Brief

Researchers have developed DeepSight, a novel world model for end-to-end autonomous driving systems that enhances decision-making by predicting future states in the bird's-eye-view (BEV) space. This model integrates Vision-Language Model (VLM) architectures with a specialized visual reasoning module designed for driving scenarios. DeepSight also incorporates an adaptive text reasoning mechanism that leverages social knowledge to improve performance in challenging long-tail situations, achieving state-of-the-art results on the Bench2drive benchmark. AI

Researchers from ICT-NLP have developed a novel system for dimensional aspect sentiment regression, achieving top rankings in the SemEval-2026 Task 3. Their approach utilizes a multilingual encoder with joint training and an adaptive ensemble, eschewing large language models for efficiency. This method demonstrated strong cross-lingual transfer capabilities and improved training stability, leading to high performance across multiple datasets. AI

Researchers have introduced MMM-Bench, a new benchmark designed to address the limitations of existing document classification systems. This benchmark features a five-level hierarchical taxonomy and a dataset of 5,990 real-world multi-modal documents from 12 commercial domains within Alibaba. MMM-Bench aims to better reflect the complexity of practical document intelligence by incorporating multi-level, multi-domain, and multi-modal aspects, and the team has released the data and evaluation toolkit to facilitate further research. AI

Researchers have developed a new supervision objective called EXACT to improve long-context adaptation in language models. This method addresses a mismatch in packed training by assigning extra weight to targets that rely on longer effective contexts. Experiments on Qwen and LLaMA models demonstrated significant improvements in benchmarks like NoLiMa and RULER, particularly when evidence was located thousands of tokens away, while preserving performance on standard QA and reasoning tasks. AI

Researchers have introduced Time Interval Encoding (TIE), a novel method to enhance video generation models like Diffusion Transformers (DiT). TIE addresses the limitation of current models that treat time as discrete points, making it difficult to represent overlapping events and extended durations. By generalizing rotary embeddings, TIE allows models to process time intervals as first-class primitives, improving temporal controllability and accuracy in video generation tasks. AI

Researchers have introduced Mela, a novel memory-augmented language model that draws inspiration from neuroscientific theories of memory consolidation. Mela utilizes a Hierarchical Memory Module (HMM) with distinct sub-modules operating at different frequencies to capture both abstract and detailed information. This architecture allows Mela to perform online memory consolidation during inference, enabling it to handle significantly longer contexts than standard Transformer models without performance degradation. AI

Researchers have developed GemDepth, a new framework designed to improve 3D-consistent video depth estimation. Unlike previous methods that often blur fine details or exhibit temporal inconsistencies, GemDepth explicitly incorporates camera motion and global 3D structure. Its Geometry-Embedding Module predicts inter-frame camera poses to create implicit geometric embeddings, enhancing the model's 3D perception and alignment capabilities. This approach allows for more precise spatial details and rigorous temporal consistency, achieving state-of-the-art results on various datasets. AI

Researchers have introduced DuetFair, a novel mechanism designed to enhance fairness in medical image segmentation models. This framework addresses the issue of "intra-group hidden failure" by simultaneously optimizing for adaptation between subgroups and robustness within each subgroup. The proposed FairDRO method, which combines distribution-aware mixture-of-experts with subgroup-conditioned distributionally robust optimization, has demonstrated improved performance on several medical imaging benchmarks, particularly in reducing worst-case subgroup disparities. AI

Researchers have developed a new framework to address class imbalance in deep learning models, particularly when dealing with multi-modal data. This approach extends multi-expert architectures to fuse information from various data sources like images and tabular data. By dynamically weighting the contribution of each modality based on its informativeness, the system aims to improve recognition accuracy in long-tailed, imbalanced scenarios. AI

Researchers have introduced M$^2$E-UAV, a new benchmark and analysis framework designed to tackle the challenge of detecting small UAVs using onboard event cameras, particularly in complex motion-on-motion scenarios. This setup addresses difficulties arising when both the observer and the target are moving simultaneously, causing background clutter to obscure the UAV. The benchmark includes a substantial dataset with over 87,000 training samples and nearly 22,000 validation samples across diverse environmental conditions. Initial analysis with a point-based event model, M$^2$E-Point, shows promising results, achieving a high F1 score, though conditioning on inertial measurement unit (IMU) data provided only minor improvements. AI

