AI × Crypto News Feed

Parameter-Efficient Neuroevolution for Diverse LLM Generation: Quality-Diversity Optimization via Prompt Embedding Evolution

Researchers introduce QD-LLM, a framework that evolves lightweight prompt embeddings (~32K parameters) to steer frozen large language models toward diverse outputs without fine-tuning. The approach outperforms existing quality-diversity optimization methods by 46.4% in coverage and demonstrates practical applications in test generation and training data improvement.

Multi-Tier Labeling and Physics-Informed Learning for Orbital Anomaly Detection at Scale

Researchers developed a multi-tier labeling system combining physics-based rules, Kalman filtering, and machine learning to detect orbital anomalies across thousands of LEO satellites. The approach generated 8.6M labeled training sequences from 232M historical records, enabling a Transformer model to achieve 55.4% maneuver recall and 62.8% decay recall—addressing a critical gap in space situational awareness infrastructure.

CrossVL: Complexity-Aware Feature Routing and Paired Curriculum for Cross-View Vision-Language Detection

CrossVL introduces a novel framework combining Complexity-Aware Pathway Aggregation and Paired Curriculum Learning to improve vision-language model performance in cross-view object detection scenarios. The approach addresses fundamental challenges when models operate across different viewpoints (ground and aerial), achieving measurable improvements in detection accuracy and consistency on the MAVREC dataset.

Insight: Enhancing Mobile Accessibility for Blind and Visually Impaired Users with LLMs

Researchers introduce Insight, an Android accessibility service leveraging large language models to provide natural language interaction and real-time screen summarization for blind and visually impaired users. A comparative study shows Insight reduces mental effort and task completion time compared to TalkBack, though users identified a need for better interruption management.

CalBench: Evaluating Coordination-Privacy Trade-offs in Multi-Agent LLMs

Researchers introduce CalBench, a controlled evaluation framework for testing multi-agent LLM coordination in calendar scheduling scenarios where agents must negotiate shared commitments while protecting private information. The benchmark measures coordination quality, communication efficiency, fairness, and privacy leakage in decentralized systems where no single agent has complete information.

Fashion Florence: Fine-Tuning Florence-2 for Structured Fashion Attribute Extraction

Researchers have fine-tuned Florence-2, a vision-language model, to extract structured fashion attributes from clothing images with 94.6% category accuracy. The resulting model, Fashion Florence, outperforms GPT-4o-mini and Gemini 2.5 Flash on fashion-specific tasks while running efficiently at 0.77B parameters, demonstrating specialized AI models can exceed general-purpose alternatives in narrow domains.

Free Energy Manifold: Score-Based Inference for Hybrid Bayesian Networks

Researchers introduce Free Energy Manifold (FEM), a score-based conditional energy model designed to improve probabilistic inference in hybrid Bayesian networks containing both discrete and continuous variables. The work identifies and addresses a critical failure mode called the mode-bridge artifact, where standard energy models create artificially low-energy paths between separated probability modes, leading to overconfident predictions in regions not seen during training.

ChladniSonify: A Visual-Acoustic Mapping Method for Chladni Patterns in New Media Art Creation

ChladniSonify presents a real-time system that maps visual Chladni patterns to acoustic frequencies using deep learning and plate theory, achieving 99.33% classification accuracy with sub-50ms latency. The engineering prototype bridges audio-visual art creation by automating the traditionally subjective mapping between vibration patterns and sound, addressing technical barriers in new media art workflows.

Probing Routing-Conditional Calibration in Attention-Residual Transformers

Researchers question whether routing traces in Attention-Residual transformers provide genuine evidence of improved post-hoc calibration beyond standard confidence metrics. Through rigorous statistical testing with matched controls, the study finds that routing-specific features offer minimal stable evidence of better calibration, suggesting previous claims of calibration improvements may reflect methodological artifacts rather than true model improvements.

MoPO: Incorporating Motion Prior for Occluded Human Mesh Recovery

Researchers introduce MoPO, a novel method for recovering human mesh models from occluded images by leveraging motion prediction from pose sequences. The approach combines spatial-temporal occlusion detection with lightweight motion prediction to estimate hidden body parts, achieving state-of-the-art results on occlusion benchmarks while reducing temporal inconsistencies.

UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning

Researchers introduce UFO, a framework addressing robust continual graph learning by simultaneously tackling catastrophic forgetting and noisy data supervision in evolving graphs. The method uses flow-based generative modeling to mitigate forgetting and instance-level reliability scoring to handle corrupted labels, demonstrating superior performance across benchmark datasets.

Nautilus Compass: Black-box Persona Drift Detection for Production LLM Agents

Nautilus Compass is a black-box persona drift detector for LLM coding agents that operates without access to model weights, making it compatible with closed APIs like Claude and GPT-4. The system detects when production agents forget user constraints or contradict prior agreements using embedding-based similarity matching, achieving 0.83 ROC AUC on drift detection while costing $3.50 per evaluation—substantially cheaper than alternatives.

