnerabilities in real time without human intervention. Applying DAS to 15 proprietary and open-source LLMs revealed a stark contrast between static benchmark performance and vulnerability under adversarial pressure. Despite a median MedQA accuracy exceeding 80\%, 94\% of previously correct answers failed our dynamic robustness tests. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86\% of scenarios, cognitive-bias priming altered clinical recommendations in 81\% of fairness tests, and we identified hallucination rates exceeding 66\% in widely used models. Such profound residual risks are incompatible with routine clinical practice. By converting red-teaming from a static checklist into a dynamic stress-test audit, DAS red-teaming offers the surveillance that hospitals/regulators/technology vendors require as LLMs become embedded in patient chatbots, decision-support dashboards, and broader healthcare workflows. Our framework delivers an evolvable, scalable, and reliable safeguard for the next generation of medical AI.

Xiaoyu Chen, Changde Du, Che Liu et al.

Decoding language information from brain signals represents a vital research area within brain-computer interfaces, particularly in the context of deciphering the semantic information from the fMRI signal. Although existing work uses LLM to achieve this goal, their method does not use an end-to-end approach and avoids the LLM in the mapping of fMRI-to-text, leaving space for the exploration of the LLM in auditory decoding. In this paper, we introduce a novel method, the Brain Prompt GPT (BP-GPT). By using the brain representation that is extracted from the fMRI as a prompt, our method can utilize GPT-2 to decode fMRI signals into stimulus text. Further, we introduce the text prompt and align the fMRI prompt to it. By introducing the text prompt, our BP-GPT can extract a more robust brain prompt and promote the decoding of pre-trained LLM. We evaluate our BP-GPT on the open-source auditory semantic decoding dataset and achieve a significant improvement up to 4.61 on METEOR and 2.43 on BERTScore across all the subjects compared to the state-of-the-art method. The experimental results demonstrate that using brain representation as a prompt to further drive LLM for auditory neural decoding is feasible and effective. The code is available at https://github.com/1994cxy/BP-GPT.

Che Liu, Lichao Ma, Xiangyu Tony Zhang et al.

Omni-modal language models are intended to jointly understand audio, visual inputs, and language, but benchmark gains can be inflated when visual evidence alone is enough to answer a query. We study whether current omni-modal benchmarks separate visual shortcuts from genuine audio-visual-language evidence integration, and how post-training behaves under a visually debiased evaluation setting. We audit nine omni-modal benchmarks with visual-only probing, remove visually solvable queries, and retain full subsets when filtering is undefined or would make comparisons unstable. This yields OmniClean, a cleaned evaluation view with 8,551 retained queries from 16,968 audited queries. On OmniClean, we evaluate OmniBoost, a three-stage post-training recipe based on Qwen2.5-Omni-3B: mixed bi-modal SFT, mixed-modality RLVR, and SFT on self-distilled data. Balanced bi-modal SFT gives limited and uneven gains, RLVR provides the first broad improvement, and self-distillation reshapes the benchmark profile. After SFT on self-distilled data, the 3B model reaches performance comparable to, and in aggregate slightly above, Qwen3-Omni-30B-A3B-Instruct without using a stronger omni-modal teacher. These results show that omni-modal progress is easier to interpret when evaluation controls visual leakage, and that small omni-modal models can benefit from staged post-training with self-distilled omni-query supervision.

Yi Zhang, Che Liu, Xiancong Ren et al.

Developing a universal and versatile embodied intelligence system presents two primary challenges: the critical embodied data bottleneck, where real-world data is scarce and expensive, and the algorithmic inefficiency of existing methods, which are resource-prohibitive. To address these limitations, we introduce Deliberate Practice Policy Optimization (DPPO), a metacognitive ``Metaloop'' training framework that dynamically alternates between supervised fine-tuning (competence expansion) and reinforcement learning (skill refinement). This enables automatic weakness identification and targeted resource allocation, specifically designed to maximize learning efficiency from sparse, finite data. Theoretically, DPPO can be formalised as a unified preference-learning framework. Empirically, training a vision-language embodied model with DPPO, referred to as Pelican-VL 1.0, yields a 20.3% performance improvement over the base model and surpasses open-source models at the 100B-parameter scale by 10.6%. We are open-sourcing both the models and code, providing the first systematic framework that alleviates the data and resource bottleneck and enables the community to build versatile embodied agents efficiently.

