This paper presents the NTIRE 2026 Cross-Domain Few-Shot Object Detection Challenge, showcasing innovative methods and pushing the performance frontier in a critical computer vision task.

Domain-specific autoencoders significantly improve medical image super-resolution fidelity in diffusion models, with code and weights available.

A systematic investigation into on-policy distillation for large language models, identifying key conditions for success and proposing strategies to improve failing distillations.

CoDe-R refines decompiler output using LLMs with rationale guidance and adaptive inference, achieving state-of-the-art re-executability for lightweight models.

Develop a cutting-edge agent-based multimodal search platform to enhance complex query resolution capabilities.

A survey and taxonomy of hallucinations in Video LLMs, analyzing root causes and proposing future research directions for more reliable systems.

A prompt learning framework that uses chain-of-thought prompts to predict missing entities, relations, and literals in knowledge graphs, outperforming traditional embedding models.

A benchmark and evaluator for LLM reasoning in coding tasks that improves accuracy and identifies limitations in existing methods.

A module that injects subcharacter compositional knowledge into Korean LLMs to improve linguistic understanding and generation without architectural changes.

A memory-augmented reinforcement learning agent for sample-efficient molecular optimization in drug discovery.

A new evaluation metric for LLMs that assesses reasoning quality beyond simple accuracy, with open-source code available.

A structured reasoning framework for multimodal LLMs to improve understanding of long documents by localizing and grounding evidence.

A novel neural compression technique for temporal lightmaps that enables high-quality dynamic global illumination with significantly reduced storage and memory.

A novel training acceleration method for vision foundation models that creates a model family chain, enabling efficient sequential knowledge transfer and reducing training costs by up to 72%.

A plug-and-play framework for reducing visual token redundancy in multimodal LLMs through class-adaptive layer fusion and dual-stage pruning.

A novel agent paradigm that grounds reasoning in deterministic computation before LLM generation for spatial-aware tasks, improving accuracy and interpretability.

A criticality-aware fine-tuning method that achieves robust Vision Transformer models by selectively updating only the most critical parameters, significantly reducing computational cost.

Generates ad personalization models using LLMs as hypernetworks, solving cold-start problems and deployed in production.

A framework that extracts LLM decision knowledge into an interpretable parametric model, enabling calibrated probabilities and quantitative human-AI collaboration.

This research provides evidence that agent identity induces attractor-like geometry in LLM activation space, offering a new way to understand and potentially control LLM behavior.

An LLM-powered multi-agent system that autonomously generates executable proofs-of-concept to validate bug reports in software, significantly improving bug detection accuracy and reducing false positives.

KG-Reasoner is an end-to-end framework using Reinforcement Learning to enable LLMs to perform dynamic, multi-hop reasoning over Knowledge Graphs.

An extension to Agent-to-Agent networks that enables modality-native routing for richer multimodal context, improving task accuracy in vision-dependent scenarios.

An offline distillation framework for large language models that significantly speeds up post-training for reasoning and code generation tasks without requiring a live teacher server.

A challenge and dataset for professional image quality assessment using multimodal large language models to provide comparative selection and expert-level reasoning.

An evolutionary algorithm enhanced with meta-learning and generative replay for robust and fast optimization of streaming data with concept drift.

A training-free framework that refines prompts for in-context image segmentation using gradient flow, significantly improving accuracy without additional training.

A novel security paradigm for AI agents that separates reasoning from execution to prevent unauthorized actions, with a 98.9% attack blocking rate.

A novel attack infers communication topologies in LLM-based multi-agent systems, revealing significant privacy risks and system vulnerabilities.

A unified loss function that improves classification accuracy and confidence calibration by leveraging uncertainty.

A pipeline for controllable molecular optimization using LLMs that preserves scaffold integrity and improves drug discovery efficiency.

MODIX is a training-free framework that scales positional indices in Vision-Language Models based on information density, improving multimodal reasoning.

