Brief

Managing prompts for large language models (LLMs) requires a structured approach similar to software development. This involves versioning prompts, implementing automated testing, and establishing deployment pipelines to ensure consistency and reliability. Tools and workflows can help teams treat prompts as code, storing them in version-controlled formats and using registries to track different versions and their statuses. AI

Researchers have developed a new pipeline to convert monocular UAV traffic video into a bird's-eye-view (BEV) representation. This method uses visible road geometry, such as lane markings, to estimate a homography that maps image coordinates to metric ground-plane coordinates. The system can then project vehicle observations into BEV, enabling the estimation of vehicle direction, speed, and dynamic 3D cuboids, which supports traffic analytics and the creation of digital-twin systems. 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

This article emphasizes the critical importance of dataset preparation before engaging in model fine-tuning. It details how a well-structured and relevant dataset is foundational for successful fine-tuning, regardless of whether the model is a large language model (LLM) or a smaller one (SLM). The author advocates for prioritizing dataset creation and refinement as the initial and most crucial step in the fine-tuning process. AI

This article explores the practice of fine-tuning smaller language models, distinguishing them from larger counterparts. It details how this process can adapt general-purpose models for specific applications, particularly in the realm of security. The author aims to provide a comprehensive understanding of fine-tuning techniques and their implications. AI

Large language models used for AI-assisted vulnerability discovery can falsely present information from their training data as novel findings. This occurs because LLMs cannot distinguish between recalling information about known vulnerabilities and reasoning about new code. To combat this, researchers propose a validation workflow that involves checking AI-generated findings against public databases like NVD and examining the code's Git history to determine if the vulnerability was previously disclosed or patched. 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 Pixal3D, a novel approach to generating high-fidelity 3D assets from images. This method directly generates 3D models in a pixel-aligned manner, ensuring greater faithfulness to the input view by explicitly mapping image features to 3D space. Pixal3D improves upon existing techniques by establishing clear pixel-to-3D correspondence, leading to more realistic and accurate 3D reconstructions. The framework also extends to multi-view generation and scene synthesis. AI

Researchers have developed a confidence-guided diffusion augmentation method to improve the recognition of handwritten Bangla compound characters. This approach uses diffusion models to generate high-quality synthetic character samples, enhanced by Squeeze-and-Excitation blocks and a confidence-based filtering mechanism. When trained on these augmented datasets, several classification architectures, including ResNet50 and Vision Transformers, showed significant performance gains. The best model achieved 89.2% accuracy on the AIBangla dataset, surpassing previous benchmarks and demonstrating the effectiveness of quality-aware augmentation in low-resource script recognition. AI

Researchers have developed DataMaster, a novel framework designed to automate the data engineering process for machine learning. This system aims to improve ML model performance by optimizing data selection, composition, and transformation, rather than altering the learning algorithm itself. DataMaster integrates tree-structured search, a shared data pool, and cumulative memory to efficiently explore the data landscape and learn from previous attempts, ultimately enhancing downstream model outcomes. AI

Researchers have developed a new method to formally verify the safety of Large Language Model (LLM) guardrail classifiers, moving beyond traditional red-teaming. This approach shifts verification from the discrete input space to the classifier's pre-activation space, defining harmful regions as convex shapes. By analyzing these regions, the researchers found verifiable safety holes in tested guardrail classifiers, revealing that empirical metrics alone can be misleading. The study also highlighted significant differences in the structural stability of safety guarantees across models like BERT, GPT-2, and Llama-3.1-8B. AI

Researchers have developed a new method for stress testing image classification models, particularly in medical imaging, to address issues arising from distribution shifts. This counterfactual stress testing framework uses causal generative models to create realistic "what if" scenarios by altering attributes like scanner type or patient sex while maintaining anatomical integrity. Experiments on chest X-ray and mammography data demonstrated that this approach provides a more accurate assessment of out-of-distribution performance compared to traditional perturbation methods, offering a more reliable evaluation for AI systems before deployment. AI

