Controlled LLM Training on Spectral Sphere

Scaling large models requires optimization strategies that ensure rapid convergence grounded in stability. Maximal Update Parametrization ($\boldsymbolμ$P) provides a theoretical safeguard for width-invariant $Θ(1)$ activation control, whereas emerging optimizers like Muon are only ``half-aligned'' with these constraints: they control updates but allow weights to drift. To address this limitation, we introduce the \textbf{Spectral Sphere Optimizer (SSO)}, which enforces strict module-wise spectral constraints on both weights and their updates. By deriving the steepest descent direction on the spectral sphere, SSO realizes a fully $\boldsymbolμ$P-aligned optimization process. To enable large-scale training, we implement SSO as an efficient parallel algorithm within Megatron. Through extensive pretraining on diverse architectures, including Dense 1.7B, MoE 8B-A1B, and 200-layer DeepNet models, SSO consistently outperforms AdamW and Muon. Furthermore, we observe significant practical stability benefits, including improved MoE router load balancing, suppressed outliers, and strictly bounded activations.

Entropy-Guided Token Dropout: Training Autoregressive Language Models with Limited Domain Data

As access to high-quality, domain-specific data grows increasingly scarce, multi-epoch training has become a practical strategy for adapting large language models (LLMs). However, autoregressive models often suffer from performance degradation under repeated data exposure, where overfitting leads to a marked decline in model capability. Through empirical analysis, we trace this degradation to an imbalance in learning dynamics: predictable, low-entropy tokens are learned quickly and come to dominate optimization, while the model's ability to generalize on high-entropy tokens deteriorates with continued training. To address this, we introduce EntroDrop, an entropy-guided token dropout method that functions as structured data regularization. EntroDrop selectively masks low-entropy tokens during training and employs a curriculum schedule to adjust regularization strength in alignment with training progress. Experiments across model scales from 0.6B to 8B parameters show that EntroDrop consistently outperforms standard regularization baselines and maintains robust performance throughout extended multi-epoch training. These findings underscore the importance of aligning regularization with token-level learning dynamics when training on limited data. Our approach offers a promising pathway toward more effective adaptation of LLMs in data-constrained domains.

ReaSeq: Unleashing World Knowledge via Reasoning for Sequential Modeling

Industrial recommender systems face two fundamental limitations under the log-driven paradigm: (1) knowledge poverty in ID-based item representations that causes brittle interest modeling under data sparsity, and (2) systemic blindness to beyond-log user interests that constrains model performance within platform boundaries. These limitations stem from an over-reliance on shallow interaction statistics and close-looped feedback while neglecting the rich world knowledge about product semantics and cross-domain behavioral patterns that Large Language Models have learned from vast corpora. To address these challenges, we introduce ReaSeq, a reasoning-enhanced framework that leverages world knowledge in Large Language Models to address both limitations through explicit and implicit reasoning. Specifically, ReaSeq employs explicit Chain-of-Thought reasoning via multi-agent collaboration to distill structured product knowledge into semantically enriched item representations, and latent reasoning via Diffusion Large Language Models to infer plausible beyond-log behaviors. Deployed on Taobao's ranking system serving hundreds of millions of users, ReaSeq achieves substantial gains: >6.0% in IPV and CTR, >2.9% in Orders, and >2.5% in GMV, validating the effectiveness of world-knowledge-enhanced reasoning over purely log-driven approaches.

A FEDformer-Based Hybrid Framework for Anomaly Detection and Risk Forecasting in Financial Time Series