Researchers have introduced OpenSGA, a novel framework for aligning 3D scene graphs, which is crucial for robots to understand and relocalize themselves in revisited environments. This new method integrates vision-language, textual, and geometric features, enhanced by a distance-gated spatial attention encoder and a minimum-cost-flow allocator for accurate object correspondence prediction. To support this work, they also released ScanNet-SG, a large-scale dataset featuring over 700,000 samples and thousands of object categories, which significantly outperforms existing scene graph alignment techniques on both frame-to-scan and scan-to-scan tasks. AI

Researchers have developed a new dataset and a transformer-based model for automated analysis of zebrafish development. This approach aims to reduce the manual labor and cost associated with current methods in drug discovery and toxicity testing. The system achieved high accuracy in classifying egg viability and detecting malformations caused by toxic exposure. AI

Researchers have identified a significant disconnect between how machine learning models for treatment effect estimation are evaluated in academic research versus industrial practice. A new study reveals that metrics used in methodological work, which rely on counterfactual outcomes, do not consistently align with observable metrics used in real-world applications. Furthermore, performance rankings on standard semi-simulated benchmarks do not reliably transfer to real-world datasets, suggesting a need to incorporate observable metrics and real-data validation into future research. AI

Researchers have introduced CoWorld-VLA, a novel framework designed to enhance end-to-end autonomous driving systems. This multi-expert world reasoning approach encodes complementary world information into expert tokens within a Vision-Language-Action model. These tokens explicitly model semantic interaction, geometric structure, dynamic evolution, and ego trajectory, serving as accessible conditioning signals for action planning. Experiments on the NAVSIM v1 benchmark demonstrate CoWorld-VLA's competitive performance in scene generation and planning, particularly in collision avoidance and trajectory accuracy. AI

DeepClaude is a new AI agent architecture that combines two distinct large language models to improve performance on complex tasks. It uses DeepSeek's R1 model for detailed reasoning and Anthropic's Claude for polished output generation. This approach aims to leverage the strengths of each model, with DeepSeek providing explicit chain-of-thought processes and Claude delivering coherent, context-aware responses. AI

Researchers have developed Gated Adaptive Positional Encoding (GAPE), a novel method to improve the performance of large language models (LLMs) with extended context lengths. GAPE addresses issues that arise when sequences exceed training limits, which can cause positional encodings like RoPE to degrade model performance. By introducing a content-aware bias into attention logits, GAPE selectively contracts irrelevant context while preserving important distant tokens, leading to sharper attention and better long-context robustness. AI

Researchers have introduced PowerStep, a novel memory-efficient optimizer for training large neural networks. Unlike traditional adaptive optimizers like Adam that store gradient statistics, PowerStep achieves adaptivity by applying a nonlinear transform to a momentum buffer. This method halves the memory required for optimizers and, when combined with quantization, can reduce memory usage by approximately eight times compared to Adam, while maintaining comparable convergence speeds. AI

Researchers have introduced a new framework for knowledge graph embedding (KGE) called KrausKGE, which leverages Kraus channel structures derived from mathematical axioms. This approach provides a principled foundation for relation operators in KGE, moving beyond externally imposed conditions. The model naturally handles complex $1$-to-$N$ and $N$-to-$N$ relations, supports multi-hop reasoning without explicit path encoders, and eliminates the need for norm constraints on entity embeddings. Empirical results show KrausKGE outperforms existing baselines, particularly on $N$-to-$N$ relations, aligning with theoretical predictions. AI

A new research paper introduces the concept of "Positive Alignment" for AI systems, moving beyond traditional safety concerns to focus on actively promoting human and ecological flourishing. This approach aims to address existing alignment failures like engagement hacking and loss of autonomy by cultivating virtues and maximizing well-being. The paper outlines technical challenges and design principles for developing AI that supports diverse values and decentralized governance. AI