GF-Score provides certified class-conditional robustness evaluation with fairness guarantees, enabling attack-free auditing of neural network vulnerabilities.

AutoDebug leverages LLMs to diagnose integration test failures, enhancing developer productivity at scale.

Adapt autonomous driving models to different vehicle dynamics with a physics-conditioned framework for improved transferability.

QuarkMedSearch is a long-horizon deep search agent for Chinese medical intelligence, achieving state-of-the-art performance with a novel data construction and training strategy.

A suite of LLM co-pilots and a unified dataset for translating natural language into formal models for optimization and satisfaction problems.

SEATrack is a multimodal tracker that uses AMG-LoRA and HMoE to achieve a balance between performance and efficiency in cross-modal fusion for tracking tasks.

A lightweight convolutional network for onboard satellite image restoration achieves competitive quality and significantly reduces latency, enabling real-time AI applications in space.

A lightweight method to control AI model access by conditioning activations on secret keys, without encrypting model weights.

A hint-assisted reasoning framework that uses cooperative small language models to improve mathematical problem-solving.

An end-to-end AI framework for real-time detection and characterization of cosmic radio signals, significantly reducing false positives and increasing speed.

WiseOWL is a methodology and Streamlit app for evaluating and recommending ontologies based on descriptiveness and semantic correctness.

A novel training methodology for LLMs that learns to generate constructive scientific paper feedback by leveraging author responses, significantly improving feedback quality and author perception.

A unified model for industrial anomaly detection, offering enhanced segmentation, understanding, and generation in manufacturing processes.

PAL is an AI platform that transforms lecture videos into interactive learning experiences, dynamically adapting to learners' understanding with personalized summaries.

ARGOS is a benchmark and framework for agentic multi-camera person search, enabling agents to reason, question, and utilize spatio-temporal tools under information asymmetry.

A graph-based soft prompting framework for multi-hop Knowledge Graph Question Answering that uses GNNs to reason over subgraphs and reduce sensitivity to incomplete KGs.

Human-centric topic modeling that uses LLM-based prompting and contrastive learning with optimal transport to produce goal-oriented topics.

A language-agnostic method for LLM safety that anchors alignment in the model's semantic bottleneck, significantly reducing attack success rates across languages.

A Gulf-focused multimodal dataset and agentic pipeline for climate decision support, integrating geospatial tools and domain-specific knowledge to improve LLM reliability.

BEAM is a bi-level evolutionary algorithm that designs high-level algorithmic structures for LLM-powered heuristic design, significantly outperforming existing methods in complex optimization problems.

A cross-lingual table understanding benchmark for Bahasa Indonesia documents, with fine-tuning insights and open-source models.

A zero-shot pipeline for frugal knowledge graph construction using local LLMs, achieving competitive results with consumer hardware.

An LLM-driven multi-agent framework that enables domain scientists to autonomously build high-quality deep learning surrogate models for subsurface flow simulations using natural language.

A minimalist conversational memory framework using only retrieval and generation, addressing signal sparsity and redundancy for robust long-term dialogue management.

Releases OpenTME, a large dataset of AI-generated tumor microenvironment profiles from histopathology images for research.

X-VC enables real-time zero-shot voice conversion to recreate any voice instantly using a neural codec.

A multi-agent deliberation system for LLM cascades that uses consensus-driven ensembles to resolve ambiguities internally, reducing costs and improving accuracy.

A new benchmark for self-evolving AI agents that iteratively optimize engineering designs using simulators and verifiers, pushing the boundaries of real-world problem-solving.

A new framework and dataset for autonomous driving that significantly improves navigation understanding by fusing global and local planning, achieving state-of-the-art results without auxiliary losses.

An open-source, efficient hybrid Mamba-Transformer model for agentic reasoning that significantly outperforms existing models in inference throughput.

A system for fine-grained simulation of cognitively impaired standardized patients using steering vectors and stochastic token modulation for precise severity control.

A novel strategy to detect and correct errors in medical image segmentation training data, improving model performance under challenging conditions.