Financial markets are inherently volatile and prone to sudden disruptions such as market crashes, flash collapses, and liquidity crises. Accurate anomaly detection and early risk forecasting in financial time series are therefore crucial for preventing systemic instability and supporting informed investment decisions. Traditional deep learning models, such as LSTM and GRU, often fail to capture long-term dependencies and complex periodic patterns in highly nonstationary financial data. To address this limitation, this study proposes a FEDformer-Based Hybrid Framework for Anomaly Detection and Risk Forecasting in Financial Time Series, which integrates the Frequency Enhanced Decomposed Transformer (FEDformer) with a residual-based anomaly detector and a risk forecasting head. The FEDformer module models temporal dynamics in both time and frequency domains, decomposing signals into trend and seasonal components for improved interpretability. The residual-based detector identifies abnormal fluctuations by analyzing prediction errors, while the risk head predicts potential financial distress using learned latent embeddings. Experiments conducted on the S&P 500, NASDAQ Composite, and Brent Crude Oil datasets (2000-2024) demonstrate the superiority of the proposed model over benchmark methods, achieving a 15.7 percent reduction in RMSE and an 11.5 percent improvement in F1-score for anomaly detection. These results confirm the effectiveness of the model in capturing financial volatility, enabling reliable early-warning systems for market crash prediction and risk management.

ICPC-Eval: Probing the Frontiers of LLM Reasoning with Competitive Programming Contests

With the significant progress of large reasoning models in complex coding and reasoning tasks, existing benchmarks, like LiveCodeBench and CodeElo, are insufficient to evaluate the coding capabilities of large language models (LLMs) in real competition environments. Moreover, current evaluation metrics such as Pass@K fail to capture the reflective abilities of reasoning models. To address these challenges, we propose \textbf{ICPC-Eval}, a top-level competitive coding benchmark designed to probing the frontiers of LLM reasoning. ICPC-Eval includes 118 carefully curated problems from 11 recent ICPC contests held in various regions of the world, offering three key contributions: 1) A challenging realistic ICPC competition scenario, featuring a problem type and difficulty distribution consistent with actual contests. 2) A robust test case generation method and a corresponding local evaluation toolkit, enabling efficient and accurate local evaluation. 3) An effective test-time scaling evaluation metric, Refine@K, which allows iterative repair of solutions based on execution feedback. The results underscore the significant challenge in evaluating complex reasoning abilities: top-tier reasoning models like DeepSeek-R1 often rely on multi-turn code feedback to fully unlock their in-context reasoning potential when compared to non-reasoning counterparts. Furthermore, despite recent advancements in code generation, these models still lag behind top-performing human teams. We release the benchmark at: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs

MARIO: A Mixed Annotation Framework For Polyp Segmentation

Existing polyp segmentation models are limited by high labeling costs and the small size of datasets. Additionally, vast polyp datasets remain underutilized because these models typically rely on a single type of annotation. To address this dilemma, we introduce MARIO, a mixed supervision model designed to accommodate various annotation types, significantly expanding the range of usable data. MARIO learns from underutilized datasets by incorporating five forms of supervision: pixel-level, box-level, polygon-level, scribblelevel, and point-level. Each form of supervision is associated with a tailored loss that effectively leverages the supervision labels while minimizing the noise. This allows MARIO to move beyond the constraints of relying on a single annotation type. Furthermore, MARIO primarily utilizes dataset with weak and cheap annotations, reducing the dependence on large-scale, fully annotated ones. Experimental results across five benchmark datasets demonstrate that MARIO consistently outperforms existing methods, highlighting its efficacy in balancing trade-offs between different forms of supervision and maximizing polyp segmentation performance

YuLan-Mini: An Open Data-efficient Language Model

Effective pre-training of large language models (LLMs) has been challenging due to the immense resource demands and the complexity of the technical processes involved. This paper presents a detailed technical report on YuLan-Mini, a highly capable base model with 2.42B parameters that achieves top-tier performance among models of similar parameter scale. Our pre-training approach focuses on enhancing training efficacy through three key technical contributions: an elaborate data pipeline combines data cleaning with data schedule strategies, a robust optimization method to mitigate training instability, and an effective annealing approach that incorporates targeted data selection and long context training. Remarkably, YuLan-Mini, trained on 1.08T tokens, achieves performance comparable to industry-leading models that require significantly more data. To facilitate reproduction, we release the full details of the data composition for each training phase. Project details can be accessed at the following link: https://github.com/RUC-GSAI/YuLan-Mini.

