Jingzhou He

Ziqi Zhao, Xinyu Ma, Liu Yang et al.

On-policy self-distillation (OPSD) improves the reasoning performance of large language models (LLMs) by providing dense token-level supervision for on-policy rollouts. However, existing OPSD methods often yield limited gains on in-domain reasoning and generalize poorly to out-of-domain problems. We identify two key causes: conditioning the self-teacher on a verified solution encourages imitation of training-domain reference trajectories rather than error-specific correction, and applying distillation to the full response can overwrite valid reasoning prefixes and reinforce overfitting. We propose Reflective On-policy Self-Distillation (ROSD), a framework that turns reference-solution imitation into targeted reasoning correction through reflection-guided, error-localized distillation. For each rollout, ROSD uses a self-reflector to extract a corrective idea and locate the first erroneous span. The corrective idea guides the self-teacher toward targeted supervision, while the localized error span restricts distillation to where correction is needed. This design corrects flawed reasoning while preserving valid prefixes. Experiments on multiple in-domain and out-of-domain reasoning benchmarks show that ROSD yields stronger in-domain reasoning performance overall and substantially better out-of-domain generalization than standard OPSD. Code is available at https://github.com/ZiqiZhao1/ROSD.

Xiaomin Fang, Fan Wang, Lihang Liu et al. · baidu

AI-based protein structure prediction pipelines, such as AlphaFold2, have achieved near-experimental accuracy. These advanced pipelines mainly rely on Multiple Sequence Alignments (MSAs) as inputs to learn the co-evolution information from the homologous sequences. Nonetheless, searching MSAs from protein databases is time-consuming, usually taking dozens of minutes. Consequently, we attempt to explore the limits of fast protein structure prediction by using only primary sequences of proteins. HelixFold-Single is proposed to combine a large-scale protein language model with the superior geometric learning capability of AlphaFold2. Our proposed method, HelixFold-Single, first pre-trains a large-scale protein language model (PLM) with thousands of millions of primary sequences utilizing the self-supervised learning paradigm, which will be used as an alternative to MSAs for learning the co-evolution information. Then, by combining the pre-trained PLM and the essential components of AlphaFold2, we obtain an end-to-end differentiable model to predict the 3D coordinates of atoms from only the primary sequence. HelixFold-Single is validated in datasets CASP14 and CAMEO, achieving competitive accuracy with the MSA-based methods on the targets with large homologous families. Furthermore, HelixFold-Single consumes much less time than the mainstream pipelines for protein structure prediction, demonstrating its potential in tasks requiring many predictions. The code of HelixFold-Single is available at https://github.com/PaddlePaddle/PaddleHelix/tree/dev/apps/protein_folding/helixfold-single, and we also provide stable web services on https://paddlehelix.baidu.com/app/drug/protein-single/forecast.

Longbin Ji, Guan Wang, Xuan Wei et al.

Joint audio-video generation aims to synthesize temporally synchronized and semantically coherent visual-acoustic content. However, existing open-source methods mainly rely on either dual-tower designs with posterior alignment or fully unified tri-modal designs that mix textual context, audio and video in one shared space. The former weakens fine-grained audio-video co-evolution, while the latter couples semantic conditioning with low-level synchronization. To address these limitations, we propose NAVA, a Native Audio-Visual Alignment framework for joint audio-video generation. NAVA is built upon context-conditioned native audio-visual alignment: it first establishes audio-video correspondence in a dedicated interaction space, and then uses external context to condition the joint denoising process. Specifically, NAVA is instantiated with an Align-then-Fuse MMDiT architecture, which transitions from modality-aware audio-video alignment to modality-shared joint denoising. Furthermore, we introduce Timbre-in-Context Conditioning to associate reference timbre cues with corresponding speech spans to achieve controllable speech timbre. Experiments on Verse-Bench and Seed-TTS, together with a user study, demonstrate that NAVA achieves superior video quality, precise audio-visual synchronization, competitive audio quality, and stronger reference-timbre controllability using only 6.3B parameters.

Haifeng Wang, Hua Wu, Tian Wu et al.

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

Lihang Liu, Donglong He, Xiaomin Fang et al.

Molecular property prediction is a fundamental task in the drug and material industries. Physically, the properties of a molecule are determined by its own electronic structure, which is a quantum many-body system and can be exactly described by the Schr"odinger equation. Full-range many-body interactions between electrons have been proven effective in obtaining an accurate solution of the Schr"odinger equation by classical computational chemistry methods, although modeling such interactions consumes an expensive computational cost. Meanwhile, deep learning methods have also demonstrated their competence in molecular property prediction tasks. Inspired by the classical computational chemistry methods, we design a novel method, namely GEM-2, which comprehensively considers full-range many-body interactions in molecules. Multiple tracks are utilized to model the full-range interactions between the many-bodies with different orders, and a novel axial attention mechanism is designed to approximate the full-range interaction modeling with much lower computational cost. Extensive experiments demonstrate the overwhelming superiority of GEM-2 over multiple baseline methods in quantum chemistry and drug discovery tasks. The ablation studies also verify the effectiveness of the full-range many-body interactions.

Lihang Liu, Shanzhuo Zhang, Donglong He et al.

