Brief

A recent analysis of Google's Gemma 4 E2B model revealed unexpected behavior at a context window of 2048 tokens. When presented with a truncated input, the model generated a three-part response: an initial summary, a self-disclaimer stating the summary was not in the transcript, and then a more cautious retry. This behavior was not observed at larger context window sizes, such as 32768 tokens, where the model correctly identified the input issue without hedging. The discovery corrected a previous assertion about the model's calibration capabilities. AI

Foundation models, while capable of generating polished medical reports, lack true biological understanding and operate by predicting likely word sequences rather than reasoning from first principles. This can lead to dangerous AI

The human brain's extreme energy efficiency, estimated to be 10,000 times greater than current AI models, is attributed to its sparse and localized processing. While techniques like mixture-of-experts offer a path toward similar efficiency in AI by using specialized sub-networks, they may reduce the benefits of superposition. Superposition, a dense shared representational space, allows neural networks to compress multiple features into the same neurons, contributing to their power but hindering interpretability. The author posits that more segmented architectures could weaken superposition, potentially making AI models easier to inspect and govern, and seeks a balance between efficiency, power, and interpretability. AI

The author argues for a shift in AI evaluation from focusing solely on capabilities to assessing model behaviors. While capability evaluations help forecast risks, they also accelerate AI development, creating a counterproductive cycle. Behavior evaluations, which measure tendencies like sycophancy or reward hacking, are presented as a more impactful and underinvested area that can better guide AI safety and governance. AI

Researchers are exploring the interoperability of minimal programs, drawing on concepts like Kolmogorov complexity and Solomonoff induction. The work proposes a method to construct a new, approximately shortest program for data by combining two existing approximate best compressions. This new program would generate an intermediate string and then the final data, potentially reusing components from the original programs if the intermediates are independent. AI

Researchers have developed a new neural network architecture called EarthquakeNet to improve the forecasting of weekly earthquake occurrences. This model addresses limitations in standard approaches by estimating a per-cell dispersion parameter, acknowledging spatial heterogeneity in seismic clustering. Evaluations show EarthquakeNet outperforms traditional negative binomial regression models, particularly in predicting extreme seismic events. AI

Researchers have developed EvoStruct, a novel method for designing antibody complementarity-determining regions (CDRs). EvoStruct combines a protein language model with an equivariant graph neural network to overcome vocabulary collapse issues common in existing GNN methods. This approach significantly improves amino acid recovery and diversity in CDR design, outperforming current baselines on the CHIMERA-Bench dataset. AI

Researchers have developed Velocityformer, a novel equivariant graph transformer architecture designed to enhance the reconstruction of galaxy velocities for cosmological studies. This model specifically addresses the broken symmetry inherent in observational data, leading to a significant 35% improvement in the correlation coefficient compared to standard linear theory baselines. Velocityformer demonstrates high data efficiency, achieving accuracy with minimal simulations, and shows strong generalization capabilities across different input geometries and cosmological parameters. AI

Researchers have introduced DeepWeb-Bench, a new benchmark designed to evaluate the deep research capabilities of advanced language models. This benchmark presents more challenging tasks than existing ones, requiring extensive evidence gathering from multiple sources, reconciliation of conflicting information, and multi-step reasoning over extended periods. Initial evaluations on nine frontier models revealed that derivation and calibration failures, rather than retrieval issues, are the primary obstacles, with models exhibiting distinct error patterns and domain specialization. AI

Researchers have developed a new machine learning framework to improve the accuracy of Global Navigation Satellite Systems (GNSS) positioning, particularly in challenging urban environments. The system uses activation functions to transform machine learning predictions about signal quality into weights for a weighted least squares algorithm. Experiments in Hong Kong and Tokyo showed that sigmoid activation functions consistently provided the most significant improvements in positioning accuracy across various machine learning models and GNSS configurations. AI

Researchers have developed HITL-D, a new shared control framework that combines human input with diffusion-based AI policies for robotic manipulation tasks. This system assists users by providing autonomous updates to the end effector's orientation, reducing the need for complex joystick controls and lowering mental workload. User studies showed that HITL-D significantly improved task completion times and user satisfaction compared to traditional teleoperation. AI

Researchers have developed a new policy called Fisher-SEP to help planners decide when to supplement simulators with real-world experiments. The policy decomposes the simulator's value error into identifiable calibration shifts and unresolvable parametric residuals. It also distinguishes between local and reachability components of the value gap between simulator-optimal and true optimal policies. Two case studies demonstrate Fisher-SEP's effectiveness in optimizing experimental strategies for supply chains and public health interventions. AI

Researchers have introduced Equilibrium Reasoners (EqR), a novel framework that enables scalable reasoning in iterative neural network models. EqR hypothesizes that generalizable reasoning emerges from learning task-conditioned attractors, which are dynamical systems that stabilize on valid solutions. This approach allows models to adaptively allocate computational resources based on task difficulty, significantly improving accuracy on complex problems like Sudoku-Extreme by scaling test-time compute. AI

