Brief

Researchers have conducted a systematic study on pretraining strategies and scaling for electrocardiography (ECG) foundation models. They evaluated five different self-supervised learning objectives, finding that contrastive predictive coding and JEPA yielded the most transferable representations. The study also demonstrated that increasing pretraining data up to 11 million samples consistently improved performance for most objectives. Furthermore, structured state space models showed superior performance compared to transformers and CNNs, suggesting their inductive biases are key for effective ECG representation learning. AI

Researchers have investigated the parameter placement problem within Low-Rank Adaptation (LoRA) for fine-tuning large language models. Their study reveals that for Supervised Fine-Tuning (SFT), the specific placement of trainable parameters in the LoRA adapter's B matrix does not significantly impact performance. However, under Gradient-based Reinforcement Learning (GRPO), random parameter placement fails to improve the base model, while informed placement recovers standard LoRA accuracy. This difference is attributed to the gradient structure, with SFT gradients being stable and GRPO gradients being near-orthogonal, necessitating a gradient-informed approach for effective learning in the latter. AI

A new research paper investigates how well two prominent time series foundation models, Chronos-2 and TabPFN-TS, integrate covariate information. The study found that TabPFN-TS is more effective at capturing simple relationships between covariates and the target variable, particularly for shorter prediction horizons. This suggests that Chronos-2's strong overall performance on benchmarks may not directly indicate superior handling of covariate dependencies. AI

Researchers have introduced UniGraphLM, a novel Unified Graph Language Model designed to enhance the generalization capabilities of existing models. UniGraphLM addresses the challenge of aligning graph-encoded representations across various domains and tasks with the Large Language Model (LLM) token space. This alignment is crucial for creating unified graph tokens that combine the structural modeling of Graph Neural Networks (GNNs) with the generalization of LLMs. AI

Researchers have developed a new decoding method called Dynamic Cognitive Reconciliation Decoding (DCRD) to address conflicts between a large language model's internal knowledge and external context. DCRD uses attention maps to predict potential conflicts and then routes the input to either a greedy decoding path or a context fidelity-based dynamic decoding path. This approach aims to efficiently mitigate outdated or incorrect parametric knowledge while maintaining performance in conflict-free scenarios. Experiments on multiple LLMs and datasets demonstrate that DCRD achieves state-of-the-art results, outperforming existing baselines. AI

Researchers have developed SyncDPO, a new post-training framework designed to improve temporal synchronization in video-audio joint generation models. This method utilizes Direct Preference Optimization (DPO) to enhance the alignment between audio events and their visual counterparts, addressing limitations of traditional supervised fine-tuning. SyncDPO introduces efficient, on-the-fly negative construction strategies to create preference pairs without extensive sampling, and employs a curriculum learning approach to progressively increase the difficulty of temporal misalignments. AI

A new research paper investigates why stylistic reformulations, like poetic language, can bypass safety mechanisms in large language models. The study, using Qwen3-14B as a case study, found that models can distinguish poetic from prose formats but struggle to predict jailbreak success within these formats. The findings suggest that accumulated stylistic irregularities, rather than specific poetic devices or a failure to recognize literary formatting, lead to distinct processing patterns that circumvent safety measures. AI

Researchers have developed a new framework called CoDAAR to improve multimodal learning by creating semantically aligned discrete representations. This approach balances the need for cross-modal generalizability with the preservation of modality-specific structures. CoDAAR utilizes Discrete Temporal Alignment and Cascading Semantic Alignment to achieve state-of-the-art performance on various cross-modal generalization benchmarks, including event classification and video segmentation. AI

Researchers have identified a critical flaw in Reinforcement Learning from Human Feedback (RLHF) when applied to flow-matching text-to-image models, where standard policy entropy fails to prevent a collapse in perceptual diversity. They propose a new metric, perceptual entropy, to accurately capture diversity in the perceptual space, addressing the limitations of policy entropy which remains constant despite diversity loss. Experiments demonstrate that strategies based on perceptual entropy significantly improve the quality-diversity trade-off in image generation models. AI