Imitate, Explore, and Self-Improve: A Reproduction Report on Slow-thinking Reasoning Systems

Recently, slow-thinking reasoning systems, such as o1, have demonstrated remarkable capabilities in solving complex reasoning tasks. These systems typically engage in an extended thinking process before responding to a query, allowing them to generate more thorough, accurate, and well-reasoned solutions. These systems are primarily developed and maintained by industry, with their core techniques not publicly disclosed. In response, an increasing number of studies from the research community aim to explore the technical foundations underlying these powerful reasoning systems. Building on these prior efforts, this paper presents a reproduction report on implementing o1-like reasoning systems. We introduce an "imitate, explore, and self-improve" framework as our primary technical approach to train the reasoning model. In the initial phase, we use distilled long-form thought data to fine-tune the reasoning model, enabling it to invoke a slow-thinking mode. The model is then encouraged to explore challenging problems by generating multiple rollouts, which can result in increasingly more high-quality trajectories that lead to correct answers. Furthermore, the model undergoes self-improvement by iteratively refining its training dataset. To verify the effectiveness of this approach, we conduct extensive experiments on three challenging benchmarks. The experimental results demonstrate that our approach achieves competitive performance compared to industry-level reasoning systems on these benchmarks.

Privacy-Preserving Federated Foundation Model for Generalist Ultrasound Artificial Intelligence

Ultrasound imaging is widely used in clinical diagnosis due to its non-invasive nature and real-time capabilities. However, conventional ultrasound diagnostics face several limitations, including high dependence on physician expertise and suboptimal image quality, which complicates interpretation and increases the likelihood of diagnostic errors. Artificial intelligence (AI) has emerged as a promising solution to enhance clinical diagnosis, particularly in detecting abnormalities across various biomedical imaging modalities. Nonetheless, current AI models for ultrasound imaging face critical challenges. First, these models often require large volumes of labeled medical data, raising concerns over patient privacy breaches. Second, most existing models are task-specific, which restricts their broader clinical utility. To overcome these challenges, we present UltraFedFM, an innovative privacy-preserving ultrasound foundation model. UltraFedFM is collaboratively pre-trained using federated learning across 16 distributed medical institutions in 9 countries, leveraging a dataset of over 1 million ultrasound images covering 19 organs and 10 ultrasound modalities. This extensive and diverse data, combined with a secure training framework, enables UltraFedFM to exhibit strong generalization and diagnostic capabilities. It achieves an average area under the receiver operating characteristic curve of 0.927 for disease diagnosis and a dice similarity coefficient of 0.878 for lesion segmentation. Notably, UltraFedFM surpasses the diagnostic accuracy of mid-level ultrasonographers and matches the performance of expert-level sonographers in the joint diagnosis of 8 common systemic diseases. These findings indicate that UltraFedFM can significantly enhance clinical diagnostics while safeguarding patient privacy, marking an advancement in AI-driven ultrasound imaging for future clinical applications.

MixPolyp: Integrating Mask, Box and Scribble Supervision for Enhanced Polyp Segmentation

Limited by the expensive labeling, polyp segmentation models are plagued by data shortages. To tackle this, we propose the mixed supervised polyp segmentation paradigm (MixPolyp). Unlike traditional models relying on a single type of annotation, MixPolyp combines diverse annotation types (mask, box, and scribble) within a single model, thereby expanding the range of available data and reducing labeling costs. To achieve this, MixPolyp introduces three novel supervision losses to handle various annotations: Subspace Projection loss (L_SP), Binary Minimum Entropy loss (L_BME), and Linear Regularization loss (L_LR). For box annotations, L_SP eliminates shape inconsistencies between the prediction and the supervision. For scribble annotations, L_BME provides supervision for unlabeled pixels through minimum entropy constraint, thereby alleviating supervision sparsity. Furthermore, L_LR provides dense supervision by enforcing consistency among the predictions, thus reducing the non-uniqueness. These losses are independent of the model structure, making them generally applicable. They are used only during training, adding no computational cost during inference. Extensive experiments on five datasets demonstrate MixPolyp's effectiveness.