Live intelligence

Latest AI updates

579 updates

Research Editor selected

BRIDGE: Balancing Representations by Generating Under-represented Data for Long-tailed Self-Supervised Learning

BRIDGE is a label-free feedback loop that improves self-supervised learning on long-tailed data by identifying under-represented images via kNN distance and generating local variants with a frozen diffusion model. It outperforms baselines across five long-tailed regimes on linear-probe and kNN tasks, and extends gains to ViT-S, low-label fine-tuning, and segmentation. The method requires no labels and leaves the pretraining objective unchanged.

Why it matters: This paper presents a concrete, novel method with strong empirical results for a common real-world problem (long-tailed data) in self-supervised learning, offering practical value for researchers and practitioners.

arxiv2d agoScore 85

self-supervised learninglong-tailed recognitiondata generation

Research Editor selected

Specification-Aware Predictive Uncertainty Decomposition

This paper introduces a probabilistic framework that decomposes predictive uncertainty into specification, residual aleatoric, and epistemic components, addressing underspecified inputs in LLMs. It shows that aleatoric uncertainty splits into specification uncertainty (expected reduction from revealing user intent) and residual aleatoric uncertainty (remaining risk after intent is known). Experiments across domains indicate specification uncertainty best detects underspecified inputs, while residual aleatoric uncertainty best detects in-distribution errors.

Why it matters: Provides a novel, formal decomposition of uncertainty with practical implications for LLM clarification and error detection, backed by experiments.

arxiv2d agoScore 82

predictive uncertaintylarge language modelsBayesian risk

Research Editor selected

CAREFUL-Bench: Effect-Aware, Judge-Free Evaluation of Multi-Turn LLM Function Calling

CAREFUL-Bench enhances the Berkeley Function Calling Leaderboard (BFCL) by making evaluation effect-aware without replacing it. It infers effect declarations for all 128 functions via simulator execution and uses a judge-free replay scorer that reproduces official results while reporting unrequested destructive calls, duplicate writes, redundant reads, and token costs. Across 11 open models, effect-aware scoring preserves ranking but reveals quality gaps, and a confirmation-aware protocol shows models rarely obtain explicit permission before destructive actions.

Why it matters: Provides a concrete, recent improvement to function-calling benchmarks, exposing critical limitations in current evaluation and offering actionable metrics for model developers.

arxiv2d agoScore 82

LLM function callingbenchmark evaluationagent safety

Research Editor selected

BLPO: Beyond-Local Policy Optimization for Long-Horizon LLM Agents

BLPO is a new reinforcement learning method for long-horizon LLM agents that estimates task-progress advantages from shared experience across tasks and updates, avoiding the need for a learned critic or extra rollouts. It summarizes trajectory prefixes into task-progress states and abstracts actions by function, using a global memory with balanced aggregation to control bias. Experiments on ALFWorld and WebShop with 1.5B and 7B policies show improved task success and faster learning under matched training budgets.

Why it matters: This paper introduces a concrete algorithmic improvement for training LLM agents on long-horizon tasks, with empirical gains on standard benchmarks, making it valuable for researchers and practitioners.

arxiv2d agoScore 85

LLM agentsreinforcement learninglong-horizon tasks

Research

Q-CamTuning: Multimodal LLM for Fine-Grained Quality Reasoning in Livestreaming Camera Tuning

The paper introduces Q-CamTuning, a multimodal large language model designed for automatic camera tuning in livestreaming. It addresses limitations of existing IQA methods by enabling fine-grained quality comparison and providing reasoning-based adjustment suggestions. The authors built a new instruction tuning dataset, FGLive-Comments, and a progressive perception-evidence-guidance framework. Experiments show Q-CamTuning outperforms existing IQA models in quality scoring and pair ranking, with user studies confirming its practical suggestions.

arxiv2d agoScore 78

Multimodal Large Language ModelsImage Quality AssessmentCamera Tuning

Research Editor selected

Global pooling beats per-ship and adaptation models for ship fuel consumption prediction

A comparative study of 11 fuel oil consumption prediction methods on 21 vessels finds that global pooling—one model using ship identity and specifications—outperforms per-ship and adaptation-based approaches in all settings. Global LightGBM achieves 8.92% MAPE vs 11.14% for per-ship LightGBM, and global models of five architectures including a Transformer beat federated models. The global model also supports zero-shot prediction for new vessels, with accuracy comparable to vessel-specific adaptation after adding first records.