OpenAI has launched a new initiative called the OpenAI Campus Network to foster AI communities within universities. This program aims to connect student clubs globally, providing them with access to AI tools and resources. The network will also support clubs in hosting events and developing AI-powered campus initiatives. AI

Researchers developed a retrieval-augmented system for Ukrainian multi-domain document understanding, achieving high accuracy in a shared task. Their pipeline incorporates contextual PDF chunking, question-aware dense retrieval, and reranking. The system utilizes Qwen models for embedding, reranking, and answer selection, demonstrating significant improvements in recall and accuracy. AI

Researchers have developed a new algorithm called Sample-Mean Anchored Thompson Sampling (Anchor-TS) to improve offline-to-online learning. This method addresses the challenge of distribution shift between offline and online data by using a novel median-based anchoring rule. Anchor-TS aims to provide more accurate estimates by correcting bias and safely leveraging offline information to accelerate online learning, with theoretical guarantees and experimental validation. AI

Researchers have developed a new system that converts expressive drum grids, a detailed MIDI format, into realistic drum audio. This method utilizes a Transformer model to predict discrete codes from a neural audio codec, which are then decoded into sound. Experiments with codecs like EnCodec, DAC, and X-Codec show that the choice of audio representation significantly impacts the quality of the synthesized drums. The system was trained and evaluated on the E-GMD dataset, demonstrating codec-token prediction as a viable approach for percussive synthesis. AI

Researchers have developed radiogenomic models capable of non-invasively predicting a specific immune cell signature in glioblastoma. These models utilize radiomic features extracted from MRI scans and transcriptomic data to identify macrophage subtype M0 immune signatures. The study, involving 176 patients across multiple datasets, demonstrated stable performance and potential for stratifying patients for immunotherapy in future clinical trials. AI

Researchers have developed a local Large Language Model (LLM) approach to classify sensitive information in government documents, specifically focusing on the deliberative process privilege for Freedom of Information Act (FOIA) requests. The study utilized the Qwen3.5 9B model, which can run on consumer-grade hardware, to avoid legal and political issues associated with cloud-based APIs. Their method, combining Chain-of-Thought and few-shot prompting with error-based examples, achieved performance comparable to commercial models and improved upon previous work in recall and F2 scores. Analysis revealed that sentences classified as deliberative often contain verbs indicating opinion and are phrased in the first person. AI

Researchers have developed a new high-probability learning theory for decentralized stochastic gradient descent (D-SGD). This theory aims to close a gap in generalization guarantees between traditional SGD and D-SGD, targeting an optimal rate of O(1/(mn) * log(1/delta)). The approach refines bounds using pointwise uniform stability and analyzes convex, strongly convex, and non-convex scenarios. It also provides high-probability results for gradient-based measures in non-convex cases and considers communication overhead for local models. AI

Researchers have introduced a new method called task calibration to improve the decision-making of large language models. This approach focuses on calibrating the model's output distribution within a task-specific latent space, rather than the entire free-form language output. By applying a decision-theoretic result, they demonstrate that Minimum Bayes Risk (MBR) decoding on this calibrated latent distribution leads to optimal generation quality across various tasks. The study also proposes Task Calibration Error (TCE) as a new metric to quantify miscalibration. AI

Professor Li Hao, chairman of MatSource, was featured in a Science magazine report highlighting his work in AI for Science. The report details three key projects from Li's team that integrate AI with material science, focusing on AI agents, machine learning potentials, and experimental material databases. MatSource is developing a closed-loop R&D system combining data, models, intelligence, and experiments to accelerate material discovery and industrial application. AI

The 2050 AGI 4 Science conference featured 17 young scholars discussing the evolving landscape of AI in scientific research. The event highlighted a shift from general AI models to deep integration within specific scientific fields, with a focus on problem-driven, interdisciplinary collaboration. Discussions explored AI's potential to tackle high-cost experimentation, reshape technical routes through first principles, and bridge the gap between academic research and industrial application. AI

Researchers have developed a new method called Graph-Structured Neural Flows (GSNF) to improve the classification of irregular multivariate time series. GSNF addresses limitations in existing Neural Flows by explicitly modeling inter-variable interactions, which were previously underexplored. The approach uses two novel self-supervision strategies: interaction-aware trajectory generation and reverse-time trajectory generation, to enhance the learning of these interactions. GSNF demonstrates state-of-the-art classification performance on multiple datasets while maintaining efficient training times and memory usage. AI