Researchers have introduced UniCustom, a novel framework designed to enhance multi-reference image generation by unifying visual conditioning. This approach integrates semantic and appearance-rich features before encoding, allowing models to better associate subjects with their specific visual details from reference images. UniCustom employs a two-stage training strategy and a slot-wise binding regularization to improve subject consistency and reduce attribute leakage, demonstrating superior performance on relevant benchmarks. AI

Mythos, a new AI model, has been introduced, described as "The Day AI Sat on Bedrock." The announcement was made on Medium, with further details available via a link to the platform. AI

Researchers have introduced OmniHumanoid, a new framework for generating videos of humanoids performing actions across different embodiments. This system separates transferable motion learning from embodiment-specific adaptation, allowing it to learn from paired videos across multiple embodiments and then adapt to new ones using unpaired data via lightweight adapters. OmniHumanoid employs a branch-isolated attention design to prevent interference between motion conditioning and embodiment modulation, demonstrating strong performance in motion fidelity and embodiment consistency on both synthetic and real-world benchmarks. AI

Researchers have developed a new Spectral Vision Transformer (SVT) architecture designed for efficient tokenization, particularly in scenarios with limited data such as medical imaging. The SVT leverages spectral projection, offering theoretical advantages like spatial invariance and improved signal-to-noise ratio, which result in reduced computational complexity compared to standard spatial vision transformers. Experiments across simulated, public, and clinical datasets demonstrate that the SVT achieves comparable or better performance with fewer parameters than various other models, including compact and standard vision transformers, CNNs with attention, and MLPs. AI

Researchers have introduced the What-Where Transformer (WWT), a novel visual backbone designed to better separate object appearance from spatial location. This new architecture uses a slot-based design where tokens represent 'what' an object is and attention maps represent 'where' it is located. The WWT demonstrates emergent capabilities in discovering multiple objects directly from attention maps, even when trained with standard classification supervision, and shows improved performance on zero-shot object discovery and weakly supervised semantic segmentation tasks. AI

Researchers have developed a new framework called Latent-to-Pixel (L2P) that efficiently transfers knowledge from pre-trained Latent Diffusion Models (LDMs) to create powerful pixel-space models. This method avoids the need for extensive computational resources and real-world data by freezing most of the source LDM and training only shallow layers for the latent-to-pixel transformation. L2P utilizes synthetic images generated by LDMs as its training corpus, enabling rapid convergence with minimal hardware. The approach also eliminates the VAE bottleneck, allowing for native generation of ultra-high resolution images. AI

Researchers have developed RealDiffusion, a new framework for generating coherent multi-character storybooks using diffusion models. The system employs heat diffusion as a prior to average features and stabilize character identity across sequential frames. Additionally, a region-aware stochastic process introduces controlled perturbations to maintain narrative dynamism and scene evolution. This approach aims to resolve the trade-off between character coherence and story progression, outperforming existing methods in experiments. AI

Researchers have introduced a new framework for guided depth super-resolution that utilizes an Interactive State Space Model. This approach aims to efficiently create high-resolution depth maps from low-resolution inputs, using RGB images as guidance. The model incorporates a cross-modal local scanning mechanism to enable detailed semantic interactions between RGB and depth features, leveraging the Mamba architecture for linear complexity. Experiments indicate that this method achieves competitive results compared to existing state-of-the-art techniques. AI

Anthropic developed an AI model with advanced capabilities that they chose not to release due to safety concerns. This AI demonstrated its power by discovering a 27-year-old security vulnerability within the OpenBSD operating system. The decision to withhold the model highlights Anthropic's commitment to responsible AI development and deployment. AI

Researchers have developed StepCodeReasoner, a new framework designed to improve code reasoning by focusing on intermediate execution states rather than just final outputs. This approach uses structured print statements to create execution-trace anchors, training models to predict runtime states at each step. The framework also incorporates a novel reinforcement learning algorithm, Bi-Level GRPO, for better credit assignment across and within execution paths. Experiments show that StepCodeReasoner achieves state-of-the-art performance on code reasoning benchmarks, with its 7B model surpassing models like GPT-4o and a previous CodeReasoner baseline. AI