Intervention-Based Time Series Causal Discovery via Simulator-Generated Interventional Distributions

Researchers introduce SVAR-FM, a framework that uses physics-based simulators to discover causal relationships in time series data by treating simulation interventions as Pearl's do operator. The method recovers correct causal directions where observational methods fail due to confounding, with theoretical guarantees and empirical validation across multiple scientific domains.

EgoMemReason: A Memory-Driven Reasoning Benchmark for Long-Horizon Egocentric Video Understanding

Researchers introduce EgoMemReason, a comprehensive benchmark for evaluating AI systems on week-long egocentric video understanding through memory-driven reasoning. The benchmark reveals that even state-of-the-art multimodal models achieve only 39.6% accuracy, indicating that long-horizon memory and temporal reasoning remain unsolved challenges for next-generation visual assistants.

Hyperbolic Distillation: Geometry-Guided Cross-Modal Transfer for Robust 3D Object Detection

Researchers propose HGC-Det, a hyperbolic geometry-based cross-modal distillation framework for 3D object detection that integrates point cloud and image data more effectively. The method addresses modality heterogeneity and spatial misalignment issues through three specialized components and demonstrates improved performance across indoor and outdoor datasets.

Rethinking Random Transformers as Adaptive Sequence Smoothers for Sleep Staging

Researchers challenge the assumption that Transformers improve sleep staging through learning complex dependencies, instead revealing that random, untrained Transformers substantially boost performance by acting as adaptive smoothers. The findings suggest sleep staging relies more on architectural inductive bias than parameter learning, enabling simpler, more efficient models suitable for edge deployment in healthcare systems.

NaiAD: Initiate Data-Driven Research for LLM Advertising

Researchers introduce NaiAD, a comprehensive dataset of nearly 59,000 ad-embedded LLM responses designed to optimize advertising within AI systems while maintaining user experience. The framework uses mechanistic analysis to identify four semantic strategies for effective ad integration and employs human-calibrated scoring to enable independent control of user and commercial utility objectives.

Verifier-Free RL for LLMs via Intrinsic Gradient-Norm Reward

Researchers propose VIGOR, a verifier-free reinforcement learning method for large language models that eliminates dependency on gold labels or domain-specific verifiers by using gradient-norm measurements as intrinsic reward signals. The approach demonstrates measurable improvements over existing baselines on mathematical reasoning and exhibits cross-domain transfer to code tasks, addressing a major scalability constraint in current RL-based LLM training.

PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning

Researchers introduce PruneTIR, an inference-time optimization framework that improves tool-integrated reasoning in large language models by pruning failed trajectories, resampling tool calls, and suspending tool usage when errors persist. The approach enhances LLM performance without requiring additional training, demonstrating significant improvements in accuracy and efficiency.

Novel GPU Boruta algorithms for feature selection from high-dimensional data

Researchers have developed GPU-accelerated versions of the Boruta feature selection algorithm, significantly improving computational efficiency for processing large-scale datasets while maintaining accuracy comparable to the original CPU-based method. The two variants—Boruta-Permut and Boruta-TreeImp—demonstrate that GPU acceleration offers a cost-effective solution for machine learning workflows on high-dimensional data.

SDTalk: Structured Facial Priors and Dual-Branch Motion Fields for Generalizable Gaussian Talking Head Synthesis

SDTalk introduces a generalizable 3D Gaussian Splatting framework for talking head synthesis that works across different identities without requiring personalized training. The method uses structured facial priors and dual-branch motion fields to achieve high-quality, real-time synthesis from single images.

HapticLDM: A Diffusion Model for Text-to-Vibrotactile Generation

Researchers introduce HapticLDM, a diffusion model that generates haptic feedback from text descriptions, outperforming previous autoregressive approaches in realism and semantic accuracy. The breakthrough enables more efficient vibration design for metaverse, gaming, and film applications by improving how AI converts natural language into precise vibrotactile experiences.

GLiNER2-PII: A Multilingual Model for Personally Identifiable Information Extraction

Researchers have developed GLiNER2-PII, a compact 0.3B-parameter multilingual model for detecting personally identifiable information across 42 entity types at character-level precision. Trained on a synthetic corpus of 4,910 annotated texts to overcome privacy constraints in real data collection, the model outperforms existing systems including OpenAI's Privacy Filter on benchmark evaluations and is now publicly available on Hugging Face.

Geometric 4D Stitching for Grounded 4D Generation

Researchers introduce Geometric 4D Stitching, a novel framework that improves 4D scene generation by explicitly identifying and filling geometric gaps with geometrically consistent components. The method achieves efficient 4D scene reconstruction in under 10 minutes on consumer hardware while supporting iterative scene expansion and editing capabilities.

Attention Drift: What Autoregressive Speculative Decoding Models Learn