Brief

Researchers have developed TractFM, a novel foundation model designed to learn representations directly from diffusion MRI tractograms. This model uniquely combines a local streamline encoder with a permutation-equivariant tractogram encoder, enabling it to process all streamlines from a subject simultaneously. By pretraining on anatomical parcellation, TractFM generates reusable embeddings for both individual streamlines and compact subject-level descriptors. The model demonstrates strong generalization capabilities, achieving accurate tract parcellation and predicting subject phenotypes like age and sex across different tractography algorithms and datasets. AI

A critical vulnerability in Zcash was discovered using Anthropic's Claude Opus model, highlighting AI's ability to find human-imperceptible flaws. This discovery led to a significant 41% drop in Zcash's value and signals a new era for automated cybersecurity. AI

A new AI model named Claude Mythos has reportedly surpassed existing benchmarks, signaling a significant advancement in AI capabilities. This development is presented as particularly relevant for small business owners considering AI adoption. The implications of this benchmark breakthrough are being analyzed for their real-world impact. AI

OpenAI has released a new tool called MCP that allows users to integrate GPT-4o, DALL-E 3, and Whisper directly into AI clients like Claude and Cursor. This integration enables AI agents to call upon OpenAI's various models for specific tasks, such as image generation or audio transcription, without leaving their primary workflow. The setup is designed to be quick, allowing for seamless multi-model orchestration and enhanced capabilities within a single AI environment. AI

Researchers have introduced TD-Grokking, a novel framework designed to enable large language models to learn from zero-reward problems. This method recursively breaks down complex, intractable problems into smaller, verifiable subproblems. These subproblems form a hierarchy, with solvable leaves providing the necessary optimization signals for model improvement. Evaluations on mathematical and medical tasks demonstrated that TD-Grokking significantly outperforms existing baseline approaches. AI

Researchers have developed a new framework called FAMOU to improve LLM-driven strategy evolution in adversarial games. This framework addresses the challenge of shifting evaluation landscapes by incorporating co-evolutionary mechanisms, hierarchical deep evaluation, and dynamic weakness pressure. Tested on the MCTF 2026 3v3 maritime capture-the-flag task, FAMOU demonstrated superior performance over existing methods, achieving the highest combined score and best generalization to unseen opponents. The evolved strategies also showcased novel algorithmic innovations, validating the approach's effectiveness and real-world transferability. AI

Researchers have investigated the impact of different normalization techniques on causal large time-series models, particularly those using transformer architectures with patching and efficient causal strategies. Their findings indicate that the choice of normalization significantly affects both the speed of training convergence and the accuracy of forecasting performance. The study highlights potential information leakage issues with standard normalization in causal settings and evaluates newer alternatives designed to mitigate this problem. AI

Researchers have developed specialized large language models for additive manufacturing by adapting open-weight models like Gemma 3, Qwen 3, and Gemma 4. These models were trained on approximately 50 million tokens of additive manufacturing journal articles, incorporating both text and visual data. Evaluations using the Additive-Manufacturing-Benchmark show these domain-adapted models achieve over 90% accuracy on additive manufacturing knowledge tasks, demonstrating an effective method for LLM specialization. AI

Researchers have developed a new formula to calculate an upper bound for local learning coefficients in three-layer neural networks. This formula addresses singular points, which were a limitation in previous methods. The new approach offers a counting rule based on budget, demand, and supply constraints and extends to a broader range of activation functions, including swish and polynomial types under specific conditions. AI

Researchers have developed a new deep learning model called the Deep Convolutional Interpreter for Time Series (DCIts). This architecture is designed to analyze nonlinear multivariate time series data and provides sample-specific, locally interpretable descriptions of interaction structures. DCIts achieves competitive forecasting accuracy while prioritizing intrinsic interpretability by explicitly learning a time- and lag-dependent transition tensor. AI

