Brief

Researchers have published a paper detailing uniform scaling limits in transformers trained with the AdamW optimizer. The study models hidden-state dynamics as an interacting particle system, demonstrating convergence to a forward-backward system of ODEs. This convergence rate is dependent on the transformer's depth and number of heads, with specific mathematical bounds derived that are independent of token count and embedding dimension. AI

Researchers have developed new techniques for improving the efficiency of training large language models (LLMs). One method, Step Rejection Fine-Tuning (SRFT), leverages unsuccessful training trajectories by assessing the correctness of each step, allowing models to learn from errors without repeating them. This approach improved resolution rates on SWE-bench tasks by 3.7%. Another development, Infinite Mask Diffusion Model (IMDM), addresses factorization errors in Masked Diffusion Models (MDMs) by introducing a stochastic infinite-state mask. IMDM demonstrates superior few-step generation capabilities and surpasses existing methods on LM1B and OpenWebText datasets when combined with distillation. AI

The Shanghai Artificial Intelligence Laboratory, in collaboration with other institutions, has developed a new method for creating high-purity KrF photoresist resin, a critical material for chip manufacturing. This AI-driven approach, utilizing the "Sheng" scientific large model and discovery platform, breaks reliance on foreign suppliers and offers a standardized, rapidly iterative path for producing advanced photoresist materials. This breakthrough is part of a national initiative aimed at advancing China's capabilities in core chip material production. AI

Researchers have introduced WorldLens, a new benchmark designed to evaluate the realism and behavioral fidelity of driving world models. Current models often excel in either visual realism or physical consistency but not both, creating a gap in how their performance is assessed. WorldLens addresses this by measuring aspects like pixel quality, 4D geometry, closed-loop driving, and human perceptual alignment across 24 dimensions. Evaluations using WorldLens revealed that no single model performs optimally across all criteria, highlighting the need for more comprehensive assessment tools. AI

Researchers have published a paper detailing concentration phenomena in mean-field transformers, specifically analyzing their behavior at low temperatures during inference. The study uses a mean-field continuity equation to model token evolution and demonstrates that token distributions rapidly concentrate under a projection map induced by the transformer's matrices. This concentration remains metastable for moderate times, with the Wasserstein distance scaling in relation to temperature and inference time. AI

Researchers have developed a novel approach to solve multi-agent path finding (MAPF) problems by reformulating them as a specific type of multi-marginal optimal transport (MMOT) problem. This method leverages a Markovian structure to reduce the computational complexity of MMOT to a polynomial-sized linear program. For large-scale applications, the approach is further adapted using Schrödinger bridges, which provide an iterative, Sinkhorn-type solution that significantly reduces complexity while maintaining near-optimal results. AI

Researchers have developed a new method called TAP (Tabular Augmentation Policy) to improve the generation of synthetic tabular data, particularly in scenarios with limited real data. This approach addresses a gap where existing methods prioritize data distribution fidelity over actual utility for downstream models. TAP combines diffusion inpainting with a policy that guides the generation process towards samples that demonstrably reduce evaluation loss, leading to significant accuracy improvements on classification and regression tasks. AI

Researchers have developed a new neural exponential tilting framework for variational inference in Lévy-driven stochastic differential equations. This method addresses the intractability of Bayesian inference for processes with heavy tails and discontinuities, which are crucial for modeling extreme events in fields like finance and AI safety. The framework uses neural networks to reweight the Lévy measure, preserving jump structures while remaining computationally efficient and enabling more reliable posterior inference than Gaussian-based methods. AI

Researchers have introduced V4FinBench, a new benchmark dataset designed to evaluate AI models on corporate bankruptcy prediction. The dataset comprises over one million company-year records from Visegràd Group economies, featuring 131 financial and non-financial features across six prediction horizons. Initial evaluations show that finetuned TabPFN models perform comparably to or better than gradient boosting methods, while Llama-3-8B models lag behind on key metrics. AI

Researchers have developed BabelDOC, a new framework designed to improve PDF translation by preserving document layout. This system uses an intermediate representation to decouple visual metadata from semantic content, allowing for better handling of terminology, cross-page context, and formulas. BabelDOC's adaptive typesetting engine then re-anchors translated text to the original layout, showing improvements in fidelity, aesthetics, and consistency. AI