Researchers have introduced Uni-Edit, a novel approach to tuning Unified Multimodal Models (UMMs) that enhances image understanding, generation, and editing simultaneously. Unlike traditional methods that use complex multi-task training, Uni-Edit employs a single editing task, a single training stage, and a single dataset. This is achieved by developing an automated data synthesis pipeline that transforms visual question-answering data into sophisticated editing instructions, creating the Uni-Edit-148k dataset. Experiments show that tuning solely on Uni-Edit leads to comprehensive improvements across all three capabilities without additional operations. AI

Researchers have developed a new method for animating full-body avatars, enhancing realism by incorporating latent dynamics. This approach uses a transformer-based decoder and a dynamics residual latent to capture temporal variations in appearance and geometry beyond simple pose information. A learned dynamics model evolves this latent state, decomposing updates into driving, restoring, and dissipative forces to produce coherent, history-dependent animations with minimal computational overhead. AI

Researchers have introduced FROG, a novel framework for Relational Deep Learning (RDL) that addresses the limitations of fixed graph structures in modeling relational databases. FROG formulates structure learning as a learnable table role modeling problem, enabling tables to function as both nodes and edges within message passing mechanisms. This approach allows for the joint optimization of graph structure and GNN representations, incorporating functional dependency constraints to maintain semantic consistency across different levels of representation. AI

Researchers have developed AiraXiv, an AI-driven platform designed to manage the increasing volume of research papers, including those generated by AI. This open-access system supports both human and AI scientists by facilitating continuous, feedback-driven iteration of research outputs. AiraXiv integrates AI-augmented analysis and review, and has been deployed as the submission platform for the ICAIS 2025 conference, showcasing its potential for scalable academic infrastructure. AI

Researchers have introduced Stream3D, a novel method designed to enable existing 3D generation models to process sequential video input without requiring retraining. This system maintains a dynamic 'evidential memory' that selectively stores the most relevant historical frames, ensuring temporal consistency in generated 3D outputs from video streams. Stream3D reportedly outperforms other methods in maintaining both photometric and geometric accuracy over extended sequences. AI

Researchers have developed agent just-in-time (JIT) compilation to optimize web agent planning and scheduling, significantly reducing latency and improving accuracy. This new approach compiles natural language task descriptions into executable code, allowing for LLM calls, tool usage, and parallelization. The system includes a JIT-Planner for generating and validating code plans, and a JIT-Scheduler for exploring parallelization strategies using Monte Carlo estimation. Tests across five web applications showed a 10.4x speedup and 28% accuracy increase over existing methods, with the scheduler providing an additional 2.4x speedup and 9% accuracy improvement. AI

Researchers have developed a new method called interpretable rubric embeddings to address label bias in AI models trained on historical human evaluations. This approach replaces standard black-box embeddings with features derived from expert-defined criteria, aiming to prevent models from inheriting biases present in past decisions. Empirical evaluations on a dataset of master's program applications demonstrated that this method reduces group disparities while enhancing cohort quality, offering a practical solution for learning with biased labels. AI

Researchers have developed a new method using Large Language Models (LLMs) to automatically adapt grammars following metamodel evolution in model-driven engineering. This LLM-based approach learns adaptations from previous versions, outperforming traditional rule-based methods in consistency and output similarity on smaller datasets. While effective for complex grammar scenarios, the study found LLMs struggled with adaptation consistency on very large grammars, indicating limitations for large-scale applications. AI

Researchers have introduced Mem-π, a novel framework designed to enhance adaptive memory capabilities in large language model (LLM) agents. Unlike traditional methods that rely on static retrieval from memory banks, Mem-π employs a separate language or vision-language model to generate context-specific guidance dynamically. This system learns to decide both when to produce guidance and what specific guidance to generate, using a reinforcement learning objective that allows it to abstain when unnecessary. In evaluations across various agentic benchmarks, including web navigation and tool use, Mem-π demonstrated significant improvements, outperforming retrieval-based and prior RL-optimized memory baselines with over a 30% relative gain in web navigation tasks. AI

Researchers have developed ProtoPathway, a novel multimodal framework designed for predicting cancer survival. This framework integrates whole slide imaging and transcriptomics data by using biologically grounded representations. ProtoPathway employs learnable morphological prototypes for image analysis and a graph neural network for genomic data, enabling cross-modal attention to model the relationship between molecular programs and tissue morphology. The system offers enhanced biological interpretability and reduced computational cost, demonstrating competitive performance on TCGA cancer cohorts. AI

A new study analyzed AI-generated Python refactoring pull requests on GitHub, finding that while these commits improve code quality in over 22% of cases, they also introduce new issues in nearly 25% of modified files. The research identified common AI refactoring operations and their impact on code quality and security, noting that developers merge these requests at a high rate despite the mixed outcomes. The findings suggest a need for enhanced tool-in-the-loop quality and security checks for AI-driven development workflows. AI

A new survey paper delves into the mathematical underpinnings of neural network expressivity, focusing on approximation theory. It reviews classical density results for single-hidden-layer networks and explores quantitative bounds that link approximation error to network size and function smoothness. The paper also highlights depth-width trade-offs and introduces recent theoretical attention on Kolmogorov-Arnold Networks (KANs) as an alternative architectural paradigm. AI