Why it matters: Provides a rigorous, fair comparison showing global models are superior for fleet-wide fuel prediction, with practical implications for voyage planning and new vessel deployment.

arxiv2d agoScore 88

ship fuel predictionglobal modelszero-shot prediction

Research

VideoEditor-MoE: Scaling Complex Video Editing with Mixture of Experts

VideoEditor-MoE is an open-source mixture-of-experts framework for complex video editing, introducing MoE-Attention with routed key-value experts to expand attention capacity. It scales from a 5B base model to 14B and 56B parameters with about 8B active per token, showing benefits over explicit expert guidance and MLP-only architectures. Analyses reveal that routing is driven by content and noise levels rather than human-defined categories, and decoupled weight decay stabilizes expert scaling.

arxiv2d agoScore 78

Video EditingMixture of ExpertsModel Scaling

Research

LLM Pipeline for Flexible Information Extraction from Criminal Court Decisions

This paper presents an LLM-based pipeline for structuring criminal verdicts into datasets for empirical legal research, using attribute exploration, span extraction, and value normalization guided by expert-refinable instructions. Evaluated on DUI and DDB traffic offenses, it shows that revising instructions improved extraction accuracy from 52% to 70% and increased human annotator agreement. The study also highlights that model consistency does not guarantee correctness)Skip, and disagreement between annotators and models can reveal unclear instructions.

arxiv2d agoScore 72

Large Language ModelsLegal Text MiningData Annotation

Research Editor selected

PRIME: Predictive Routing of Interventions for Scientific Ideation

PRIME is a plug-and-play decision layer that improves LLM-generated scientific ideas by routing feedback to the right revision scope. It addresses the diagnosis-intervention mismatch where the location of a weakness doesn't determine the best revision target. PRIME uses structural attribution, counterfactual routing, and intervention execution, showing gains across benchmarks and ideators.

Why it matters: This paper introduces a novel method to improve AI-driven scientific ideation, with concrete validation and clear performance gains, making it valuable for researchers and practitioners.

arxiv2d agoScore 82

AI researchscientific discoveryLLM evaluation

Research Editor selected

Networked AI: A New Architecture for Scalable, Adaptive Intelligence

This paper introduces Networked AI, a system of interacting, heterogeneous AI agents organized in a hierarchical layered architecture (Agent, Interaction, Regulation layers). It synthesizes data science, signal processing, and control theory to achieve near-linear scaling in resource demand, unlike exponential growth in monolithic systems. The framework enables online adaptation to novel environments and autonomous evolution toward application-specific intelligence, with applications in IoT, embodied, spatial, and industrial intelligence.

Why it matters: Presents a concrete architectural and theoretical framework addressing scalability and adaptability limitations of current AI, with broad application implications.

arxiv2d agoScore 85

AI architecturemulti-agent systemsscalability

Research Editor selected

Towards a Trustworthy and Governable Internet of Agentics: An Architectural Perspective

This perspective paper proposes an architectural framework for the Internet of Agentics (IoA), treating heterogeneous AI agents as governable service participants. It organizes the landscape into a three-layer progression from fragmented ecosystems to interoperable agents and trustworthy autonomy, covering protocol stacks (MCP, A2A, AGENTS.md), systems abstractions (orchestration, gateways, registries), and services computing principles. The paper argues that agentic systems reinterpret SOA and microservices concerns at the level of cognitive capabilities.

Why it matters: Provides a structured architectural view of emerging agent interoperability and governance standards, valuable for researchers and practitioners navigating the agent ecosystem.

arxiv2d agoScore 82

agent interoperabilityagent governanceservice-oriented architecture

Research Editor selected

PACERec: Progressive Adaptation of LLM Item Embeddings with Collaborative Evidence for Recommendation

PACERec is an ID-free framework that adapts static LLM item embeddings for sequential recommendation by learning a shared velocity field via frequency-weighted flow matching. It transfers interaction contexts from frequent to long-tail items, improving recommendation across the catalog. Experiments on three datasets show gains over baselines in overall and long-tail performance.