Researchers have developed a new framework called BICR (Blind-Image Contrastive Ranking) to assess the confidence of Large Vision-Language Models (LVLMs). This method helps distinguish between predictions genuinely informed by visual input and those relying solely on language priors. BICR trains a lightweight probe to contrast hidden states from the LVLM with and without the image, penalizing higher confidence when the image is obscured. Evaluated on multiple LVLMs and diverse tasks, BICR demonstrated superior calibration and discrimination with significantly fewer parameters than existing baselines. AI

Researchers have developed a new formal framework for probabilistic safety shields in Markov Decision Processes (MDPs). This framework addresses the complexities of ensuring safety when a certain probability of undesirable events is acceptable. The paper introduces constructions for both offline and online shields that maintain strong safety guarantees, supported by empirical evaluations demonstrating their practical advantages and computational feasibility. 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 developed LoKA, a framework designed to make low-precision arithmetic, specifically FP8, practical for large recommendation models (LRMs). Unlike previous attempts that often degraded model quality, LoKA employs a system-model co-design approach. It achieves this through statistical profiling to identify safe FP8 adoption points, model adaptations for improved stability and efficiency, and a runtime that selects optimal FP8 kernels based on accuracy requirements. 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 RUBEN, a new tool designed to generate rule-based explanations for retrieval-augmented large language models. This system uses pruning strategies to identify a minimal set of rules that effectively explain the model's outputs. The paper also highlights RUBEN's utility in enhancing LLM safety by testing the robustness of safety training and the impact of adversarial prompts. AI

Researchers have introduced the Generalized Turing Test (GTT), a new formal framework designed to compare the intelligence of arbitrary agents through indistinguishability. This framework defines a 'Turing comparator' to determine if one agent cannot be reliably distinguished from another, offering a task- and dataset-agnostic measure of relative intelligence. Initial empirical evaluations on modern AI models using the GTT framework suggest it yields meaningful comparative orderings that align with existing rankings. AI

A new research paper introduces a diagnostic framework called [METHOD NAME] to expose the unreliability of self-verification in medical visual question answering (VQA) systems. The study argues that current self-verification methods, where a vision-language model (VLM) checks its own answers, create a "verification mirage" by falsely accepting incorrect responses. This phenomenon is particularly pronounced in knowledge-intensive clinical tasks and is exacerbated by a "lazy verifier" that under-attends to image evidence. AI

Researchers have developed a new evaluation protocol for AI pentesting agents that moves beyond simplified benchmarks to assess real-world vulnerability discovery. This protocol incorporates structured ground-truth, LLM-based semantic matching, and methods to handle ambiguity and stochasticity for more operationally relevant comparisons. The team has also released the code and expert-annotated ground truth to ensure reproducibility. 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 a novel method for 3D scene reconstruction by integrating diffusion models with Neural Radiance Fields (NeRF). This approach treats 3D reconstruction as a probabilistic problem, using a stochastic latent variable to represent the scene. The model learns a prior over these latents and performs posterior inference using diffusion models combined with a reconstruction likelihood term derived from volumetric rendering. The system demonstrates accurate 3D structure prediction from various inputs, including single-view, multi-view, noisy images, sparse pixels, and sparse depth data, effectively modeling the uncertainty associated with each observation type. AI

Researchers have developed SLIM, a novel framework designed to enhance the interpretability and property-directed editing capabilities of large language models in molecular design. SLIM utilizes a Sparse Autoencoder with learnable gates to decompose the model's hidden states into sparse, property-aligned features. This approach allows for precise steering of property-relevant dimensions without altering the model's core parameters, significantly improving editing success rates. Experiments on the MolEditRL benchmark demonstrated substantial gains, with improvements up to 42.4 points across various molecular properties and model architectures. AI

Researchers have introduced NoRIN, a novel non-linear reversible normalization technique for time-series forecasting that goes beyond the linear affine transformations of existing methods like RevIN. NoRIN utilizes a Johnson $S_U$ transform with parameters that can adjust for distribution tails and skewness, unlike RevIN's limitations. The method decouples shape parameter optimization from gradient training, using a quantile fit and Bayesian optimization to prevent the model from defaulting to a linear form, demonstrating that different network architectures benefit from distinct normalization parameters. AI