Enabling smaller LLMs to predict their own performance limitations, paving the way for more efficient and self-aware AI systems.

A novel reinforcement learning paradigm for LLMs that bifurcates policy into normal and high-entropy modes to improve exploration without sacrificing accuracy.

An Opinion-Aware RAG architecture enhances LLM synthesis of subjective content by preserving opinion diversity, addressing limitations of current factual RAG systems.

MemJack is a memory-augmented multi-agent framework that uses visual semantics to orchestrate automated jailbreak attacks on VLMs, achieving high success rates and releasing a comprehensive benchmark dataset.

A benchmark and framework for evaluating the investigation depth of security incident response agents, distinguishing genuine forensic analysis from simple alert parroting.

A multimodal chunking pipeline that leverages document structure and vision to significantly improve RAG performance on long industrial documents.

A cognitive diagnostic framework for LLMs that moves beyond single scores to provide fine-grained ability assessments across multiple scientific domains, enabling targeted improvement and selection.

This paper reveals that multi-token prediction in Transformers enables more robust planning by inducing a reverse reasoning process, outperforming next-token prediction on various reasoning benchmarks.

Round-trip translation reveals limitations of current multilingual benchmarks and introduces 'Lost in Translation' (LiT) for realistic evaluation of multilingual LLMs.

A two-stage LLM framework that verifies and refines AI explanations for accuracy and accessibility, making XAI systems more trustworthy.

A novel reward mechanism for text-to-image models that eliminates the need for human annotation or reward model training, improving prompt following capabilities.

An intelligent framework automatically identifies optimal regions for vehicle counting in traffic videos, achieving near-perfect accuracy and four times faster processing.

A deterministic framework that enhances LLM text categorization accuracy and reproducibility by prioritizing high-value semantic features.

TRUST Agents is a multi-agent framework for explainable fake news detection and claim reasoning, offering improved interpretability and evidence transparency.

A training-free framework for efficient and temporally smooth video generation by dynamically allocating computation budget and using stochastic attention routing.

A new benchmark and entropy-guided search algorithm to enable LLM agents to execute long-horizon plans in large tool spaces.

A benchmark and data curation pipeline for evaluating deflection and hallucination in large vision-language models, focusing on retrieval-dependent samples and insufficient evidence scenarios.

A resilient framework for autonomous cars that uses sensor fusion and a novel correction algorithm to ensure accurate object distance estimation in real-time.

A training-free framework that mitigates multimodal LLM hallucinations by dynamically perturbing text to stabilize visual grounding.

A meta-learning framework that generates synthetic tasks to enable black-box optimization from small, offline datasets.

BID-LoRA, a parameter-efficient framework for Continual Learning and Unlearning, enabling precise knowledge deletion and efficient integration of new knowledge with minimal leakage.

This research systematically analyzes how well AI agents follow instructed plans, revealing critical insights for improving autonomous reasoning and task completion.

A framework for evaluating the robustness of LLM explanations in black-box enterprise systems, enabling better user trust and compliance.

A training-free framework for multi-style image transfer that uses mask-guided attention to seamlessly blend multiple styles without artifacts or structural inconsistencies.

A framework for improving visual preference optimization in multimodal AI using instance-specific rubrics, outperforming existing methods and approaching GPT-5.4 quality.

A foundation model for relational data that natively processes connected tables, outperforming supervised methods in few-shot learning and scaling to billion-scale datasets.

A hardware-efficient framework for LLM quantization that suppresses activation outliers to maintain accuracy and achieve significant speedups on modern AI accelerators.

MISID, a multimodal dataset and FRACTAM framework for complex intent recognition in strategic deception games, addressing deficiencies in current MLLMs for long-context discourse and cross-modal synergy.

A novel prompting schema that integrates expert system heuristics to improve LLM reasoning and problem-solving efficiency.

BayMOTH is a novel meta-Bayesian optimization approach that intelligently uses related-task information or falls back to lookahead for efficient sequential optimization.