Researchers have developed a method to test the robustness of driving-focused Vision-Language-Action (VLA) models by applying sensor perturbations. Their study on the Alpamayo R1 model revealed that changes in Chain-of-Causation (CoC) explanations directly correlate with significant deviations in driving trajectories. The findings suggest that reasoning consistency can serve as a reliable indicator for planning safety in autonomous driving systems. AI

Researchers have introduced TempGlitch, a new benchmark designed to evaluate how well vision-language models (VLMs) can detect temporal glitches in gameplay videos. Unlike previous methods that focused on static frame anomalies, TempGlitch specifically targets glitches that only become apparent when observing changes across sequential frames. Initial tests with 12 different VLMs revealed that current models struggle significantly with this task, often exhibiting either overly cautious or overly sensitive detection, with neither larger model size nor denser frame sampling reliably improving performance. AI

A new PyTorch-native library called torchtune has been introduced to simplify the post-training phase for large language models. This library focuses on modularity and direct access to PyTorch components, aiming to facilitate efficient fine-tuning, experimentation, and deployment. Torchtune is designed to be highly flexible for research iteration and has demonstrated competitive performance and memory efficiency compared to existing frameworks like Axolotl and Unsloth. AI

Researchers have developed ReMATF, a new recurrent framework designed to mitigate atmospheric turbulence in videos. This lightweight system processes only two frames at a time, reducing computational cost and memory usage compared to existing transformer-based methods. ReMATF enhances video quality by combining a multi-scale encoder-decoder with temporal warping and a motion-adaptive fusion module, improving spatial detail and temporal stability while minimizing flicker. AI

Researchers have introduced Gaussian Sheaf Neural Networks (GSNNs), a novel framework designed for learning on relational data where node features are represented by probability distributions, specifically Gaussian distributions. Traditional Graph Neural Networks (GNNs) struggle with the geometric and algebraic structure of Gaussian means and covariances by treating them as simple vectors. GSNNs address this by incorporating these inductive biases through a new Laplacian operator derived from cellular sheaf theory, which preserves key properties relevant to Gaussian data structures. Experiments on both synthetic and real-world datasets demonstrate the practical utility of this new approach. AI

Researchers have introduced roto 2.0, a new benchmark for tactile-based reinforcement learning in robotics. This benchmark utilizes GPU parallelism and focuses on end-to-end "blind" manipulation tasks across four different robotic morphologies. The team demonstrated a significant performance improvement, with their agents achieving 13 Baoding ball rotations in 10 seconds, which is substantially faster than existing methods. By open-sourcing the environments and baseline models, they aim to lower the entry barrier for researchers in this field. AI

Researchers have developed Adaptive Signal Resuscitation (ASR), a novel training-free method to repair sparse vision networks after pruning. ASR addresses the accuracy collapse seen in high-sparsity scenarios by applying channel-wise corrections, unlike previous layer-wise methods that can over-correct damaged channels. This technique estimates variance-matching corrections for each output channel and uses a data-driven shrinkage rule to stabilize them, improving accuracy significantly, especially in high-sparsity regimes. AI

Researchers have developed PRISM, a novel method for efficient fine-tuning of large language models by prioritizing data samples that most effectively guide the model toward a desired behavior. Unlike previous approaches that treat all target examples equally, PRISM weights these examples based on the current model's preference, creating a more precise target representation. This allows PRISM to concentrate the training budget on the most impactful data, leading to improved performance in both general fine-tuning and safety-oriented tasks. AI

Researchers have developed FedCritic, a novel serverless federated learning framework for resource allocation in 6G networks. This approach addresses the challenges of inter-cell interference in ultra-dense networks by enabling decentralized critic learning through parameter averaging. FedCritic aims to improve signal quality, cell-edge rates, and overall network throughput and fairness compared to existing methods. AI

Researchers have developed a novel framework for recognizing blended emotions by selectively fusing information from multiple pre-extracted video and audio encoders. This rank-aware approach uses an attention-based gating module to identify and combine the most informative encoders, improving accuracy in distinguishing subtle and overlapping multimodal cues. The system also incorporates unsupervised domain adaptation to enhance robustness and was recognized with a second-place ranking in the BlEmoRE challenge. AI

Researchers have introduced iTryOn, a new framework designed to enhance interactive virtual try-on experiences in videos. This system addresses the limitations of current methods by enabling subjects to actively interact with their clothing, a feature previously overlooked. iTryOn utilizes a video diffusion Transformer with a multi-level interaction injection mechanism, incorporating a 3D hand prior for spatial guidance and global/action captions for semantic understanding. AI

Researchers have developed AIGaitor, a novel system for motion analysis that operates entirely on a smartphone, eliminating the need for cloud processing. This approach addresses key barriers in clinical motion capture, such as cost, complexity, and privacy concerns, as identified by rehabilitation clinicians. AIGaitor utilizes on-device neural accelerators to perform markerless monocular motion capture and deep-learning analysis, achieving processing times comparable to cloud-based systems. AI