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.

Spatiotemporal-Untrammelled Mixture of Experts for Multi-Person Motion Prediction

Comprehensively and flexibly capturing the complex spatio-temporal dependencies of human motion is critical for multi-person motion prediction. Existing methods grapple with two primary limitations: i) Inflexible spatiotemporal representation due to reliance on positional encodings for capturing spatiotemporal information. ii) High computational costs stemming from the quadratic time complexity of conventional attention mechanisms. To overcome these limitations, we propose the Spatiotemporal-Untrammelled Mixture of Experts (ST-MoE), which flexibly explores complex spatio-temporal dependencies in human motion and significantly reduces computational cost. To adaptively mine complex spatio-temporal patterns from human motion, our model incorporates four distinct types of spatiotemporal experts, each specializing in capturing different spatial or temporal dependencies. To reduce the potential computational overhead while integrating multiple experts, we introduce bidirectional spatiotemporal Mamba as experts, each sharing bidirectional temporal and spatial Mamba in distinct combinations to achieve model efficiency and parameter economy. Extensive experiments on four multi-person benchmark datasets demonstrate that our approach not only outperforms state-of-art in accuracy but also reduces model parameter by 41.38% and achieves a 3.6x speedup in training. The code is available at https://github.com/alanyz106/ST-MoE.

From Retrieval to Reasoning: A Framework for Cyber Threat Intelligence NER with Explicit and Adaptive Instructions

The automation of Cyber Threat Intelligence (CTI) relies heavily on Named Entity Recognition (NER) to extract critical entities from unstructured text. Currently, Large Language Models (LLMs) primarily address this task through retrieval-based In-Context Learning (ICL). This paper analyzes this mainstream paradigm, revealing a fundamental flaw: its success stems not from global semantic similarity but largely from the incidental overlap of entity types within retrieved examples. This exposes the limitations of relying on unreliable implicit induction. To address this, we propose TTPrompt, a framework shifting from implicit induction to explicit instruction. TTPrompt maps the core concepts of CTI's Tactics, Techniques, and Procedures (TTPs) into an instruction hierarchy: formulating task definitions as Tactics, guiding strategies as Techniques, and annotation guidelines as Procedures. Furthermore, to handle the adaptability challenge of static guidelines, we introduce Feedback-driven Instruction Refinement (FIR). FIR enables LLMs to self-refine guidelines by learning from errors on minimal labeled data, adapting to distinct annotation dialects. Experiments on five CTI NER benchmarks demonstrate that TTPrompt consistently surpasses retrieval-based baselines. Notably, with refinement on just 1% of training data, it rivals models fine-tuned on the full dataset. For instance, on LADDER, its Micro F1 of 71.96% approaches the fine-tuned baseline, and on the complex CTINexus, its Macro F1 exceeds the fine-tuned ACLM model by 10.91%.

MPD-SGR: Robust Spiking Neural Networks with Membrane Potential Distribution-Driven Surrogate Gradient Regularization

The surrogate gradient (SG) method has shown significant promise in enhancing the performance of deep spiking neural networks (SNNs), but it also introduces vulnerabilities to adversarial attacks. Although spike coding strategies and neural dynamics parameters have been extensively studied for their impact on robustness, the critical role of gradient magnitude, which reflects the model's sensitivity to input perturbations, remains underexplored. In SNNs, the gradient magnitude is primarily determined by the interaction between the membrane potential distribution (MPD) and the SG function. In this study, we investigate the relationship between the MPD and SG and their implications for improving the robustness of SNNs. Our theoretical analysis reveals that reducing the proportion of membrane potentials lying within the gradient-available range of the SG function effectively mitigates the sensitivity of SNNs to input perturbations. Building upon this insight, we propose a novel MPD-driven surrogate gradient regularization (MPD-SGR) method, which enhances robustness by explicitly regularizing the MPD based on its interaction with the SG function. Extensive experiments across multiple image classification benchmarks and diverse network architectures confirm that the MPD-SGR method significantly enhances the resilience of SNNs to adversarial perturbations and exhibits strong generalizability across diverse network configurations, SG functions, and spike encoding schemes.

DS-ATGO: Dual-Stage Synergistic Learning via Forward Adaptive Threshold and Backward Gradient Optimization for Spiking Neural Networks

Brain-inspired spiking neural networks (SNNs) are recognized as a promising avenue for achieving efficient, low-energy neuromorphic computing. Direct training of SNNs typically relies on surrogate gradient (SG) learning to estimate derivatives of non-differentiable spiking activity. However, during training, the distribution of neuronal membrane potentials varies across timesteps and progressively deviates toward both sides of the firing threshold. When the firing threshold and SG remain fixed, this may lead to imbalanced spike firing and diminished gradient signals, preventing SNNs from performing well. To address these issues, we propose a novel dual-stage synergistic learning algorithm that achieves forward adaptive thresholding and backward dynamic SG. In forward propagation, we adaptively adjust thresholds based on the distribution of membrane potential dynamics (MPD) at each timestep, which enriches neuronal diversity and effectively balances firing rates across timesteps and layers. In backward propagation, drawing from the underlying association between MPD, threshold, and SG, we dynamically optimize SG to enhance gradient estimation through spatio-temporal alignment, effectively mitigating gradient information loss. Experimental results demonstrate that our method achieves significant performance improvements. Moreover, it allows neurons to fire stable proportions of spikes at each timestep and increases the proportion of neurons that obtain gradients in deeper layers.