Jiarui Jin, Xiaocheng Fang, Haoyu Wang et al.

Myocardial infarction is a critical manifestation of coronary artery disease, yet detecting it from single-lead electrocardiogram (ECG) remains challenging due to limited spatial information. An intuitive idea is to convert single-lead into multiple-lead ECG for classification by pre-trained models, but generative methods optimized at the signal level in most cases leave a large latent space gap, ultimately degrading diagnostic performance. This naturally raises the question of whether latent space alignment could help. However, most prior ECG alignment methods focus on learning transformation invariance, which mismatches the goal of single-lead detection. To address this issue, we propose SelfMIS, a simple yet effective alignment learning framework to improve myocardial infarction detection from single-lead ECG. Discarding manual data augmentations, SelfMIS employs a self-cutting strategy to pair multiple-lead ECG with their corresponding single-lead segments and directly align them in the latent space. This design shifts the learning objective from pursuing transformation invariance to enriching the single-lead representation, explicitly driving the single-lead ECG encoder to learn a representation capable of inferring global cardiac context from the local signal. Experimentally, SelfMIS achieves superior performance over baseline models across nine myocardial infarction types while maintaining a simpler architecture and lower computational overhead, thereby substantiating the efficacy of direct latent space alignment. Our code and checkpoint will be publicly available after acceptance.

Che Liu, Jiazhen Pan, Weixiang Shen et al.

Vision-Language Models (VLMs) trained on web-scale corpora excel at natural image tasks and are increasingly repurposed for healthcare; however, their competence in medical tasks remains underexplored. We present a comprehensive evaluation of open-source general-purpose and medically specialised VLMs, ranging from 3B to 72B parameters, across eight benchmarks: MedXpert, OmniMedVQA, PMC-VQA, PathVQA, MMMU, SLAKE, and VQA-RAD. To observe model performance across different aspects, we first separate it into understanding and reasoning components. Three salient findings emerge. First, large general-purpose models already match or surpass medical-specific counterparts on several benchmarks, demonstrating strong zero-shot transfer from natural to medical images. Second, reasoning performance is consistently lower than understanding, highlighting a critical barrier to safe decision support. Third, performance varies widely across benchmarks, reflecting differences in task design, annotation quality, and knowledge demands. No model yet reaches the reliability threshold for clinical deployment, underscoring the need for stronger multimodal alignment and more rigorous, fine-grained evaluation protocols.

Weiying Lin, Che Liu, Xin Zhang et al.

Implementing precise detection of oil leaks in peak load equipment through image analysis can significantly enhance inspection quality and ensure the system's safety and reliability. However, challenges such as varying shapes of oil-stained regions, background noise, and fluctuating lighting conditions complicate the detection process. To address this, the integration of logical rule-based discrimination into image recognition has been proposed. This approach involves recognizing the spatial relationships among objects to semantically segment images of oil spills using a Mask RCNN network. The process begins with histogram equalization to enhance the original image, followed by the use of Mask RCNN to identify the preliminary positions and outlines of oil tanks, the ground, and areas of potential oil contamination. Subsequent to this identification, the spatial relationships between these objects are analyzed. Logical rules are then applied to ascertain whether the suspected areas are indeed oil spills. This method's effectiveness has been confirmed by testing on images captured from peak power equipment in the field. The results indicate that this approach can adeptly tackle the challenges in identifying oil-contaminated areas, showing a substantial improvement in accuracy compared to existing methods.

Zhongwei Wan, Zhihao Dou, Che Liu et al.

Multimodal large language models (MLLMs) have shown promising capabilities in reasoning tasks, yet still struggle with complex problems requiring explicit self-reflection and self-correction, especially compared to their unimodal text-based counterparts. Existing reflection methods are simplistic and struggle to generate meaningful and instructive feedback, as the reasoning ability and knowledge limits of pre-trained models are largely fixed during initial training. To overcome these challenges, we propose Multimodal Self-Reflection enhanced reasoning with Group Relative Policy Optimization (SRPO), a two-stage reflection-aware reinforcement learning (RL) framework explicitly designed to enhance multimodal LLM reasoning. In the first stage, we construct a high-quality, reflection-focused dataset under the guidance of an advanced MLLM, which generates reflections based on initial responses to help the policy model learn both reasoning and self-reflection. In the second stage, we introduce a novel reward mechanism within the GRPO framework that encourages concise and cognitively meaningful reflection while avoiding redundancy. Extensive experiments across multiple multimodal reasoning benchmarks, including MathVista, MathVision, MathVerse, and MMMU-Pro, using Qwen-2.5-VL-7B and Qwen-2.5-VL-32B demonstrate that SRPO significantly outperforms state-of-the-art models, achieving notable improvements in both reasoning accuracy and reflection quality.