Researchers have developed V-ABS, a novel beam search framework designed to improve multi-step visual reasoning in multimodal large language models. This approach addresses the imagination-action-observer bias by iteratively refining reasoning through thinker-actor-observer cycles. V-ABS also incorporates an entropy-based adaptive weighting algorithm and a large dataset of over 80,000 samples to better balance policy priors with observational feedback. Experiments demonstrate significant performance gains, with an average improvement of 19.7% on the Qwen3-VL-8B baseline across various benchmarks. AI

Researchers have developed a new framework called IMPACT to analyze disagreements within scientific peer reviews, moving beyond simple binary contradiction detection. This system identifies specific evidence spans and assigns graded scores for the intensity of disagreement. To make this practical, IMPACT has been distilled into a smaller language model named TIDE, which can predict contradiction evidence and intensity efficiently. AI

Researchers have developed a new automated algorithm for solving infinite-state polynomial reachability games, which have applications in artificial intelligence and reactive synthesis. The proposed method utilizes ranking certificates as a proof rule to demonstrate winning strategies for the 'REACH' player. This algorithm is sound, semi-complete, and runs in sub-exponential time, outperforming existing methods on complex examples. AI

Researchers have introduced ASTRA-QA, a new benchmark designed to evaluate abstract question answering capabilities over documents. This benchmark addresses limitations in existing methods by providing explicit evaluation annotations, including answer topic sets and curated unsupported topics, to enable more robust scoring. ASTRA-QA aims to assess how well models synthesize information and avoid generating unsupported content, offering diagnostics for coverage and hallucination. AI

Researchers have developed a new framework called Standardized Loss Aggregation (SLA) to identify noisy labels in large datasets, particularly in medical imaging. SLA quantifies label reliability by aggregating standardized validation losses from repeated cross-validation runs, providing a continuous and interpretable score. This method is more efficient than existing hard-counting approaches, especially in low-noise scenarios, and can help improve dataset quality for various classification tasks. AI

Researchers have developed new methods for hyperparameter transfer specifically for Dense Associative Memories (DenseAMs). These AI architectures, characterized by neural networks with temporal dynamics on an energy landscape, present unique challenges due to shared weights and rapidly peaking activation functions. The new techniques provide explicit guidance on scaling hyperparameters from smaller models to larger ones, with theoretical findings validated by empirical results. AI

Researchers have developed Active-SAOOD, a novel method to reduce the cost of annotating oriented objects in remote sensing images. This active learning approach intelligently selects the most informative sparse samples for annotation, considering factors like orientation, classification, and localization uncertainty. Experiments show Active-SAOOD significantly boosts performance and stability, achieving a 9% gain with only 1% of data annotated. AI

Researchers have introduced OUIDecay, a novel adaptive weight decay method for convolutional neural networks. This technique dynamically adjusts regularization strength for each layer based on online activation patterns, aiming to improve training efficiency and performance. Unlike existing methods, OUIDecay does not require a validation set and has demonstrated superior results across multiple benchmark datasets and network architectures. AI

Researchers have developed MolSight, a novel approach to predicting molecular properties using only 2D images of molecular structures. This method leverages vision architectures and a chemistry-informed curriculum to analyze molecule images, achieving competitive results across various prediction tasks. MolSight demonstrates that visual analysis of molecular diagrams can be sufficient for property prediction, offering a significantly more computationally efficient alternative to existing multi-modal or graph-based methods. AI

Researchers have developed a new neural network architecture called the Phase-Coherent Transformer (PCT). This model modifies the attention mechanism of standard Transformers to better preserve phase information across layers, which is crucial for certain types of computation. Experiments show that PCT outperforms existing real-valued and complex-valued Transformers on various benchmarks, including those involving long-range memory and reasoning, without suffering from accuracy collapse at greater depths. AI