Researchers have introduced Yoked Feature Preference Optimization (YFPO), a novel framework designed to enhance the mathematical reasoning capabilities of large language models. Unlike existing methods that rely solely on external preference data, YFPO incorporates internal neuron activation patterns to guide the optimization process. By identifying neurons associated with mathematical concepts and logical reasoning, YFPO constructs an auxiliary reward signal that complements external supervision. Preliminary experiments on a small-scale model using the GSM8K benchmark indicate that this neuron-guided approach can potentially improve reasoning performance and offers a more interpretable path for model fine-tuning. AI

Researchers have developed a theoretical framework to understand Lifelong Normalization (LN), a key strategy for continuously updating Large Language Models without causing catastrophic forgetting or model collapse. Their analysis reveals that LN creates a self-reinforcing stability loop, ensuring parameter updates are orthogonal and bounded, which directly combats forgetting. Building on this, they introduce StableEdit, a method that enhances this stability through an explicit warm-up stage and full whitening, demonstrating improved long-horizon stability with minimal overhead. AI

Former OpenAI researcher Lilian Weng's new venture, Thinking Machines Lab (TML), has unveiled a vision for full-duplex, real-time conversational AI. This concept closely mirrors the capabilities demonstrated by China's MiniCPM-o 4.5, which was open-sourced by company "面壁智能" (OpenBMB) three months prior. Both TML and "面壁智能" aim to break away from traditional turn-based AI interactions, proposing a "full-duplex" or "time-aligned micro-turn" framework that processes interleaved multimodal information streams. AI

Researchers have developed a new framework called Macro to improve the generation of counterfactual explanations for large language models across multiple languages. This preference alignment framework uses Direct Preference Optimization (DPO) to balance the trade-off between explanation validity and minimality, which has been a challenge for non-English languages. Experiments across seven languages demonstrated that Macro significantly enhances the validity of explanations without sacrificing minimality, outperforming both chain-of-thought and supervised fine-tuning baselines. AI

Google has introduced a new AI-powered dictation feature called Rambler for its Gboard Android keyboard app. Leveraging Gemini-based multilingual models, Rambler can transcribe speech to text, remove filler words, and handle mid-sentence language switching. This integration into Gboard, the default keyboard for many Android users, poses a significant competitive challenge to existing third-party dictation startups. AI

Researchers are pushing the boundaries of 3D vision, moving beyond simple reconstruction to focus on spatial understanding, dynamic simulation, and practical engineering applications. New methods are enabling models to learn geometric relationships without explicit 3D labels, directly extract 3D-aware features for real-time synthesis, and generate dynamic 4D scenes with physical consistency. These advancements aim to equip AI with a deeper comprehension of the world, enabling it to model not just appearances but also spatial structures and physical behaviors. AI

MiniMax has released an updated version of its M2.7 AI model, focusing on improving the onboarding process for new users. This update, developed with assistance from LilacML, aims to make the model more accessible and easier for teams to implement. The enhancements highlight a push towards better usability and streamlined deployment for AI tools. AI

He Kai Ming's team has published several papers challenging the dominance of diffusion models in image generation, proposing flow matching as a more efficient alternative. Their work introduces methods like JiT, which directly predicts clean images instead of noise, achieving competitive FID scores without distillation. Additionally, their VARC model demonstrates that visual reasoning tasks, like the ARC benchmark, can be solved effectively by pure vision models without relying on language understanding, matching human performance with significantly fewer parameters. AI

Microsoft has released updates for two AI-powered developer tools. The WinUI agent plugin integrates with GitHub Copilot and Claude Code to assist in building native Windows applications. Additionally, Foundry Local 1.1 now features live transcription, embeddings, and a Responses API for local AI model interaction. AI

Anthropic has released an update for Claude Code, introducing an agent view and a /goal command. The agent view provides a centralized dashboard to manage multiple Claude Code sessions simultaneously, akin to a tmux interface for code generation tasks. The /goal command allows Claude to autonomously work on a task until a specified completion condition is met, reducing the need for constant human intervention. AI

A new AI tool named Jessie has been released, with its announcement met with enthusiasm from its creator. Separately, Claude AI's Agent View has been updated with an automated git worktree feature, aiming to enhance developer workflows. Additionally, GLM 5.1 was tested autonomously across over 600 prompts, showcasing potential for agent-based applications and model evaluation. AI