A multi-agent framework for generating presentations from academic papers by reconstructing narrative flow, with a new benchmark.

A diagnostic benchmark and evaluation pipeline for understanding and improving LLM agent performance on long-horizon tasks.

A new benchmark and dataset for evaluating Video-LLMs in high-velocity esports environments, revealing significant performance gaps and guiding future development.

A physics-inspired neural operator efficiently synthesizes high-resolution multispectral images from panchromatic and low-resolution multispectral inputs.

A new method that simultaneously constrains LLM weights and activations to prevent safety degradation during fine-tuning.

A novel policy optimization method that jointly improves LLM reasoning accuracy and calibration, mitigating overconfidence and hallucination.

A novel time series forecasting framework that uses LLM-guided semantic asynchronous fusion to improve accuracy and generalization by decoupling semantic guidance from temporal dynamics.

A case-based learning framework for LLM agents that transfers prior task experience to new, complex real-world scenarios, improving structured analysis and performance.

An AI system that automates clinical research workflows, from planning to report generation, without requiring coding or direct patient data access.

A bidirectional attention classifier that improves LLMs' abstract meaning comprehension by mimicking human cognitive strategies.

A framework that adaptively generates biologically-informed features for decision trees, enabling interpretable and predictive DNA sequence analysis.

FastGrasp is a learning-based framework for fast, dexterous grasping with mobile manipulators, integrating grasp guidance, whole-body control, and tactile feedback for robust manipulation.

A framework for LLM agents that enforces organizational policies by simulating world-state changes, significantly outperforming content-only baselines.

Interactive machine unlearning for language models to empower users with personal data control.

A hierarchical spatial-aware algorithm with efficient reinforcement learning for human-robot task planning and allocation in production.

Develops a human-inspired multimodal memory system for social robots to enable personalized, context-aware interactions.

A proactive design approach for autonomous vehicles integrating redundancy and anomaly detection to ensure security and operational continuity against cyberattacks.

A new framework for evaluating LLM agents by profiling their execution-level behavior and refusal signals, crucial for safe organizational deployment.

A clustering-enhanced domain adaptation method for industrial control systems that significantly improves unknown attack detection with limited labeled data.

A novel LLM memory system using keyword bookmarks and a recall tool to dramatically improve long-conversation recall, outperforming existing methods.

An end-to-end LLM research loop autonomously reproduces, critiques, and extends computational physics papers, demonstrating significant potential for accelerating scientific discovery.

A lightweight sequential unlearning framework for LLMs that enables privacy-aligned deployment by selectively suppressing sensitive data without compromising general language capabilities.

LLM agents with dual-trace memory encoding significantly improve cross-session recall and temporal reasoning by pairing facts with contextual scene traces.

A novel deepfake detection method using Curvelet Transform and attention mechanisms to enhance feature robustness against compression.

Vector Relieved Flash Attention (VFA) optimizes attention computation by reducing vector operations, achieving significant speedups on modern hardware without performance loss.

Reveals that instruction-tuned LLMs are fragile to simple lexical constraints, leading to significant response collapse, and identifies a planning failure as the root cause.

A new benchmark framework that measures relational reasoning in LLMs by varying the complexity of entity binding, revealing consistent performance degradation.

A gold-supervision-free framework for training search agents using cycle-consistency to reconstruct questions from search trajectories.

Empirical assessment of LLM-generated code for construction safety reveals significant silent failure rates, highlighting the need for deterministic wrappers.

A fuzzing framework to find vulnerabilities in LLM chat templates for jailbreaking and red teaming.

Evaluates LLM and back-translation data augmentation for Hausa and Fongbe NLP, finding task-specific effectiveness for NER and POS tagging.

Extracting low-resource language data from LLMs using strategic prompting, with released corpora and code.

Integrates post-quantum homomorphic encryption into Llama 3 inference for privacy-preserving LLM applications with high accuracy and reasonable latency.