Canyu Chen, Jian Yu, Shan Chen et al. · harvard

Large Language Models (LLMs) hold great promise to revolutionize current clinical systems for their superior capacities on medical text processing tasks and medical licensing exams. Meanwhile, traditional ML models such as SVM and XGBoost have still been mainly adopted in clinical prediction tasks. An emerging question is Can LLMs beat traditional ML models in clinical prediction? Thus, we build a new benchmark ClinicalBench to comprehensively study the clinical predictive modeling capacities of both general-purpose and medical LLMs, and compare them with traditional ML models. ClinicalBench embraces three common clinical prediction tasks, two databases, 14 general-purpose LLMs, 8 medical LLMs, and 11 traditional ML models. Through extensive empirical investigation, we discover that both general-purpose and medical LLMs, even with different model scales, diverse prompting or fine-tuning strategies, still cannot beat traditional ML models in clinical prediction yet, shedding light on their potential deficiency in clinical reasoning and decision-making. We call for caution when practitioners adopt LLMs in clinical applications. ClinicalBench can be utilized to bridge the gap between LLMs' development for healthcare and real-world clinical practice.

Hongjin Su, Shuyang Jiang, Yuhang Lai et al.

Recently the retrieval-augmented generation (RAG) has been successfully applied in code generation. However, existing pipelines for retrieval-augmented code generation (RACG) employ static knowledge bases with a single source, limiting the adaptation capabilities of Large Language Models (LLMs) to domains they have insufficient knowledge of. In this work, we develop a novel pipeline, EVOR, that employs the synchronous evolution of both queries and diverse knowledge bases. On two realistic settings where the external knowledge is required to solve code generation tasks, we compile four new datasets associated with frequently updated libraries and long-tail programming languages, named EVOR-BENCH. Extensive experiments demonstrate that EVOR achieves two to four times of execution accuracy compared to other methods such as Reflexion (Shinn et al., 2024), DocPrompting (Zhou et al., 2023), etc. We demonstrate that EVOR is flexible and can be easily combined with them to achieve further improvement. Further analysis reveals that EVOR benefits from the synchronous evolution of queries and documents and the diverse information sources in the knowledge base. We hope that our studies will inspire more insights into the design of advanced RACG pipelines in future research. Our model, code, and data are available at https://arks-codegen.github.io.

Haozhe Wang, Qixin Xu, Che Liu et al.

Reinforcement Learning (RL) has proven highly effective at enhancing the complex reasoning abilities of Large Language Models (LLMs), yet underlying mechanisms driving this success remain largely opaque. Our analysis reveals that puzzling phenomena like ``aha moments", ``length-scaling'' and entropy dynamics are not disparate occurrences but hallmarks of an emergent reasoning hierarchy, akin to the separation of high-level strategic planning from low-level procedural execution in human cognition. We uncover a compelling two-phase dynamic: initially, a model is constrained by procedural correctness and must improve its low-level skills. The learning bottleneck then decisively shifts, with performance gains being driven by the exploration and mastery of high-level strategic planning. This insight exposes a core inefficiency in prevailing RL algorithms like GRPO, which apply optimization pressure agnostically and dilute the learning signal across all tokens. To address this, we propose Hierarchy-Aware Credit Assignment (HICRA), an algorithm that concentrates optimization efforts on high-impact planning tokens. Our extensive experiments validate that HICRA significantly outperforms strong baselines, and offer deep insights into how reasoning advances through the lens of strategic exploration.

Che Liu, Haozhe Wang, Jiazhen Pan et al.