Researchers have developed Transcoda, a novel system for Optical Music Recognition (OMR) that can transcribe sheet music into a textual format. The system addresses the scarcity of annotated datasets by employing an advanced synthetic data generation pipeline and a grammar-based decoding approach. Transcoda, with its compact 59M-parameter model, achieves state-of-the-art performance, outperforming larger models and significantly reducing error rates on historical music scans. AI

Researchers have developed DECO-MWE, a new linguistic resource for analyzing sentiment in Korean text, specifically focusing on multiword expressions (MWEs). This resource utilizes the Local Grammar Graph (LGG) methodology, formalizing MWEs as a Finite-State Transducer. The DECO-MWE lexicon categorizes MWEs into four types, including standard polarity, domain-dependent polarity, named entity, and feature MWEs, achieving an f-measure of 0.806 in test corpora. The methodology and lexicon are intended for broad application in feature-based sentiment analysis. AI

Two new benchmarks, MMCL-Bench and Personal-VCL-Bench, have been introduced to evaluate the multimodal context learning capabilities of large language models. MMCL-Bench focuses on learning from visual rules, procedures, and evidence, while Personal-VCL-Bench assesses the ability of models to utilize user-specific visual context for personalized queries. Both benchmarks reveal significant limitations in current frontier multimodal models, indicating a substantial gap in their ability to effectively extract, reason over, and apply visual information. AI

Researchers have analyzed the regularization effects of data augmentation on supervised regression methods, particularly in scenarios where the number of covariates scales with the number of samples. The study provides a precise characterization of test error, using mean squared error, based on population quantities of the true data and statistics of the augmentation process. These findings apply to models with misspecified feature maps and architectures where only the final layer is trained, with the rest of the network being fixed or randomly initialized. AI

Researchers have developed a theoretical framework for operator learning applied to nonlinear parabolic partial differential equations (PDEs). This approach focuses on learning solution operators from finite data, emphasizing discretization invariance and PDE-specific structures. The study derives generalization error bounds that distinguish between implementation and estimation errors, showing that increased "Picard depth" can reduce truncation errors without inflating estimation errors. AI

Researchers are exploring new methods to optimize Sparse Mixture-of-Experts (SMoE) models, which are crucial for scaling large language models efficiently. One paper reveals a geometric coupling between routers and experts, suggesting that matched directions accumulate similar routed token histories and that auxiliary load-balancing losses can disrupt this structure. Another study systematically analyzed over 2,000 pretraining runs to optimize design choices like expert count and granularity, finding that these factors have a greater impact than others such as shared experts or load-balancing mechanisms. A third paper introduces DECO, an SMoE architecture designed for end-side devices that matches dense Transformer performance with significantly fewer active parameters and offers hardware acceleration. AI

Researchers have introduced MulTaBench, a new benchmark designed to evaluate multimodal tabular learning. This benchmark comprises 40 datasets that combine tabular data with either text or images, focusing on tasks where these modalities offer complementary predictive signals. The goal is to encourage the development of foundation models that can effectively integrate and leverage diverse data types for improved performance. AI

Two new benchmarks, CADBench and BenchCAD, have been released to evaluate AI's ability to generate Computer-Aided Design (CAD) programs from various inputs. These benchmarks aim to standardize the assessment of multimodal AI systems in tasks like reconstructing editable CAD programs from images or 3D models. Early evaluations show that while specialized models perform better on mesh-to-CAD tasks, current general-purpose vision-language models struggle with complex geometric details and industrial design parameters, indicating a gap in their industrial readiness. AI

Two new research papers challenge the current direction of video anomaly detection (VAD). The first paper argues that the field's focus on general models and multi-modal large language models (MLLMs) has shifted focus away from scene-specific, context-dependent anomaly identification. The second paper introduces MMVIAD, a new dataset and benchmark for industrial VAD, and presents a model called VISTA that improves performance on multi-task evaluation, outperforming GPT-5.4. AI

Researchers have developed Muown, a novel optimization method designed to improve the training of large language models. Muown addresses issues with the Muon optimizer, specifically the upward drift of spectral norms in weight matrices during training. By treating row-magnitude vectors as explicit variables, Muown enhances perplexity and learning rate stability across various model scales, outperforming existing optimizers like AdamW and Lion. AI