Researchers have developed PRISM, a novel sequence modeling architecture designed to balance the expressivity of Transformers with the efficiency of linear models. PRISM addresses the serial dependencies found in iterative methods like Test-Time Training by reconstructing the iterative process in a parallelizable form. This is achieved through a Write-Forget Decoupling strategy and a two-stage proxy architecture, enabling significantly higher throughput compared to existing optimization methods. AI

Researchers have introduced LongMoE, a novel framework designed to tackle the complexities of multimodal clinical learning. This approach effectively addresses two key challenges: missing data across different patient modalities and the temporal dynamics of disease progression. By integrating context-aware imputation with trajectory-aware encoding and a sparse Mixture-of-Experts system, LongMoE can model disease evolution over time even with incomplete or inconsistent patient data. AI

Researchers have developed WorldPlay, a novel streaming video diffusion model designed for real-time interactive world modeling. This model addresses the speed-memory trade-off in current systems by employing a Dual Action Representation for robust input control and a Reconstituted Context Memory with temporal reframing to maintain long-term geometric consistency. Additionally, Context Forcing, a distillation method, ensures the model can effectively utilize long-range information, enabling real-time 720p video generation at 24 FPS with improved consistency and generalization. AI

Researchers have developed a new approach called FoA-SR for image super-resolution that can generate distinct restoration profiles. This method allows for either faithful reconstructions that prioritize structural integrity and reference consistency, or aesthetic reconstructions that focus on visually pleasing details. The system uses a supervised SR adapter trained with various losses, then fine-tunes separate LoRA adapters using profile-specific rewards to achieve these different objectives. AI

Researchers have developed a new sampling method called Entropy-Guided Power Sampling (EGPS) to improve the reasoning capabilities of base language models. This method addresses the inefficiencies of traditional Metropolis-Hastings samplers by focusing on high-entropy regions within sequences, leading to faster and more effective sampling. EGPS demonstrated strong performance on benchmarks like MATH500, HumanEval, and GPQA, achieving significant speedups over existing techniques. AI

A new research paper explores the critical role of synthetic data composition in pretraining time series foundation models. The study found that the choice of synthetic data generator can lead to a twofold difference in forecasting error, and these generator rankings are not consistent across different model architectures. Researchers propose that mixing multiple generators with real data creates the strongest pretraining corpora, framing the problem as one of corpus composition rather than generator selection. AI

Researchers have developed a new forward-only learning algorithm for convolutional neural networks (CNNs) that improves upon existing methods. This approach introduces a learnable mechanism for assigning channels to classes, allowing for more adaptive and data-driven specialization. Additionally, a loss-aware layer contribution strategy weights intermediate predictions based on their validation performance, enhancing inference. When integrated into residual CNNs, this method achieves state-of-the-art performance among forward-only models on several image datasets, significantly closing the gap with traditional backpropagation techniques. AI

Researchers have developed a new framework for post-training augmentation invariance, allowing pretrained neural networks to gain new invariance properties without affecting their performance on original data. This method uses lightweight adapter networks appended to the latent space, trained with novel Markov-Wasserstein minimization or Wasserstein correlation maximization losses. Empirical results show significant improvements in classification accuracy for rotated and noisy images, with minimal corruption to the original features and no fine-tuning of the base network. AI

Researchers have developed a new deep learning model called the Time-Aware Small-Attention U-Net (TA-SmaAt-UNet) to improve precipitation nowcasting, particularly for high-intensity rainfall events. This model incorporates lightweight temporal conditioning layers that use cyclical encodings of time-of-day and time-of-year to enhance feature representations. Experiments demonstrated that this temporal context is most beneficial for rare, intense rainfall, while also improving the representation of seasonal variability and rainfall intensity distributions. AI

Researchers have introduced ELF-S2T, a novel approach to speech-to-text systems that operates in a continuous latent space rather than discrete text tokens. This model, built on the Embedded Language Flows (ELF) backbone, uses audio conditioning and flow-matching denoising for both speech recognition and translation tasks. Experiments on standard datasets demonstrate competitive performance and reveal that errors in both recognition and translation stem from similar confusions within this continuous latent space. AI