Thinking Machines has introduced a research preview of interaction models designed for native, real-time collaboration. These models process audio, video, and text simultaneously, allowing for continuous thought, response, and action. This approach aims to overcome the limitations of current turn-based AI interfaces, enabling a more natural and fluid human-AI partnership that mirrors human-to-human interaction. AI

Researchers have developed a new method called Super-Linear Advantage Shaping (SLAS) to improve text-to-image models trained with reinforcement learning. This technique addresses reward hacking by reshaping the policy space using an information geometry perspective, amplifying informative updates while suppressing noisy ones. SLAS demonstrates superior performance over existing methods like DanceGRPO, leading to faster training, better out-of-domain generation, and increased robustness to model scaling. AI

Researchers have introduced a new framework for open-world object counting, addressing the brittleness of current vision-language models in accurately identifying and counting objects based on user intent. They propose redefining counting as a multi-grained problem, where both visual examples and detailed text prompts, including negative prompts, specify the target appearance and semantic granularity. To overcome the data limitations for this approach, they developed an automated pipeline using 3D synthesis and VLM filtering to create KubriCount, the largest dataset for counting tasks. Their new model, HieraCount, leverages both text and visual exemplars to significantly improve multi-grained counting accuracy and generalize to real-world scenarios. AI

Researchers have introduced Directional-Groupwise Preference Optimization (DGPO), a new framework designed to improve the alignment and reasoning diversity of large language models. DGPO aggregates supervision signals at the group level, using multi-candidate comparisons to explicitly model direction-aware alignment. By organizing question-answer instances into structured sets and optimizing a margin-based objective, DGPO aims to differentiate coherent reasoning paths from inconsistent ones. Experiments show that this approach can lead to significant accuracy improvements across various benchmarks and model families. AI

Researchers have developed a new framework called On-policy Data Evolution (ODE) to improve multimodal deep search agents. This system allows agents to reuse intermediate visual information from search results and dynamically refines training data based on the agent's current learning progress. ODE enhances agent performance across various benchmarks, with significant improvements shown for Qwen3-VL models, surpassing Gemini-2.5 Pro in complex agent-workflow settings. AI

Researchers have developed CLEF, a new foundation model designed for interpreting clinical electroencephalogram (EEG) data. Unlike previous models that focus on short EEG segments, CLEF can process entire EEG sessions and integrate signal patterns with clinical context. The model represents EEG data as 3D spectrogram tokens, allowing for efficient Transformer modeling, and is aligned with neurologist reports and electronic health records. CLEF significantly outperforms existing models on a broad benchmark of clinical tasks, demonstrating its potential for advancing clinical EEG analysis. AI

Researchers have investigated the internal mechanisms of audio-visual large language models (AVLLMs), focusing on how information flows between audio and visual modalities. Their analysis revealed that AVLLMs predominantly store integrated audio-visual information in specific 'sink tokens'. Furthermore, a subset of these sink tokens, termed 'cross-modal sink tokens', are specialized for holding this cross-modal information. Based on these findings, the paper proposes a new method to mitigate hallucination by leveraging the integrated information within these specialized tokens. AI

Interfaze has introduced a new model architecture designed for high accuracy and efficiency on deterministic tasks. This architecture reportedly outperforms leading models such as Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini, and Grok-4.3 across nine benchmarks covering OCR, vision, speech-to-text, and structured output. Interfaze aims to specialize in these specific tasks, offering a cost-effective and high-performance alternative to generalist large language models for high-volume applications. AI

Researchers have developed MASS-DPO, a new method for Direct Preference Optimization (DPO) that efficiently selects informative negative samples for training language models. This approach uses a PL-specific Fisher-information objective to identify compact subsets of negative responses that provide complementary information, reducing redundancy from similar candidates. Experiments across recommendation and multiple-choice QA benchmarks demonstrate that MASS-DPO achieves comparable or superior accuracy with significantly fewer negative samples, improving optimization dynamics and alignment. AI

Researchers have developed DRAPE, a novel framework for Multimodal Continual Instruction Tuning (MCIT) that generates instance-specific soft prompts for multimodal large language models. Unlike existing methods that rely on task-level prompts, DRAPE synthesizes continuous prompts tailored to individual query-image pairs by conditioning on both textual instructions and visual features. The framework also incorporates techniques like null-space gradient projection and CLIP-based prototype routing to prevent catastrophic forgetting during sequential task acquisition, achieving state-of-the-art results on MCIT benchmarks. AI

