98 items (43 posts + 55 reading notes)
The authors analyze geospatial representations in pure vision architectures, vision-language models, and multimodal foundation models. They find that textual supervision effectively enhances geospatial information learning, with language serving as a complementary modality to encode spatial context.
The paper conducts a large-scale analysis of papers involving long-form text generation in *CL conferences from 2023 to 2025, defining 20 reportable standards for human evaluation to reveal transparency and reproducibility issues in current evaluation protocols.
The authors propose IEA, a conversational image editing agent that performs editing by manipulating parameterized tools within an explicit action space. It employs a three-stage multitask alignment training process: expert editing distilled SFT, GRPO-based reinforcement learning, and large-scale synthetic fine-tuning, to achieve interpretable editing trajectories and user intent summarization.
The authors propose the Robust-U1 framework, endowing Multimodal Large Language Models (MLLMs) with explicit visual self-recovery capabilities. The method reconstructs corrupted visual content through three stages: supervised fine-tuning, dual-reward reinforcement learning, and multimodal reasoning, to enhance robust understanding.
The authors propose MemoPilot, a plug-and-play memory copilot, which formulates memory updates as a multi-turn decision problem. It utilizes multi-turn GRPO to end-to-end optimize the memory update process, enhancing the test-time learning performance of frozen LLMs in sequential interactions.
The authors transform five established benchmarks into free-form generation evaluations to demonstrate that model rankings derived from pairwise comparisons (e.g., Elo ratings) align highly with accuracy rankings based on ground truth. The study finds that pairwise comparison rankings accurately reflect model performance even with weaker evaluators, while style and bias exert limited influence on overall rankings.
The authors investigate the relationship between token-level log-probability distributions, LLM-as-judge scores, and final task accuracy within multi-agent debates. The paper utilizes a two-agent (Constructor and Auditor) debate architecture, incorporating LLM judges for fine-grained scoring of the reasoning process. It aims to diagnose reasoning quality and task correctness in external evaluations via internal confidence signals.
The authors utilize a multi-agent debate framework to investigate whether token-level log-probabilities during decoding can predict reasoning quality. The study finds that early token confidence at the beginning of generation is the strongest predictor of reasoning quality, outperforming full-sequence statistics.
The authors propose the ERAlign framework, which utilizes an Energy-Based Model (EBM) to project graph structures encoded by GNNs and text embeddings derived from LLMs into a shared latent space, achieving distribution consistency through layer-wise alignment and energy difference optimization.
The authors propose the Path-Aligned Decompression Distillation (PADD) framework, aiming to distill knowledge from dense teacher models without explicit routing to Mixture of Experts (MoE) student models. The method initializes expert functionality through teacher neuron clustering and combines online adaptive distillation with path refinement strategy optimization to achieve efficient knowledge transfer and routing learning under a fixed computational budget.
The authors propose the UniSVQ framework, unifying scalar and vector quantization by parameterizing codebooks as affine transformations of integer lattices. This approach maintains compatibility with optimized integer kernels while preserving vector quantization flexibility, introducing a data-driven block-level fine-tuning strategy to minimize reconstruction error.
The authors propose the LC-QAT framework, achieving 2-bit weight quantization-aware training through linear-constrained vector quantization. This method uses learned affine mappings to represent quantized weights, avoiding explicit codebook lookup and enabling fully differentiable end-to-end optimization.
The authors propose the SDBN framework, integrating adversarial training into Parameter-Efficient Fine-Tuning (PEFT) to enhance model robustness and generalization in low-resource and noisy environments using discrete uncertainty sets.
The authors propose the FlowTracer framework, which traces answer-oriented reasoning flows by constructing attention-induced directed acyclic graphs. The method utilizes aggregated attention weights as edge capacities and extracts the information flow backbone through flow conservation constraints, thereby assigning credit to tokens based on flow throughput.
The authors propose the Dynamic Linear Attention (DLA) framework, which employs information-aware dynamic state merging and capacity-bounded memory modeling to address the inability of fixed state merging strategies in existing linear attention methods to adapt to dynamic token importance. The method adaptively determines state boundaries based on token-level information changes, preserving high-resolution representations at semantic transitions while compressing stable regions.
The authors propose a class of sequence kernels utilizing evolutionary substitution matrices and local linearity, integrated with Gaussian Processes to construct protein property prediction models. This approach exhibits efficiency under sparse data conditions and can incorporate structural information from foundation models through structure-aware substitution matrices.
The authors propose a Social World Model (SWM) framework utilizing Large Language Models (LLMs) as the backbone. By mining temporal patterns in social data and optimizing the evidence lower bound, the paper learns state transition functions for social beliefs to predict belief evolution following major events.