Protein-ligand structure prediction is an essential task in drug discovery, predicting the binding interactions between small molecules (ligands) and target proteins (receptors). Recent advances have incorporated deep learning techniques to improve the accuracy of protein-ligand structure prediction. Nevertheless, the experimental validation of docking conformations remains costly, it raises concerns regarding the generalizability of these deep learning-based methods due to the limited training data. In this work, we show that by pre-training on a large-scale docking conformation generated by traditional physics-based docking tools and then fine-tuning with a limited set of experimentally validated receptor-ligand complexes, we can obtain a protein-ligand structure prediction model with outstanding performance. Specifically, this process involved the generation of 100 million docking conformations for protein-ligand pairings, an endeavor consuming roughly 1 million CPU core days. The proposed model, HelixDock, aims to acquire the physical knowledge encapsulated by the physics-based docking tools during the pre-training phase. HelixDock has been rigorously benchmarked against both physics-based and deep learning-based baselines, demonstrating its exceptional precision and robust transferability in predicting binding confirmation. In addition, our investigation reveals the scaling laws governing pre-trained protein-ligand structure prediction models, indicating a consistent enhancement in performance with increases in model parameters and the volume of pre-training data. Moreover, we applied HelixDock to several drug discovery-related tasks to validate its practical utility. HelixDock demonstrates outstanding capabilities on both cross-docking and structure-based virtual screening benchmarks.

Jing Hu, Danxiang Zhu, Xianlong Luo et al.

Large Audio Language Models (LALMs) have garnered significant research interest. Despite being built upon text-based large language models (LLMs), LALMs frequently exhibit a degradation in knowledge and reasoning capabilities. We hypothesize that this limitation stems from the failure of current training paradigms to effectively bridge the acoustic-semantic gap within the feature representation space. To address this challenge, we propose CORD, a unified alignment framework that performs online cross-modal self-distillation. Specifically, it aligns audio-conditioned reasoning with its text-conditioned counterpart within a unified model. Leveraging the text modality as an internal teacher, CORD performs multi-granularity alignment throughout the audio rollout process. At the token level, it employs on-policy reverse KL divergence with importance-aware weighting to prioritize early and semantically critical tokens. At the sequence level, CORD introduces a judge-based global reward to optimize complete reasoning trajectories via Group Relative Policy Optimization (GRPO). Empirical results across multiple benchmarks demonstrate that CORD consistently enhances audio-conditioned reasoning and substantially bridges the audio-text performance gap with only 80k synthetic training samples, validating the efficacy and data efficiency of our on-policy, multi-level cross-modal alignment approach.

Hongwei Chen, Yishu Lei, Dan Zhang et al.

Test-time scaling has emerged as a promising paradigm in language modeling, wherein additional computational resources are allocated during inference to enhance model performance. Recent approaches, such as DeepConf, have demonstrated the efficacy of this strategy, however, they often incur substantial computational overhead to achieve competitive results. In this work, we propose MatryoshkaThinking, a novel method that significantly reduces computational cost while maintaining state-of-the-art performance. Specifically, MatryoshkaThinking attains a score of 99.79 on AIME2025 using only 4% of the computation required by DeepConf. The core of our approach lies in the recursive exploitation of the model's intrinsic capabilities in reasoning, verification, and summarization, which collectively enhance the retention of correct solutions and reduce the disparity between Pass@k and Pass@1. Comprehensive evaluations across multiple open-source models and challenging multi-modal reasoning benchmarks validate the effectiveness and generality of our method. These findings offer new insights into the design of efficient and scalable test-time inference strategies for advanced language models.

Dan Zhang, Yishu Lei, Jing Hu et al.

We present Eureka-Audio, a compact yet high-performance audio language model that achieves competitive performance against models that are 4 to 18 times larger across a broad range of audio understanding benchmarks. Despite containing only 1.7B parameters, Eureka-Audio demonstrates strong performance on automatic speech recognition (ASR), audio understanding, and dense audio captioning, matching or surpassing multiple 7B to 30B audio and omni-modal baselines. The model adopts a unified end-to-end architecture composed of a lightweight language backbone, a Whisper-based audio encoder, and a sparsely activated Mixture-of-Experts (MoE) adapter that explicitly accounts for audio heterogeneity and alleviates cross-modal optimization conflicts under limited capacity. To further enhance paralinguistic reasoning, we introduce DataFlux, a closed loop audio instruction data synthesis and verification pipeline that constructs high quality, logically consistent supervision from raw audio. Extensive evaluations across ASR, knowledge reasoning, safety, instruction following, and paralinguistic benchmarks, demonstrate that Eureka-Audio achieves an efficient balance between computational cost and performance. These results establish Eureka Audio as a strong and practical baseline for lightweight audio understanding models.

Yuchen Zhu, Jihong Chen, Yitong Li et al.

Structural assessment of biomolecular complexes is vital for translating molecular models into functional insights, shaping our understanding of biology and aiding drug discovery. However, current structure-based scoring functions often lack generalizability across diverse biomolecular systems. We present BioScore, a foundational scoring function that addresses key challenges -- data sparsity, cross-system representation, and task compatibility -- through a dual-scale geometric graph learning framework with tailored modules for structure assessment and affinity prediction. BioScore supports a wide range of tasks, including affinity prediction, conformation ranking, and structure-based virtual screening. Evaluated on 16 benchmarks spanning proteins, nucleic acids, small molecules, and carbohydrates, BioScore consistently outperforms or matches 70 traditional and deep learning methods. Our newly proposed PPI Benchmark further enables comprehensive evaluation of protein-protein complex scoring. BioScore demonstrates broad applicability: (1) pretraining on mixed-structure data boosts protein-protein affinity prediction by up to 40% and antigen-antibody binding correlation by over 90%; (2) cross-system generalizability enables zero- and few-shot prediction with up to 71% correlation gain; and (3) its unified representation captures chemically challenging systems such as cyclic peptides, improving affinity prediction by over 60%. BioScore establishes a robust and generalizable framework for structural assessment across complex biomolecular landscapes.