Researchers have developed a new method called Untargeted Jailbreak via Entropy Maximization (UJEM-KL) to bypass safety measures in vision-language models (VLMs). This technique focuses on manipulating high-entropy tokens during decoding to flip refusal outcomes, rather than relying on fixed patterns. UJEM-KL demonstrates improved transferability across different VLMs and remains effective against common defenses, suggesting that previous limitations in multimodal jailbreaks were due to overly constrained optimization objectives. AI

Researchers have identified joint sparse coding and temporal dynamics as key mechanisms for how the brain reconfigures neural representations to adapt to new contexts without losing prior knowledge. This balance is crucial for lifelong learning in dynamic environments and has implications for artificial intelligence systems struggling with catastrophic forgetting. The study found that sparsity in representations reduces interference between contexts, while temporal dynamics enhance context separation over time, leading to more stable adaptation. AI

Researchers have developed MTA-RL, a novel framework that integrates multi-modal transformer-based 3D affordances with reinforcement learning for robust urban autonomous driving. This approach fuses RGB images and LiDAR data to predict explicit, geometry-aware affordances, creating a structured observation space for the RL policy. Evaluations in the CARLA simulator demonstrate MTA-RL's superior performance in sample efficiency, stability, and zero-shot generalization compared to existing baselines. AI

Researchers have developed a new security framework to combat SQL injection attacks in applications that use large language models (LLMs) to interact with databases. These attacks exploit the translation process from natural language prompts to SQL queries, allowing malicious users to generate unsafe commands. The proposed multi-layered system includes prompt sanitization, anomaly detection, and signature-based controls to identify and block these threats, aiming to enhance the security of LLM-driven database applications. AI

Researchers have developed a new Post-Recurrent Module (PRM) to enhance the explainability and performance of Recurrent Neural Networks (RNNs) used in P300-based Brain-Computer Interfaces (BCIs). This module improves classification accuracy by 9% over existing methods while also providing insights into the spatio-temporal patterns of EEG data that contribute to model decisions. The framework aims to make EEG-based models more transparent and can be applied to various neurological tasks beyond P300 detection. AI

Researchers have developed an adaptive perception system for autonomous driving that dynamically adjusts its computational resources based on scene complexity, significantly reducing latency without sacrificing accuracy. This system, called Enhanced HOPE, also incorporates a novel linear-time interaction model and a temporal memory module to track objects through occlusions for extended periods. Separately, another research paper introduces a new adversarial attack method that leverages view-dependent camouflage on static objects to trick autonomous vehicles into inferring incorrect trajectories, potentially causing dangerous braking maneuvers. AI

Two new arXiv papers explore theoretical frameworks for sequential decision-making in machine learning. The first paper introduces a "mechanistic information" metric to quantify the value of hybrid models that combine physical priors with learned residuals, demonstrating sample-efficiency gains in simulations and cautioning against LLM priors in safety-critical applications. The second paper develops a sequential supersample framework to establish information-theoretic generalization bounds for adaptive data settings, applicable to online learning, streaming active learning, and bandits. AI

Two new research papers explore novel approaches to generative modeling, aiming to significantly speed up the process. One paper introduces W-Flow, a framework that uses Wasserstein gradient flows to compress complex evolutionary paths into a single-step generation, achieving state-of-the-art results on ImageNet with drastically reduced sampling times. The second paper investigates the theoretical underpinnings of one-shot generative flows, characterizing when such direct transport maps exist and identifying obstructions for targets with well-separated modes, particularly for Gaussian distributions. AI

Researchers have developed NCO, a new decoding strategy designed to enhance control over Large Language Model (LLM) outputs. This plug-in addresses the challenge of preventing multiple forbidden patterns, such as profanity or personally identifiable information (PII), from appearing in generated text. NCO achieves this by performing efficient online pattern matching, avoiding the state explosion issues common with converting multiple constraints into a single automaton. The strategy is compatible with standard inference methods and has demonstrated effectiveness in practical applications. AI

Researchers have developed a new topological method for analyzing dynamic Bayesian networks (DBNs). This approach associates a time-varying graph with each DBN, highlighting strong dependencies between variables. By applying persistent homology, the method generates a barcode that tracks the evolution of these dependency structures over time, offering a stable and noise-resistant summary. AI