Improving performance on complex tasks and enabling interpretable decision making in large language models (LLMs), especially for clinical applications, requires effective reasoning. Yet this remains challenging without supervised fine-tuning (SFT) on costly chain-of-thought (CoT) data distilled from closed-source models (e.g., GPT-4o). In this work, we present AlphaMed, the first medical LLM to show that reasoning capability can emerge purely through reinforcement learning (RL), using minimalist rule-based rewards on public multiple-choice QA datasets, without relying on SFT or distilled CoT data. AlphaMed achieves state-of-the-art results on six medical QA benchmarks, outperforming models trained with conventional SFT+RL pipelines. On challenging benchmarks (e.g., MedXpert), AlphaMed even surpasses larger or closed-source models such as DeepSeek-V3-671B and Claude-3.5-Sonnet. To understand the factors behind this success, we conduct a comprehensive data-centric analysis guided by three questions: (i) Can minimalist rule-based RL incentivize reasoning without distilled CoT supervision? (ii) How do dataset quantity and diversity impact reasoning? (iii) How does question difficulty shape the emergence and generalization of reasoning? Our findings show that dataset informativeness is a key driver of reasoning performance, and that minimalist RL on informative, multiple-choice QA data is effective at inducing reasoning without CoT supervision. We also observe divergent trends across benchmarks, underscoring limitations in current evaluation and the need for more challenging, reasoning-oriented medical QA benchmarks.

Sibo Cheng, Marc Bocquet, Weiping Ding et al.

Unstructured grid data are essential for modelling complex geometries and dynamics in computational physics. Yet, their inherent irregularity presents significant challenges for conventional machine learning (ML) techniques. This paper provides a comprehensive review of advanced ML methodologies designed to handle unstructured grid data in high-dimensional dynamical systems. Key approaches discussed include graph neural networks, transformer models with spatial attention mechanisms, interpolation-integrated ML methods, and meshless techniques such as physics-informed neural networks. These methodologies have proven effective across diverse fields, including fluid dynamics and environmental simulations. This review is intended as a guidebook for computational scientists seeking to apply ML approaches to unstructured grid data in their domains, as well as for ML researchers looking to address challenges in computational physics. It places special focus on how ML methods can overcome the inherent limitations of traditional numerical techniques and, conversely, how insights from computational physics can inform ML development. To support benchmarking, this review also provides a summary of open-access datasets of unstructured grid data in computational physics. Finally, emerging directions such as generative models with unstructured data, reinforcement learning for mesh generation, and hybrid physics-data-driven paradigms are discussed to inspire future advancements in this evolving field.

Che Liu, Zhongwei Wan, Haozhe Wang et al.

Medical Vision-Language Pre-training (MedVLP) has made significant progress in enabling zero-shot tasks for medical image understanding. However, training MedVLP models typically requires large-scale datasets with paired, high-quality image-text data, which are scarce in the medical domain. Recent advancements in Large Language Models (LLMs) and diffusion models have made it possible to generate large-scale synthetic image-text pairs. This raises the question: "Can MedVLP succeed using purely synthetic data?" To address this, we use off-the-shelf generative models to create synthetic radiology reports and paired Chest X-ray (CXR) images, and propose an automated pipeline to build a diverse, high-quality synthetic dataset, enabling a rigorous study that isolates model and training settings, focusing entirely from the data perspective. Our results show that MedVLP models trained exclusively on synthetic data outperform those trained on real data by 3.8% in averaged AUC on zero-shot classification. Moreover, using a combination of synthetic and real data leads to a further improvement of 9.07%. Additionally, MedVLP models trained on synthetic or mixed data consistently outperform those trained on real data in zero-shot grounding, as well as in fine-tuned classification and segmentation tasks. Our analysis suggests MedVLP trained on well-designed synthetic data can outperform models trained on real datasets, which may be limited by low-quality samples and long-tailed distributions.

Jiao Ou, Jiayu Wu, Che Liu et al.

Aligning large language models (LLMs) with human expectations requires high-quality instructional dialogues, which usually require instructions that are diverse and in-depth. Existing methods leverage two LLMs to interact for automatic collection: one simulating a user to pose instructions, and the other acting as a system agent to respond. However, these user simulators struggle to model the rules behind how dialogues can pose different instructions without explicit guidance, resulting in general instructions. In this paper, we propose to explicitly capture the complex rules to help the user simulator pose diverse and in-depth instruction. Specifically, we first induce high-level instruction strategies from various real instruction dialogues serving as rules. Afterward, different possible strategies are applied to the newly given dialogue scenario deductively to pose various instructions. Experimental results show that our method can generate diverse and in-depth instructions. The constructed multi-turn instructional dialogues can outperform competitive baselines on the downstream chat model.