The authors propose an automated, domain-agnostic framework to quantify the creativity of large language models in open-ended tasks. The framework decouples the measurement apparatus from the task, using semantic entropy to measure divergent creativity and a retrieval-based multi-agent framework to evaluate convergent creativity.
The authors conduct the first systematic multilingual evaluation of VLA models, finding that non-English instructions lead to a significant drop in success rates. The study reveals that language influence is highly uneven across task execution steps and proposes a reasoning-time intervention method based on step-wise language sensitivity to align representations.
The authors propose the AGDO framework, which utilizes attention structures to guide the denoising order in diffusion language models. Additionally, it emphasizes attention-critical tokens during fine-tuning and reinforcement learning to align training and inference dynamics.
The authors investigate the impact of model scale on the security of linear multi-agent workflows, discovering a compliance-correction symmetry. Experiments show that larger models are more prone to executing malicious instructions, but adding a lightweight terminal correction phase can restore system resilience.
The authors construct the PaperGuard benchmark to systematically evaluate and defend against adversarial attacks in multimodal AI peer review. The core involves integrating text and image attack methods and proposing a defense strategy based on chunked embedding search.
The authors propose MLUBench, a large-scale benchmark comprising 127 entities and 9 categories to evaluate the lifelong unlearning capabilities of Multimodal Large Language Models (MLLMs). Additionally, they introduce the LUMoE method, which leverages a Mixture of Experts mechanism to mitigate cumulative performance degradation resulting from continuous unlearning.
The authors propose the Human-like Criteria Probing for Hallucination Detection (HCPD) paradigm, simulating the multi-faceted reasoning of human evaluators. The core is the Human-like Criteria Probing (HCP) mechanism, where an LLM agent adaptively decomposes judgments into a weighted set of interpretable criteria and aggregates scores. It employs a reward-based alignment scheme utilizing semantic consistency for weak supervision, and uses multi-sampling aggregation during inference to ensure robustness.
The authors propose the TWLA post-training quantization framework, aiming to achieve joint compression of 1.58-bit weights and 4-bit activations. The framework collaboratively optimizes through three modules: Euclidean-to-Manifold Asymmetric Ternary Quantizer, Kronecker Orthogonal Trimodal Reshaping, and Inter-Layer Aware Activation Mixed Precision, addressing weight distribution mismatch and activation long-tail problems.
The authors propose the FCGraft framework, which maintains function-level verified code skeletons and their associated Transformer KV cache library to synthesize new policies using retrieval and grafting mechanisms. This method eliminates redundant prefilling computations through cache stitching and locally adapts to task parameters via cache patching, thereby reducing generation latency and improving robustness.
The authors construct the AuthorityBench benchmark using a 2x2 factorial design to independently manipulate claim truthfulness and citation truthfulness, aiming to isolate the influence of citation authority signals on the epistemic behavior of large language models.
The authors propose the UXBench benchmark and the UI-UX model, aiming to evaluate and enhance the user experience reasoning capabilities of multimodal large language models on UI screenshots. UXBench contains 2,000 VQA samples covering 8 tasks, while UI-UX is based on Qwen3-VL and enhances the reasoning process through reinforcement learning.
The authors propose the MACCO framework, which captures cross-modal compositional structures by masking compositional concepts in one modality and reconstructing them using the full context of the other modality.
The authors propose the PolyFlow framework, which directly embeds polytope constraints into the flow matching model and dynamics. Through discrete-time flow formulation and a projection-free architecture, it eliminates numerical integration errors and strictly satisfies arbitrary polytope constraints without iterative solvers.
The authors propose a reinforcement learning system deployed at DoorDash that leverages delayed marketplace feedback to adaptively adjust dispatch objective weights. Rather than replacing the combinatorial allocation optimizer, store-level policies select discrete multipliers to balance the trade-off between delivery quality and batching efficiency.
The authors propose a benchmark and framework for evaluating next action predictions in spreadsheets. By constructing an edit history dataset containing 12K actions and designing an online evaluation mechanism, they address the issues of lacking fine-grained operation records in public corpora and the complexity of the action space.
The authors propose the MINIM framework, serving as a trusted local agent to perform privacy-aware minimization of UI states on the client side. Based on the theory of contextual integrity, the method learns sensitivity scores and task-conditional necessity scores for each UI element, driving a ternary disclosure strategy to balance privacy and utility.
The authors propose the DRIVE framework, which optimizes offline automatic bidding strategies by decoupling action generation from decision-making through distributional action modeling, retrieval-augmented candidate generation, and value-based evaluation.