Researchers have developed MAGE, a framework that uses a co-evolutionary knowledge graph to manage self-evolving language model agents. This approach externalizes the agent's knowledge into a graph, allowing it to learn and adapt without altering its core model. The framework has demonstrated strong performance across nine diverse benchmarks, outperforming existing methods that rely on natural language feedback or implicit reinforcement signals. AI

Researchers have developed new AI frameworks for molecular optimization, aiming to improve molecule properties while maintaining structural similarity. One approach, FORGE, uses a two-stage process that ranks and generates fragment replacements, outperforming larger models by leveraging explicit fragment-level supervision. Another method, SMER-Opt, employs a response-oriented discrete edit strategy with a single-step predictor and a multi-step planner to guide optimization trajectories through guided tree search. AI

Researchers have developed CausalGS, a new framework capable of learning the physical causality of 3D dynamic scenes directly from multi-view videos. This approach avoids the need for explicit physical priors or high-quality geometry reconstruction, instead inferring initial velocities and intrinsic material properties. The system then uses this inferred information within a differentiable physics simulator to achieve state-of-the-art performance in long-term future frame extrapolation and novel view interpolation. AI

Two new research papers explore advancements in hypergraph neural networks (HGNNs), a type of AI model designed to learn from complex, higher-order interactions. The first paper introduces the "WidthWall" concept, establishing a fundamental hierarchy of expressivity for HGNNs based on their ability to detect and count structural patterns. The second paper presents "Anchor-guided Hypergraph Condensation" (AHGCDD), a method to distill large hypergraphs into smaller, more manageable synthetic ones for efficient training of HGNNs. Both studies aim to improve the capabilities and efficiency of HGNNs for various applications. AI

Researchers have developed a novel method for training physics-informed neural networks (PINNs) by formulating the update-direction selection as a Chebyshev-center problem. This approach aims to simplify the simultaneous optimization of multiple loss terms inherent in PINNs, which often complicates their training. The new method selects a normalized direction that maximizes the minimum distance to cone facets, offering a unified geometric principle that recovers desirable properties of existing techniques without explicit imposition. Experiments indicate strong empirical performance on PINN benchmarks. AI

Researchers have developed Phoenix-VL 1.5 Medium, a 123-billion parameter multimodal and multilingual foundation model specifically adapted for the Singaporean context. This model was pre-trained on a massive 1-trillion token multimodal corpus, extended for long-context understanding, and further refined with Singapore-specific cultural, legal, and legislative data. Phoenix-VL 1.5 Medium demonstrates state-of-the-art performance on localized benchmarks while maintaining global competitiveness in general intelligence and STEM fields. AI

Researchers have developed a novel method for sensor design that synthesizes measurement likelihoods to meet specific accuracy bounds, even when sensor models are uncertain. This approach inverts the traditional design flow by starting with an error budget and then constructing the necessary likelihood function. The framework accommodates various discrepancy metrics and includes a two-layer architecture for integrating the synthesized likelihood into sensor placement and configuration. AI

Researchers have introduced Sens-VisualNews, a new benchmark dataset designed for detecting sensational content in images. The dataset comprises over 9,500 images from news items, annotated for various sensational concepts. This resource aims to advance research into identifying shocking or emotionally charged visuals that can bypass critical evaluation and accelerate viral sharing, potentially aiding in the detection of disinformation. AI

Two new research papers explore the nuances of the Adam optimizer, a popular tool in deep learning. The first paper proposes a "refresh rule" for Adam's momentum parameter, suggesting it should scale with training data size to optimize performance and robustness across different scales. The second paper delves into how mini-batch noise, influenced by batch size and Adam's hyperparameters, affects the optimizer's implicit bias and generalization capabilities, particularly in multi-epoch training scenarios. AI

Researchers have developed new methods for Langevin dynamics, a technique used in generative AI models. One paper introduces Slowly Annealed Langevin Dynamics (SALD) and Velocity-Aware SALD (VA-SALD) for training-free guided generation with diffusion models, providing theoretical convergence guarantees. Another paper presents a way to use higher-order Langevin dynamics for faster and more efficient parallel sampling from complex distributions, reducing memory and gradient-evaluation costs for models like Bayesian logistic regression and two-layer neural networks. AI