A survey of AI methods for modeling mixed automated and human traffic in simulation, identifying gaps and future directions.

Leveraging large-scale travel surveys and graph neural networks, this research uncovers nuanced patterns of urban social mixing and place exposure across five global cities, revealing mobility's greater influence than sociodemographics.

Analyzing architecture-dependent sample forgetting in fine-tuned image classifiers to inform data pruning and curriculum design.

A reproducible workflow for building and validating a legally screened chemistry corpus from S2ORC, enriched with embeddings and annotations for downstream text mining.

A hypothesis and framework for permutation-invariant table representation learning to build robust table retrieval systems.

Accelerating LLM inference with adaptive bounded self-speculation and layer-wise confidence calibration.

Reports on the first LLM Testing competition focused on benchmarking an LLM-based automotive assistant for car manual information retrieval failures.

An information-theoretic framework for task-adapted compressed sensing MRI that enables probabilistic inference for uncertainty prediction and adaptive sampling for clinical tasks.

Introduces LogicEval, a framework and dataset for evaluating automated repair techniques for logical vulnerabilities in real-world software, highlighting limitations of current LLM-based approaches.

This research introduces a thermodynamically consistent deep learning network for solar forecasting in off-grid microgrids, eliminating nocturnal generation anomalies and achieving zero-lag synchronization.

A measurement study of seven tactics to reduce cloud LLM token usage for coding agents, offering workload-specific strategies for cost savings.

A framework to infer and balance latent prognostic factors from real-world survival data to improve treatment-effect estimation.

A cognition-inspired dual-stream model that enhances dynamic emotion recognition by integrating semantic and contextual knowledge with facial dynamics, achieving state-of-the-art performance.

A longitudinal health agent framework designed for sustained user engagement and personalized decision-making in health tasks.

A new dataset and challenge for AI-accelerated photon and proton dose calculation in radiotherapy, enabling development of faster and more accurate dose estimation methods.

A multi-agent system that uses structured reflection notes to improve the factuality and coverage of image captions generated by large vision-language models.

A design for companion knowledge systems that proposes a governance profile for personal LLM memory wikis to prevent entrenchment and mirror user operational dimensions.

A self-distillation framework that improves model robustness and generalization by enforcing consistency along the optimization trajectory.

A bilevel optimization algorithm for the Electric Capacitated Vehicle Routing Problem that separates routing and charging decisions to accelerate convergence and achieve near-optimal solutions.

Designing reliable LLM-assisted rubric scoring for constructed responses in STEM exams, focusing on rubric design.

CycloneMAE is a scalable multi-task learning model for global tropical cyclone probabilistic forecasting that outperforms NWP systems.

This paper introduces Memory Worth, a lightweight primitive for agent memory governance that tracks memory success/failure co-occurrence to improve decision-making.

A framework that generates realistic facial expressions for training AI models to better perceive emotions from video.

Optimizing LLM prompts using evolutionary computation to enhance password guessing capabilities for security auditing.

Develops semantic image communication pipelines for edge traffic monitoring that drastically reduce data transmission costs by replacing full images with compact representations.

An AI system automates patient summary generation for thoracic tumor boards, improving efficiency and accuracy in clinical practice.

Introduces SOAR, a novel post-training method for diffusion models that corrects exposure bias and improves alignment without relying on reward models.

Evaluating protein sequence representations for Parkinson's disease classification shows limited discriminative power, requiring more informative biological features.

Investigating representation geometry in vision-language models for CT enterography, finding that mean pooling and per-slice contrast are key for disease assessment and retrieval.

An interactive interface for simulating citizen emotional responses to bureaucratic red tape across cultures, aiming to improve policymaking.

Improving audio source separation in reverberant environments using beta-divergence based nonnegative factorization.

An evaluation of AI tools' capabilities in transforming low-demand mathematics tasks, revealing moderate success rates and distinct generative versus classification abilities.