The paper reveals dual-sided data leakage risks in Split-LLMs, proposes the PIDI attack method to simultaneously reconstruct input prompts and output responses, and designs the ADMI defense mechanism combining Adapter warmup and mutual information regularization to protect privacy.
The authors propose an Adaptive Layer-wise Visual Token Selection (ALVTS) framework that utilizes a lightweight selector to dynamically identify important tokens for computation, while secondary tokens skip the current layer and are reintegrated in subsequent layers. Based on low-rank approximation with importance consistency constraints, this method simulates full attention mechanisms without requiring model retraining.
The authors propose a Riemannian metric matching framework utilizing neural networks to learn the Riemannian geometric structure of data. The core involves estimating the carré du champ operator via a denoising probabilistic framework, expressed as the conditional expectation of data random perturbations. This enables sample-level training without explicit kernel construction and amortized inference with constant cost.
The authors propose the ForceForget framework, leveraging reinforcement learning to optimize the Concept Erasing Reward (CER), eliminating unsafe content while maintaining the model's interpretability of safe semantics. By introducing a Safe Adapter module to project partial text embeddings, it achieves efficient concept modulation in cross-attention layers, avoiding over-erasure of benign concepts.
The authors propose the Persona-Pruner framework, which isolates role-specific subnetworks from general large models through role-driven data synthesis and learning-based subnetwork discovery, enabling lightweight deployment.
The authors propose Constraint-Sensitive Policy Optimization (CSPO), a first-order primal-dual method that incorporates local constraint sensitivity into policy updates. The method adds a correction term based on the shortest directed distance to the safety boundary in the primal objective, adaptively adjusting the correction magnitude using the constraint gradient norm.
The authors propose a fully self-supervised training method that eliminates the need for expensive physiological ground truth data.
Breaks through labeled data limitations; introduces the first non-contrastive unsupervised physiological signal regression framework.
The paper employs state space models to balance long-context capabilities with linear computational complexity.
The paper introduces the first non-contrastive unsupervised learning framework that discovers pulse signals without labeled data.
The authors introduce the first long-term daily care rPPG dataset with multi-scenario synchronized data.
Hybridizes handcrafted features with deep learning for a motion-robust lightweight model.
The authors propose a physics-driven model based on the Navier-Stokes equations to enhance interpretability.
The paper proposes a high-fidelity rPPG model supporting robust Heart Rate Variability (HRV) estimation.
The paper utilizes Relational Temporal Graph Neural Networks to mine the inherent periodicity within physiological signals.
The paper utilizes 3D to 2D knowledge distillation to achieve real-time video-based physiological measurement.
The paper presents non-contact physiological monitoring in pediatric intensive care units using adaptive masking and self-supervised pre-training.
The paper utilizes dual-branch structured attention fusion to balance noise and details.
The authors propose harmonic-constrained optimal transport to address cross-domain appearance variation issues.
Orthogonal Butterfly Temporal Shifting to expand the temporal receptive field.
The authors propose a label quantization framework to address noise in contact PPG labels.
A complete log of setting up a PaddleOCR development environment from scratch: Conda isolation, PaddlePaddle GPU installation, CLI verification, and Python API integration.
From web-based chat editing to Cline, then to Cursor — a personal account of embracing AI coding agents and how they genuinely transformed my development workflow.
A log of pitfalls encountered upgrading from Fedora 36 all the way to 43, plus a surprisingly pleasant experience with the stability of the latest KDE Plasma.
This video tutorial demonstrates how to configure your development environment using VSCode, Qt, and CMake for efficient C++ cross-platform development.
This article provides a comprehensive overview of recent breakthroughs in emotion recognition, covering multimodal reasoning, audio-driven facial animation, EEG analysis, and the ethics of affective computing.
This edition summarizes key papers in contrastive learning, medical image segmentation, graph neural networks, and generative models, highlighting the shift toward robust, cross-modal, and theory-driven machine learning.
This article provides an in-depth look at the definition, psychological foundations, and current state of micro-expression research in computer vision, highlighting its applications in public safety, mental health, and human-computer interaction.
This article reviews the latest research progress in diffusion models as of late 2025, covering key advancements in video super-resolution, 3D generation, robotic control, physical simulation, and model safety.
By deconstructing the Hook Model, this article explores the psychological mechanisms behind digital addiction and offers actionable strategies to regain control over your attention.
This article summarizes key recent advancements in multimodal learning, covering controllable generation, autonomous memory, robustness, multimodal reasoning, and safety detection, while analyzing evolving research trends.
Encountering an 'undefined symbol' error when importing TensorFlow I/O in Kaggle? Learn how to resolve this compatibility issue by matching the correct library versions.