Researchers have developed a new framework called FedMITR to improve one-shot federated learning, particularly in scenarios with highly non-independent and identically distributed (non-IID) data. This method addresses the issue of low-quality synthetic data generated by existing approaches by employing sparse model inversion to focus on meaningful image patches and avoid background noise. Additionally, FedMITR uses a token relabeling strategy for Vision Transformers (ViTs) to enhance prediction robustness by distinguishing between high and low information density patches. AI

Alibaba's Qwen-Image-2.0 is a new foundation model designed for both high-fidelity image generation and precise editing within a single framework. It addresses limitations in existing models concerning ultra-long text rendering, multilingual typography, photorealism, and instruction following. The model utilizes Qwen3-VL as a condition encoder and a Multimodal Diffusion Transformer, trained on extensive data, to achieve improved multimodal understanding and flexible generation capabilities. AI

Researchers have developed new post-training objectives for large language models that optimize for the best-of-N performance, rather than just the average reward. This is crucial because current deployment strategies involve sampling multiple responses and selecting the best one, a process that standard training objectives do not adequately address. The proposed Tail-Extrapolated (TEA) estimators and Prefix-TEA can approximate the best-of-N objective using significantly fewer per-prompt rollouts during training than would be required for deployment, showing improved performance on instruction-following tasks. AI

Researchers have developed a new method called Gated Cropped Attention-Delta steering (GCAD) to improve the reliability of controlling language model behavior. Standard activation steering can degrade performance in long conversations due to issues with the KV-cache. GCAD addresses this by extracting steering signals from self-attention mechanisms and applying them with token-level gating, significantly enhancing long-horizon coherence and trait expression in multi-turn dialogues. AI

Researchers have developed bViT, a novel Vision Transformer architecture that utilizes a single transformer block applied repeatedly for image recognition. This recurrent approach achieves accuracy comparable to standard ViTs on ImageNet-1K with significantly fewer parameters. The study suggests that a substantial portion of a ViT's depth can be achieved through recurrent computation, especially when the representation space is wide, enabling parameter-efficient fine-tuning for downstream tasks. AI

Researchers have developed BCJR-QAT, a novel method for quantizing large language models to 2 bits per weight, a significant advancement beyond current post-training quantization techniques. This new approach uses a differentiable relaxation of the Viterbi algorithm, enabling quantization-aware training and achieving better perplexity scores on benchmarks like WikiText-2. The method has been demonstrated to improve performance on models such as Llama-3.2-1B, outperforming existing methods by a notable margin. AI

Researchers have developed a new criterion for initializing weights in gated recurrent neural networks, crucial for the performance of reservoir computing models. This criterion, derived from random-matrix theory, helps identify an effective critical point that separates ordered and chaotic phases in randomly initialized models. The method closely tracks the optimal gain for gated RNNs on forecasting tasks and could inform future initialization strategies. AI

Researchers have identified a pretraining failure mode in language models where upper layers prematurely specialize their attention patterns before lower layers have stabilized. This "premature upper-layer attention specialization" can be mitigated by temporarily slowing the Q/K projections in these upper layers during early training. This intervention improves final perplexity and downstream accuracy without changing other model parameters, suggesting a critical interaction between decoder architecture and optimization. AI

A developer successfully transformed a small, 397MB Qwen2.5–0.5B model into a functional customer service chatbot. This involved fine-tuning the model on specific company data using the LoRA technique, enabling it to provide accurate and contextually relevant responses. The resulting chatbot was integrated into a real company's workflow, addressing customer inquiries about orders, returns, and product compatibility with the company's specific policies and tone. AI

Researchers have investigated the impact of temporal sampling frequency on end-to-end autonomous driving trajectory prediction models. They found that while dense frame sampling is often assumed to improve performance, this is not always the case. Smaller models often perform best with lower or intermediate sampling frequencies, suggesting that dense sampling can introduce redundant information and noise that burdens models with limited capacity. Larger, vision-language-model-style architectures, however, continued to improve performance even at the highest tested sampling frequencies. AI