AISafetyBenchExplorer is a catalogue of AI safety benchmarks that reveals fragmented measurement and weak governance, providing a structured approach for benchmark discovery and meta-evaluation.

Optimizing deep learning models with elastic net regularization and Gabor dictionaries for improved heart sound signal classification.

This work shows that self-monitoring modules in reinforcement learning agents only provide benefits when structurally integrated into the decision-making pathway, not as auxiliary add-ons.

Investigating how latent planning abilities emerge and scale in Large Language Models through analysis of their internal representations.

A hybrid cognitive architecture using spiking neural networks to enable autonomous, emergent reasoning and LLM actions without explicit prompting.

Exploring contextual multi-task reinforcement learning for autonomous underwater reef monitoring to improve policy reusability and robustness.

A new metric for evaluating Natural Language to SQL systems that focuses on user intent rather than exact SQL syntax.

A hierarchical optimization framework using LLM-driven evolutionary search to efficiently tune GPU kernels across various dimensions.

Developing social learning strategies for virtual soft robots to accelerate brain optimization by leveraging peer experience.

A framework that transfers LLM knowledge into symbolic Tsetlin Machines for interpretable and accurate text classification without runtime LLM calls.

A novel method for continual machine unlearning that uses SVD-guided orthogonal subspace projection to prevent parameter collision and maintain model performance across sequential deletion requests.

A hierarchical graph-based memory framework for LLM agents that decouples memory encoding and consolidation to improve long-term coherence and reduce noise interference.

A theoretical framework for reducing redundancy in explainable clustering by formally characterizing and removing duplicate pattern representations.

Investigating how memory length impacts cooperative behavior in LLM agents within a social particle swarm model.

A framework for continuously updating a knowledge base and generating scientific hypotheses from evolving literature.

Safe reinforcement learning with online filtering for fatigue-predictive human-robot task planning and allocation in production.

Examining teacher interactions with a multi-agent system for personalized math problem generation, highlighting areas for improvement in authenticity and fit.

A novel neural network architecture designed for efficient low-bandwidth pipeline parallelism in large-scale decentralized training.

Aethon introduces a reference-based replication primitive for near-constant-time instantiation of stateful AI agents.

A biologically inspired digital twin of the visual hierarchy that generates contour sketches resembling ancient pictographs.

Investigates the Identifiable Victim Effect in LLMs, revealing how alignment and reasoning training modulate moral biases.

A framework for energy-efficient continual learning in HPC systems using agentic AI and brain-inspired architectures.

Theoretical analysis of loop corrections to training and generalization errors in random feature models.

Reinforcement learning refines robotic motion plans to ensure physical feasibility, bridging the gap between planning and real-world execution.

Guidelines and a formal model to maintain academic integrity in computing colleges amidst the rise of AI tools like ChatGPT.

This research demonstrates that current AI methods for detecting extraterrestrial life are prone to significant false positives due to their susceptibility to out-of-distribution samples.

A tutorial bridging operations research with deep learning for sequential decision-making under uncertainty.

This paper investigates methods for training behavioral dispositions into small language models, ultimately reporting negative results across multiple experimental arcs.

This paper analyzes resource sharing practices in underground forums for synthetic non-consensual intimate imagery creation and dissemination to identify intervention points for deterrence.

This paper provides theoretical guarantees for the convergence and robustness of Dense Associative Memory, a generalization of Hopfield networks, with potential applications in memory systems.

A theoretical generalization of conditional syntax splitting for nonmonotonic reasoning from conditional belief bases.

A theoretical framework for context-sensitive, multi-level similarity in First-Order Logic arguments, extending axiomatic foundations and parametric models.

This paper analyzes the efficiency of proportional mechanisms in online advertising auto-bidding, proposing a modified mechanism with improved price of anarchy bounds.

A review of how Large Language Models are being used to create AI agents for teaching and learning across different educational contexts.

Deepfakes are wrong because they usurp our authority over the permissible uses of our image and identity by exploiting biometric features as a generative resource.