Researchers have introduced SensorFault-Bench, a new protocol designed to evaluate the robustness of forecasting models in cyber-physical systems. This benchmark addresses the common issue where models perform well under ideal conditions but degrade significantly when faced with noisy, missing, or misaligned sensor data. The protocol uses real-world datasets and a standardized severity model to assess model performance under various fault scenarios, providing metrics like worst-scenario degradation and fault-time MSE. Initial evaluations showed that models favored by clean MSE metrics can perform poorly under faults, and even advanced models like Chronos-2 struggled compared to simpler methods in certain fault conditions. AI

Researchers have introduced MaD Physics, a new benchmark designed to evaluate AI agents' ability to conduct scientific discovery under real-world constraints. This benchmark focuses on how agents make measurements and draw conclusions when faced with limitations on the quality and quantity of data they can collect. The system includes three environments based on altered physical laws to prevent prior knowledge contamination, challenging agents to infer underlying principles and make future predictions within a set budget. Initial evaluations using various Gemini models revealed shortcomings in their structured exploration and data collection abilities, indicating areas for improvement in scientific reasoning. AI

Researchers have developed a method for creating periodic distributed representations using Fourier embeddings, which can better handle and distinguish nearby angles compared to traditional scalar representations. This approach allows for control over dot product similarities and the construction of various kernel shapes. The work formalizes Dirichlet and periodic Gaussian kernels within the Spatial Semantic Pointers framework. 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

Researchers have developed NanoResearch, a multi-agent framework designed to personalize AI-driven research automation. This system addresses limitations in current tools by accumulating procedural knowledge, retaining user-specific experience, and learning implicit preferences. Through a tri-level co-evolution of skills, memory, and policy, NanoResearch aims to make AI research assistants more usable by adapting to individual researchers' needs and workflows. AI

This paper introduces a new framework for analyzing Q-value iteration in Markov decision processes, focusing on a technique called rank-one deflation. The authors interpret the algorithm's behavior through the geometry of switching systems, providing a novel JSR-based convergence analysis. Their findings suggest that deflation offers a more precise characterization of convergence rates by removing a redundant component, without altering the fundamental decision-making problem or the resulting policy sequence. AI

Researchers have developed a new self-supervised benchmark for evaluating language models on mathematical text continuations. This benchmark uses likelihood scoring to assess how well a model's auxiliary forecast string transmits information about a hidden continuation, such as the rest of a displayed equation. Tests on models like GPT-5.5 and Opus 4.7 showed they could distinguish between model families and reasoning efforts, even when scorers were fine-tuned to emulate shortcut vulnerabilities. The findings suggest cross-model likelihood scoring is a viable method for static benchmarking and probing shortcut vulnerabilities before further optimization. AI

Researchers evaluated domain-adapted language models for threat modeling in 5G security using the STRIDE approach. Their empirical study, involving 52 configurations across 8 language models, found that domain adaptation did not consistently improve performance over general-purpose models. Decoding strategies and model scale showed significant impact, but larger models did not guarantee reliable threat modeling, suggesting a need for better task-specific reasoning and security grounding. AI

A new paper explores the dual nature of Large Language Models (LLMs) in hardware design, highlighting both their potential to revolutionize the semiconductor industry and the significant security risks they introduce. The research details how LLMs can accelerate tasks like RTL code generation and testbench automation, but also warns of vulnerabilities such as data contamination and adversarial evasion. The paper proposes countermeasures including dynamic benchmarking and red-teaming to foster secure and trustworthy design ecosystems. AI

A new research paper introduces a method called RACER (Robust Adaptive Cost-Efficient Routing) to optimize the use of large language models (LLMs) as judges. The study found that while explicit reasoning in LLMs significantly improves accuracy for complex tasks like math and coding, it offers minimal gains for simpler evaluations and incurs higher computational costs. RACER dynamically selects between reasoning and non-reasoning LLM judges within a fixed budget, addressing potential distribution shifts and aiming for superior accuracy-cost trade-offs. AI