This article provides a comprehensive overview of recent breakthroughs in Embodied AI, covering key areas such as 3D scene generation, Vision-Language-Action (VLA) models, multi-agent systems, spatial intelligence, and AI safety.
This article surveys key 2025 research in remote photoplethysmography (rPPG), covering multimodal fusion, lightweight model architectures, robustness in dynamic environments, and clinical validation.
This summary highlights recent breakthroughs in reinforcement learning, covering embodied AI, multi-agent systems, offline learning, communication optimization, and the integration of generative models.
Learn how to force update your Git repository to match the remote branch by discarding local changes, along with safer alternatives for preserving your work.
This summary covers recent breakthroughs in remote sensing, including large-scale disaster datasets, multimodal geospatial reasoning models, hyperspectral image restoration, and autonomous drone navigation.
Laplace's Demon was once the ultimate symbol of scientific determinism, but quantum mechanics dismantled this fantasy. This article explores how Heisenberg's uncertainty principle, quantum randomness, and entanglement prove that an omniscient predictor is physically impossible.
This roundup highlights recent advancements in video generation, multimodal understanding, and benchmarking, covering instruction-guided editing, scientific video reasoning, and efficient compression.
Why is the speed of light a universal constant? This article explores the core principles of Special Relativity and how the invariance of light speed reshaped our understanding of space and time.
This article provides a curated overview of the latest research in Vision-Language Models (VLM) and Embodied AI as of May 2025, covering key areas such as 3D scene generation, multi-agent collaboration, robotic control, and security.
If you encounter a curl 56 network error while cloning the OpenVINO repository from Gitee, you can resolve it by adjusting your Git buffer configuration.
By default, Docker requires sudo to execute commands. This guide explains how to enable non-root access by adding your user to the docker group and troubleshooting common permission issues.
The face_recognition library relies on CPU by default. This guide explains how to enable GPU acceleration by recompiling dlib with CUDA support.
This guide explains how to containerize Python applications that require GPU acceleration and USB camera access using Docker, including private registry configuration and disk management.
This guide explains how to compile and install GNUPlot 5.4.3 from source on CentOS 8 or AWS Linux 2, including how to resolve libgd compatibility issues.
This guide provides step-by-step instructions for compiling and installing FFmpeg with CUDA hardware acceleration support on CentOS 8.
This article provides an introduction to using the C++ `<regex>` standard library for string pattern matching and manipulation, covering basic usage and core components.
This article explains how to update legacy CvxText code to function correctly in OpenCV 4.5 by resolving header file inclusions and data type conversion issues.
This guide demonstrates how to implement FTP file uploads using POCO, a lightweight and flexible C++ network library.
Text detection is a classic computer vision task, and curved text presents a unique challenge due to its free-form nature. This guide demonstrates how to perform curved text detection using PaddleOCR.
This guide explores how to map common NumPy array operations—such as Sigmoid functions, channel slicing, and conditional filtering—to C++ using OpenCV's cv::Mat.
Learn how to display real-time video streams from a camera in a Qt interface using OpenCV and the official PySide6 library in this concise guide.
This guide walks you through installing the MMOCR framework and demonstrates how to convert PASCAL VOC annotations to COCO format for custom model training.
A concise guide to installing and configuring GitLab on a local Ubuntu 16.04 server without a domain name, bypassing the complexity of standard documentation.
Learn how to use the `&` operator in OpenCV to perform intersection operations on `cv::Rect`, allowing for elegant image ROI clipping and boundary validation without verbose conditional logic.
This post introduces the MNIST-ROT (Rotated MNIST) dataset, a standard benchmark for evaluating rotation-equivariant algorithms, and provides the necessary download links.
An overview of common PyTorch learning rate scheduling strategies and how to implement them to optimize the training process, using StepLR as a practical example.
This guide demonstrates how to use the OpenCV DNN module for image classification, covering core APIs, the differences between Mat and Blob formats, and the implementation of both single-image and batch inference.
A guide on how to perform HTTP POST requests with JSON payloads in Qt and how to build a corresponding local server using the QtHttpServer module.
A practical guide on using CMake to manage C/C++ projects, covering project setup and integration with popular libraries like OpenCV, Boost, Qt, and CUDA.
Learn how to compile the QtHttpServer module and solve the QCoreApplication event loop initialization challenge when wrapping Qt functionality into a C-compatible dynamic library.
A step-by-step guide on how to compile a custom OpenCV library with CUDA and DNN acceleration support from source using CMake and Visual Studio 2019.
A step-by-step guide on how to install and configure the FFTW 3.3.2 library in a Visual Studio 2010 environment, including .lib file generation and project property setup.