Researchers have introduced GLiNER-Relex, a novel unified framework designed to simultaneously perform named entity recognition and relation extraction. This approach extends the existing GLiNER architecture, utilizing a shared transformer encoder to process text, entity labels, and relation labels. The model is capable of zero-shot extraction for arbitrary entity and relation types specified during inference, demonstrating competitive performance on several benchmarks while maintaining computational efficiency. The framework is publicly available as an open-source Python package. AI

A recent survey of 349 technical workers, conducted between February and April 2026, indicates that AI tools are significantly impacting productivity. Participants self-reported a median increase of 1.4 to 2 times in the value of their work due to AI, with a median speed increase of 3 times. However, the researchers caution that these self-reported figures may be overstated, citing previous findings where perceived AI impact was overestimated. AI

Researchers have introduced TimeClaw, a novel AI agent designed for time-series analysis that goes beyond simple execution by learning from exploratory processes. This framework employs a four-stage loop—Explore, Compare, Distill, and Reinject—to transform exploratory executions into reusable hierarchical experience. By keeping the base model frozen and avoiding online adaptation, TimeClaw demonstrated consistent performance gains across 17 finance and weather prediction tasks in an MTBench-aligned evaluation, highlighting the importance of experience reuse in AI systems. AI

Researchers have developed TrajDLM, a new framework for generating synthetic GPS trajectories that balances efficiency with adherence to road network topology. This model treats trajectories as sequences of discrete road segments, employing a block diffusion backbone for rapid denoising and incorporating topology-aware embeddings. TrajDLM generates realistic and coherent trajectories significantly faster than previous methods and shows promise for zero-shot transfer across different transportation modes. AI

Researchers have identified two distinct timescales in generative model training: the point at which generations become rule-valid ($\tau_{\mathrm{rule}}$) and the point at which models begin reproducing training samples ($\tau_{\mathrm{mem}}$). The interval between these, termed the 'innovation window,' widens with larger datasets and narrows with increased rule complexity. This phenomenon, observed in both diffusion and autoregressive models, explains when and how these models demonstrate genuine innovation. AI

Researchers have introduced a new framework for differentially private sampling from distributions, utilizing Wasserstein distance as the primary utility measure. This approach addresses limitations of prior methods that relied on KL divergence, particularly when dealing with differing distribution supports or when geometric structure is important. The proposed Wasserstein Projection Mechanism (WPM) is designed to be minimax optimal, with accompanying algorithms for approximate computation and convergence guarantees. AI

Researchers have developed a new method called annotation-anchored training to address semantic mode collapse in large language models. This technique involves pretraining models on documents paired with semantic annotations, which helps maintain the diversity of the original pretraining data during fine-tuning. The approach allows models to generate more diverse outputs by using these annotations as anchors, reportedly reducing diversity collapse by six times compared to standard supervised fine-tuning and showing improved performance with increased model scale. AI

The AI creation platform Lingzhu has launched its second internal beta, featuring significant upgrades. Users can now access the platform without an invitation code and experience a notable performance boost due to the integration of the DeepSeek V4 large language model. This integration reportedly triples efficiency in the demand analysis phase, reducing processing time from nearly 20 seconds to under 5 seconds, alongside optimizations to the user interface. AI

Researchers have introduced R-GFM, a novel Graph Foundation Model that utilizes a Riemannian Graph-of-Graphs approach to address limitations in existing models. Unlike previous methods that use fixed-hop subgraph sampling, R-GFM models structural scale as a primary element, constructing multi-scale graphs and learning representations from Riemannian manifolds. This new architecture reportedly reduces structural domain generalization error and has achieved state-of-the-art performance, with relative improvements up to 49% on downstream tasks. AI

Researchers have explored the role of optimizers in mode connectivity within neural networks, a concept previously underexplored. Their work demonstrates that solutions generated by a single optimizer, such as AdamW or Muon, form a connected set in two-layer ReLU networks at sufficient width. The study further characterizes how regions from different optimizers interact, showing they can be disjoint or overlapping depending on regularization and network width. Empirical tests on GPT-2 pretraining revealed that paths using the same optimizer maintain spectral properties, while cross-optimizer paths exhibit smoother transitions, highlighting optimizer-dependent structures. AI