A new research paper identifies a significant flaw in chain-of-thought (CoT) corruption studies, which are used to evaluate the faithfulness of AI reasoning. The study found that these evaluations often mistakenly identify the location of the final answer as the most computationally important part of the reasoning process, rather than the actual computational steps. This format confound was demonstrated by ablating the answer statement, which drastically reduced sensitivity to corruption in the reasoning steps. AI

Researchers have analyzed the limits of factual recall in linear associative memories, a simplified model for understanding how neural networks store and retrieve information. They found that a decoupled model accurately represents the original model's storage capacity and learning mechanisms. Using statistical physics, the study determined that these networks can store up to approximately half an association per dimension squared, offering insights into the memory capabilities of more complex neural architectures. AI

Researchers have developed a new automated pipeline for estimating forest fuel load using virtual remote sensing data from Google Earth Studio. This method employs a feed-forward Transformer model called Pi-Long for 3D reconstruction and introduces a metric recovery module to resolve scale ambiguity. The system then generates height and density maps to classify tree species, calculate Leaf Area Index, and estimate total fuel load, offering a cost-effective alternative to traditional methods. 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 introduced TrajPrism, a new benchmark designed to evaluate language-grounded urban trajectory understanding. This benchmark addresses the gap in prior work by unifying instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning. TrajPrism comprises 300,000 trajectories from three cities, yielding over 2.1 million task instances, and includes proof-of-concept models to demonstrate its utility. AI

Researchers have introduced LITMUS, a new benchmark designed to evaluate the behavioral safety of LLM agents operating within real OS environments. This benchmark addresses limitations in existing safety evaluations by incorporating a dual verification mechanism that assesses both semantic and physical-layer OS operations, along with OS-level state rollback to prevent test contamination. Initial evaluations using LITMUS revealed that current frontier agents, including strong models like Claude Sonnet 4.6, exhibit significant safety vulnerabilities, with a high percentage of dangerous operations being executed and a phenomenon termed 'Execution Hallucination' where agents verbally refuse but still perform harmful actions. AI

Researchers have developed a new method called DLR-Lock to prevent unauthorized modifications of open-weight language models. This technique replaces standard MLPs with deep low-rank residual networks, which increase memory usage during backpropagation and complicate the fine-tuning optimization landscape. DLR-Lock aims to defend against adaptive attackers who have full knowledge of the model and defense strategy, while preserving the original model's capabilities, as validated by experiments on LLMs. AI

Researchers have theoretically analyzed the convergence properties of gradient descent in training wide, shallow neural networks with bounded nonlinearities. Their work extends previous findings beyond simple ReLU or sigmoid activations to more complex architectures like multi-head attention layers and two-layer sigmoid networks with vector output weights. The study proves that non-global minimizers are unstable under gradient descent dynamics, ensuring convergence to global minimizers when initial parameters have full support. AI

Researchers have developed a new benchmark and training methodology for applying large language-vision models (LLVMs) to automatic target recognition (ATR) using synthetic aperture radar (SAR) imagery. The study leverages transformer-based LLVMs like CLIP and LLaVA, extending the MSTAR dataset with text captions and question-answer pairs. Using parameter-efficient fine-tuning, an LLVM achieved 98% accuracy in identifying fine-grained target qualities, aiming to enhance machine-assisted remote sensing for military and intelligence applications. 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 GridProbe, a novel method to improve the efficiency of long-video Visual Language Models (VLMs). This technique adaptively selects relevant frames during inference, reducing the computational cost associated with processing thousands of frames. GridProbe achieves this by probing frame importance in the answer space, allowing for a dynamic adjustment of the number of frames processed based on question difficulty without sacrificing accuracy. AI

Researchers have introduced RadThinking, a new dataset designed to train AI systems in longitudinal clinical reasoning for radiology. The dataset includes visual question-answering pairs across three difficulty levels, focusing on atomic perception, single-step reasoning, and multi-step compositional reasoning. RadThinking aims to enable AI to not just detect cancer but also reason about it, using over 20,000 CT scans and incorporating clinical reporting standards. AI