Why it matters: This paper presents a concrete, novel method addressing a known limitation of LLM embeddings in recommendation, with experimental validation.

arxiv2d agoScore 85

Large Language ModelsRecommendation SystemsEmbedding Adaptation

Research

MIRA: Reinforcing Multimodal Reasoning via Deceptive Contextual Augmentation

MIRA is a training framework that improves multimodal reasoning robustness by injecting deceptive visual contexts during training, creating 'hard traps' that stress-test models against misleading information. It also introduces a group-wise reflection-triggering mechanism for autonomous error detection and correction. Experiments show an average 7% improvement on in-domain benchmarks without human annotations.

arxiv2d agoScore 78

multimodal reasoningreinforcement learningrobustness

Research

Avatar Intelligence: A Paradigm for Persistent Personal Agents

This paper proposes Avatar Intelligence, a paradigm for persistent personal AI agents that represent users' goals, values, and constraints through sustained decision delegation. It introduces a maturity roadmap from human-directed execution to autonomous operation and an Agentic Oversight Index (AOI) for evaluating value against oversight burden. The authors identify three core research challenges: identity-continuous personalization, user-situated evolution via Rehearsal Engineering, and safety mechanisms combining technical control with accountability.

arxiv2d agoScore 78

AI agentshuman-AI interactionagentic oversight

Research

Hybrid Retrieval, NLI and Open-source LLMs for Scientific Claim Verification

This experimental study evaluates claim verification on scientific texts using a hybrid retrieval method (semantic + BM25) and compares an NLI baseline (DeBERTa) with three open-source LLMs (Llama 3.1, Qwen3, Mistral). Qwen3 achieves the highest Macro-F1 (0.7937), while the NLI baseline lags at 0.3157. A conservative fusion rule that requires agreement between NLI and LLM degrades performance, as correct LLM predictions are often overturned by NLI disagreements.

arxiv2d agoScore 78

scientific claim verificationnatural language inferencelarge language models

Research Editor selected

Fine-Grained Hierarchy and Benchmark for State Tracking in Language Models

This paper introduces a hierarchy of word problems for state tracking, from simple immutable assignments to general finite-state tracking, and proves that transformers cannot length-generalize on all aperiodic problems while Gated DeltaNet can. The authors build StateBench, a 4-tier benchmark, and show that both trained and pretrained models, including frontier ones, fail to robustly solve even soluble state tracking.

Why it matters: Provides a new theoretical hierarchy and benchmark that clarifies architectural differences in state tracking, with practical implications for model evaluation and development.

arxiv2d agoScore 85

state trackinglanguage model architectureslength generalization

Research Editor selected

LUCA: A Synthetic World for Evaluating Experimental Reasoning in AI Agents

LUCA is a configurable synthetic environment that tests AI agents' experimental reasoning by having them learn unknown regulatory systems through budgeted experimentsholistically. It separates evaluation of experiment selection, prediction, diagnosis, and intervention, revealing that these capabilities dissociate sharply across frontier LLMs. For example, diagnosis varies 7x across models while prediction varies only 1.4x, and agents often rescue systems while misdiagnosing faults. The framework exposes hidden failures like performing the right experiment but failing to use the evidence.

Why it matters: Provides a novel benchmark for dissecting AI agent reasoning processes, with concrete findings on capability dissociation and failure modes, valuable for AI safety and evaluation research.

arxiv2d agoScore 82

AI evaluationexperimental agentsLLM capabilities

Research Editor selected

Multimodal Prompt Learning with Dual-Granularity Contrastive Alignment for Preoperative Prediction of Lung Adenocarcinoma Invasiveness Using 3D CT

This paper presents a multimodal prompt-learning framework that combines 3D CT images with text prompts derived from clinical reports to predict lung adenocarcinoma invasiveness. The method uses a 3D ResNet-CBAM encoder and a frozen RadBERT with learnable prompts, plus global and local contrastive alignment modules. On a test set of 1,593 nodules, it achieved 87.92% accuracy, 87.78% F1-score, and 0.9499 AUC, significantly outperforming an image-only baseline. The framework requires only CT images at inference, preserving clinical usability.

Why it matters: The paper reports a concrete, recent AI development with strong quantitative results and a practical CT-only inference workflow, offering clear value for medical imaging and multimodal learning readers.

arxiv2d agoScore 82

Multimodal AIMedical Image AnalysisPrompt Learning