Making Every Head Count: Sparse Attention Without the Speed-Performance Trade-off

The design of Large Language Models (LLMs) has long been hampered by a fundamental conflict within their core attention mechanism: its remarkable expressivity is built upon a computational complexity of $O(H \cdot N^2)$ that grows quadratically with the context size ($N$) and linearly with the number of heads ($H$). This standard implementation harbors significant computational redundancy, as all heads independently compute attention over the same sequence space. Existing sparse methods, meanwhile, often trade information integrity for computational efficiency. To resolve this efficiency-performance trade-off, we propose SPAttention, whose core contribution is the introduction of a new paradigm we term Principled Structural Sparsity. SPAttention does not merely drop connections but instead reorganizes the computational task by partitioning the total attention workload into balanced, non-overlapping distance bands, assigning each head a unique segment. This approach transforms the multi-head attention mechanism from $H$ independent $O(N^2)$ computations into a single, collaborative $O(N^2)$ computation, fundamentally reducing complexity by a factor of $H$. The structured inductive bias compels functional specialization among heads, enabling a more efficient allocation of computational resources from redundant modeling to distinct dependencies across the entire sequence span. Extensive empirical validation on the OLMoE-1B-7B and 0.25B-1.75B model series demonstrates that while delivering an approximately two-fold increase in training throughput, its performance is on par with standard dense attention, even surpassing it on select key metrics, while consistently outperforming representative sparse attention methods including Longformer, Reformer, and BigBird across all evaluation metrics.

FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation

While LLMs have shown great success in financial tasks like stock prediction and question answering, their application in fully automating Equity Research Report generation remains uncharted territory. In this paper, we formulate the Equity Research Report (ERR) Generation task for the first time. To address the data scarcity and the evaluation metrics absence, we present an open-source evaluation benchmark for ERR generation - FinRpt. We frame a Dataset Construction Pipeline that integrates 7 financial data types and produces a high-quality ERR dataset automatically, which could be used for model training and evaluation. We also introduce a comprehensive evaluation system including 11 metrics to assess the generated ERRs. Moreover, we propose a multi-agent framework specifically tailored to address this task, named FinRpt-Gen, and train several LLM-based agents on the proposed datasets using Supervised Fine-Tuning and Reinforcement Learning. Experimental results indicate the data quality and metrics effectiveness of the benchmark FinRpt and the strong performance of FinRpt-Gen, showcasing their potential to drive innovation in the ERR generation field. All code and datasets are publicly available.

Cross-domain EEG-based Emotion Recognition with Contrastive Learning

Electroencephalogram (EEG)-based emotion recognition is vital for affective computing but faces challenges in feature utilization and cross-domain generalization. This work introduces EmotionCLIP, which reformulates recognition as an EEG-text matching task within the CLIP framework. A tailored backbone, SST-LegoViT, captures spatial, spectral, and temporal features using multi-scale convolution and Transformer modules. Experiments on SEED and SEED-IV datasets show superior cross-subject accuracies of 88.69% and 73.50%, and cross-time accuracies of 88.46% and 77.54%, outperforming existing models. Results demonstrate the effectiveness of multimodal contrastive learning for robust EEG emotion recognition.

Fusion of Heterogeneous Pathology Foundation Models for Whole Slide Image Analysis

Whole slide image (WSI) analysis has emerged as an increasingly essential technique in computational pathology. Recent advances in the pathological foundation models (FMs) have demonstrated significant advantages in deriving meaningful patch-level or slide-level feature representations from WSIs. However, current pathological FMs have exhibited substantial heterogeneity caused by diverse private training datasets and different network architectures. This heterogeneity introduces performance variability when we utilize the extracted features from different FMs in the downstream tasks. To fully explore the advantage of multiple FMs effectively, in this work, we propose a novel framework for the fusion of heterogeneous pathological FMs, called FuseCPath, yielding a model with a superior ensemble performance. The main contributions of our framework can be summarized as follows: (i) To guarantee the representativeness of the training patches, we propose a multi-view clustering-based method to filter out the discriminative patches via multiple FMs' embeddings. (ii) To effectively fuse the heterogeneous patch-level FMs, we devise a cluster-level re-embedding strategy to online capture patch-level local features. (iii) To effectively fuse the heterogeneous slide-level FMs, we devise a collaborative distillation strategy to explore the connections between slide-level FMs. Extensive experiments conducted on lung cancer, bladder cancer, and colorectal cancer datasets from The Cancer Genome Atlas (TCGA) have demonstrated that the proposed FuseCPath achieves state-of-the-art performance across multiple tasks on these public datasets.

See the Text: From Tokenization to Visual Reading

People see text. Humans read by recognizing words as visual objects, including their shapes, layouts, and patterns, before connecting them to meaning, which enables us to handle typos, distorted fonts, and various scripts effectively. Modern large language models (LLMs), however, rely on subword tokenization, fragmenting text into pieces from a fixed vocabulary. While effective for high-resource languages, this approach over-segments low-resource languages, yielding long, linguistically meaningless sequences and inflating computation. In this work, we challenge this entrenched paradigm and move toward a vision-centric alternative. Our method, SeeTok, renders text as images (visual-text) and leverages pretrained multimodal LLMs to interpret them, reusing strong OCR and text-vision alignment abilities learned from large-scale multimodal training. Across three different language tasks, SeeTok matches or surpasses subword tokenizers while requiring 4.43 times fewer tokens and reducing FLOPs by 70.5%, with additional gains in cross-lingual generalization, robustness to typographic noise, and linguistic hierarchy. SeeTok signals a shift from symbolic tokenization to human-like visual reading, and takes a step toward more natural and cognitively inspired language models.