Researchers have developed MoE-FedTP, a new framework for spatiotemporal prediction that uses a Mixture-of-Experts (MoE) approach within a federated learning system. This method aims to improve traffic prediction accuracy, especially in cities with limited data, by enabling knowledge transfer from data-rich cities without compromising privacy. Experiments show MoE-FedTP outperforms existing cross-city and federated learning techniques. AI

Researchers have developed Whisper-GPT, a novel language model designed for generating speech and music. This model uniquely integrates continuous audio representations, like spectrograms, with discrete tokens derived from neural compression algorithms. This hybrid approach aims to overcome the context length limitations often encountered with purely discrete token models, while retaining the predictive benefits of discrete spaces for tasks like sampling. AI

Researchers have developed Whisfusion, a novel non-autoregressive system for automatic speech recognition (ASR) that utilizes masked diffusion models. This approach aims to match the accuracy of traditional autoregressive models while significantly improving inference speed. Whisfusion achieves this by training a diffusion decoder on top of frozen Whisper-large-v3 audio embeddings, enabling parallel decoding and outperforming existing models in both speed and accuracy across multiple languages. AI

Researchers have developed a new method called Streaming Knowledge Compilation to update LLM wikis with evolving information. This technique uses a "materiality signal" to proactively pin important documents within a fixed token budget, aiming to minimize regret against future queries. The system was tested on financial news and Wikipedia, demonstrating its ability to adapt to new information and providing a more reliable evaluation metric than standard QA scores for post-training knowledge. AI

Researchers have identified a "prompt forgetting" issue in Multimodal Diffusion Transformers (MMDiTs) used for text-to-image generation. This phenomenon occurs because the text prompt's semantic representation degrades as it passes through deeper layers of the model. To address this, a new training-free method called "prompt reinjection" has been proposed, which reintroduces early-layer prompt representations into later layers. Experiments on models like SD3, SD3.5, and FLUX.1 demonstrate that this technique improves instruction-following capabilities and overall generation quality. AI

Researchers have developed SARA, a new method for improving video diffusion models by focusing supervision on semantically relevant parts of the video. This approach uses text-conditioned saliency to determine which token pairs in the video generation process are most important for aligning with the prompt. SARA demonstrates improved text alignment and motion quality compared to existing methods in evaluations. AI

Researchers have developed a new method called S3-GFN for generating molecules that are both synthesizable and possess desirable properties. This approach uses a sequence-based Generative Flow Network (GFlowNet) with soft regularization, incorporating rich molecular priors learned from large datasets. By employing contrastive learning with separate buffers of synthesizable and unsynthesizable molecules, S3-GFN effectively guides the generation process towards high-reward chemical spaces, achieving over 95% synthesizability in experiments. AI

Researchers have developed a new framework called SHAPE for pruning experts in sparse Mixture-of-Experts (MoE) large language models. Unlike previous methods that evaluated experts independently, SHAPE considers the cooperative nature of MoE inference, where experts work in coalitions. The framework uses a Shapley-style attribution to identify experts crucial for high-utility collaborations, leading to more effective pruning. Experiments on models like Qwen3-30B-A3B, GPT-OSS-20B, and DeepSeek-V2-Lite demonstrated that SHAPE can significantly reduce memory footprint without substantial accuracy loss, even with up to 40% expert pruning. AI

Researchers have developed a new framework for improving the accuracy of reduced-order models (ROMs) used in complex multiscale systems. This uncertainty-aware approach utilizes conditional normalizing flows to learn a probabilistic mapping between low-fidelity and high-fidelity model coefficients. The method aims to enhance predictive accuracy while also quantifying the uncertainty in the learned closure, which is crucial for reliable application of ROMs. Experiments on a vortex merging problem demonstrated that this technique significantly improves ROM accuracy over uncorrected models. AI