Researchers are exploring advanced techniques for interpreting the internal workings of complex AI models. One paper details the application of Sparse Autoencoders (SAEs) to Automatic Speech Recognition (ASR) systems like Whisper, revealing linguistic and non-linguistic features and demonstrating cross-lingual capabilities. Another study introduces Sparse Autoencoder Neural Operators (SAE-NOs), which represent concepts as functions rather than fixed-dimensional vectors, allowing for a more nuanced understanding of how and where concepts are expressed across input domains, particularly beneficial for data with spatial or frequency structures. AI

A new position paper published on arXiv warns that academic conferences, particularly in AI, are vulnerable to a novel threat called "Agentic Denominator Gaming." This involves using AI agents to flood conferences with low-quality submissions, not for acceptance, but to inflate the denominator of total submissions. This tactic can artificially increase the acceptance rate for legitimate papers by overwhelming reviewer capacity and degrading review quality. The paper suggests that mitigating this requires systemic policy and incentive reforms beyond just technical detection methods. AI

Researchers have developed a new annotation scheme and classifier for personal facts within dialogue systems, aiming to improve LLM personalization. The scheme expands on existing methods by adding categories like Demographics and Possessions, along with attributes for duration and validity. A classifier trained using this scheme, combined with the Gemma-300M encoder, achieved an 81.6% macro F1 score, significantly outperforming few-shot LLM baselines like GPT-5.4-mini. AI

Researchers have developed novel approaches to zero-shot anomaly detection, a technique for identifying defects in unseen categories without specific training. One method, AVA-DINO, utilizes dual specialized branches for normal and anomalous patterns, adapting frozen visual features to exploit the asymmetric distributions of normal versus anomalous data. Another approach, AnomalyClaw, frames anomaly judgment as a multi-round refutation process using a library of tools to verify against normal-sample references, improving the reliability of vision-language models for cross-domain anomaly detection. AI

Researchers are developing new methods to improve how large language models (LLMs) interact with databases. One approach focuses on enabling LLMs to query across multiple, distributed graph databases by introducing database routing and multi-database decomposition. Another study enhances existing Text2Cypher systems by incorporating grammar and schema-aware filtering during test-time inference to ensure generated queries are syntactically valid and consistent with database structures. AI

Two new research papers introduce frameworks for evaluating the metacognitive abilities of large language models. The first, TRIAGE, assesses an LLM's capacity to strategically select and sequence tasks under resource constraints, revealing significant gaps in current models' prospective control. The second, The Metacognitive Probe, offers a diagnostic tool to decompose an LLM's confidence behavior into five distinct dimensions, highlighting that standard benchmarks fail to capture a model's self-awareness of its own errors. AI

Researchers have developed new methods for Fitted Q-Evaluation (FQE) and soft Fitted Q-Iteration (soft FQI) that do not require Bellman completeness, a condition often unmet with function approximation. The proposed techniques, stationary-weighted FQE and stationary-reweighted soft FQI, address instability issues by reweighting regression steps to align with the target policy's stationary distribution. These approaches aim to improve stability and reduce value error in off-policy evaluation for reinforcement learning. AI

Researchers have demonstrated that training AI vision systems on a "human developmental visual diet" can lead to more robust and shape-based perception. This approach mimics how infants learn to see, focusing on the gradual development of visual understanding. The findings suggest that incorporating principles of human visual development can significantly enhance AI's ability to interpret visual information. AI

This post clarifies the behavioral selection model, emphasizing why distinguishing between AI motivations is crucial for predicting deployment outcomes. While the model is useful for short-to-medium term predictions, it omits significant factors like reflection and deliberation, which could be dominant drivers of AI motivations. The author presents an updated causal graph to illustrate how cognitive patterns that ensure their own influence during training are more likely to persist in deployment. AI

This article cluster explores various strategies for chunking data, a crucial step in Retrieval-Augmented Generation (RAG) systems. It details methods like fixed-size chunking, recursive character splitting, and semantic chunking, which uses embedding similarity to identify natural topic boundaries. The cluster also delves into multi-modal RAG, discussing techniques to incorporate images, tables, and other non-textual data by converting them to text, using multi-vector retrieval, or employing specialized multi-modal embeddings. AI