Researchers have developed a new method called SPGCL to improve Graph Contrastive Learning (GCL). They found that existing GCL methods often fail to effectively learn from positive samples due to the message-passing mechanism in graph encoders. SPGCL aims to fix this by selectively propagating high-energy features and using low-energy features for more reliable positive sampling, leading to better performance in experiments. AI

Researchers have developed a new two-stage framework for time-series forecasting that aims to improve accuracy by explicitly modeling and correcting systematic residual biases. The approach uses a base transformer model for initial predictions, followed by a dedicated meta-corrector that learns to refine these predictions. This method has demonstrated state-of-the-art performance on eight benchmark datasets, showing significant improvements in standard metrics like MSE and MAE. AI

Researchers have developed ANCHOR, a novel autoregressive model designed for incremental speech quality assessment. Unlike previous methods that require complete utterances, ANCHOR can estimate quality from partial audio streams, making it suitable for real-time applications. The model employs a dual-resolution token system and a hierarchical structure to refine quality predictions from coarse to fine, demonstrating a significant reduction in error on short audio prefixes. AI

Researchers have developed a novel framework called the Biological-Informed Recurrent Neural Network (BIRNN) to improve the modeling of glucose-insulin dynamics for Type 1 Diabetes management. This approach integrates a Gated Recurrent Units (GRU) architecture with physics-informed loss functions that embed physiological constraints. The BIRNN framework demonstrated superior glucose prediction accuracy compared to traditional linear models, even accounting for circadian variations in insulin sensitivity, according to validation using the UVA/Padova simulator. AI

Researchers have identified a phenomenon called "model plasticity loss" that hinders the effectiveness of Reinforcement Learning (RL) after Supervised Fine-Tuning (SFT) for large language models. Excessive SFT can lead to over-confident token distributions and difficult optimization landscapes, limiting RL's ability to further enhance model capabilities. To address this, a new method called "Rejuvenation" has been proposed, which uses base-anchored model fusion and targeted neuron resets to restore plasticity while retaining SFT benefits, showing improved performance on reasoning and agentic tasks. AI

Researchers have developed VFUSE, a new method using sparse autoencoders to interpret generative models for protein design. This approach audits models like RoseTTAFold3 and RFDiffusion3 for potentially hazardous features. VFUSE's analysis in the latent space of these models improved the detection of dangerous protein designs, identifying specific features that activate only for hazardous outputs with high accuracy. AI

Researchers have introduced Conditional-Vendi and Conditional-RKE, new metrics designed to evaluate the diversity of outputs from generative AI models, specifically when guided by text prompts. These methods build upon existing diversity measures by isolating variability that originates from the model itself, rather than just the prompts. The new scores have demonstrated effectiveness in tasks involving text-to-image generation, image captioning, and large language models, showing they can accurately reflect ground-truth diversity and even guide models to produce more varied outputs. AI

Researchers have developed CodeAlchemy, a framework for generating large-scale synthetic code data to improve AI model training. The system employs five strategies, including code rewriting, question answering, developer tasks, conversational dialogues, and execution traces, producing over 500 billion tokens of synthetic code and 350 billion reasoning tokens. This extensive dataset aims to address the limitations of current models in understanding real-world code tasks, with new benchmarks like DevEval and TraceEval highlighting significant gaps in semantic comprehension among even frontier models. AI

Researchers have developed Density Field State Space Models (DF-SSM), a novel framework for compressing large SSMs into a 1-bit scaffold with minimal performance loss. Applied to Mamba-2 1.3B, this method resulted in a model that is over nine times smaller and significantly faster for inference, while retaining performance close to a 1.58-bit model. The distillation process is remarkably efficient, requiring limited data and computational resources. Beyond compression, the study also analyzed the model's internal knowledge organization, revealing distinct phases for intent classification, knowledge retrieval, and output formatting, suggesting that representational structure can develop independently of strong factual recall. AI