Robustness Frontiers: From LLMs to Robotic Perception

Latest 100 papers on robustness: Aug. 8, 2026

In the fast-evolving landscape of AI and Machine Learning, achieving robust and reliable system performance is paramount, especially as models are deployed in complex, real-world environments. This means going beyond impressive average accuracy to ensure models are resilient to unexpected inputs, subtle perturbations, and shifts in operating conditions. Recent research highlights a crucial shift towards understanding and engineering this robustness, delving into its fundamental mechanisms across diverse applications from large language models (LLMs) to robotic control and medical diagnostics. This digest explores several cutting-edge breakthroughs that are pushing these robustness frontiers.

The Big Idea(s) & Core Innovations

Enhancing LLM Reliability and Security

Several papers tackle the critical challenge of making LLMs more reliable and secure. The position paper “Position: It’s Time to Optimize LLMs for Self-Consistency” from MIT CSAIL and Goodfire AI argues that many LLM failures stem from evaluating single outputs in isolation. They propose self-consistency as a unifying framework, suggesting that optimizing for consistency across related inputs (paraphrases, counterfactuals) can address issues like sycophancy and factual inconsistency. Building on this, “Robust Context-Aware Detection of Malicious Instructions in Text” by Washington University in St. Louis introduces CAD (Context-Aware Detection), a lightweight classifier that uses query and context to detect indirect prompt injection attacks. Their work highlights that feature-space adversarial training can transfer robustness to actual language attacks, a crucial finding for hardening LLM agents. Further exposing vulnerabilities, “Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks” from Imperial College London reveals that even state-of-the-art models suffer significant faithfulness degradation (up to 50%) under meaning-preserving query perturbations in RAG systems. This underscores the fragility of LLM grounding. Addressing another aspect of LLM security, “Breaking Customized LLMs for Coding: Automated Red Teaming for Instruction Backdoor Attacks” by Nanjing University proposes ARIA, the first automated red-teaming framework that generates stealthy instruction backdoor attacks against customized coding LLMs, achieving high attack success rates while evading detection. Finally, “PURPOSE: Poisoning Conflict Resolution in RAG via Proxy-Fact-Grounded Updates” by Nanjing University demonstrates a black-box poisoning attack on RAG systems that bypasses conflict resolution by framing malicious injections as “updates” rather than contradictions, emphasizing the need for more sophisticated provenance checking.

Revolutionizing Robotic Perception and Control

Robotics research is making significant strides in robust perception and control. “GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions” from Tsinghua University and Tencent Robotics X introduces GeniWorld, an interactive world model that uses URDF-based visual action representations for robotic manipulation. This approach achieves remarkable zero-shot generalization to unseen environments and can synthesize diverse training data, significantly boosting downstream policy performance. For compliant manipulation, the University of Waterloo presents VIDP (Variable Impedance Diffusion Policy) in “VIDP: Variable Impedance Diffusion Policy for Compliant Robot Manipulation from Diverse Demonstrations”. VIDP learns variable impedance control from kinematic demonstrations without force sensors, dynamically modulating compliance to reduce interaction forces and improve success rates in contact-rich tasks. In navigation, “TRACE: Learned Proprioceptive Odometry for Legged Robots under Unreliable Contact Conditions” by KAIST introduces TRACE, an end-to-end learned proprioceptive odometry estimator for legged robots. It uses foot-aware cross-attention to adaptively weight sensor data, achieving up to 53.8% reduction in position drift on challenging terrains. For humanoid robots, “KILVO: Kinematic-Inertial-LiDAR-Visual Odometry with Robust Multimodal Adaptation for Humanoid Robots” from Harbin Institute of Technology presents KILVO, a multisensor fusion framework tightly coupling various data streams in an ESIKF, achieving 1 kHz state estimation and robust adaptation to sensor failures through seamless modality switching. Lastly, “DreamWAM: Beyond RGB Future Prediction for World Action Models” and “Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models” both from Huazhong University of Science and Technology, emphasize predicting future states not just as RGB images, but through complementary views of appearance, motion, geometry, and semantics. This future conditioning significantly improves robustness under visual distribution shifts, particularly for robot manipulation.

Resilient Perception and AI Systems

Beyond LLMs and robotics, research is enhancing the robustness of various AI systems. “UQ-Loc: Uncertainty-Aware LiDAR Scene Coordinate Regression” by Warsaw University of Technology improves LiDAR localization by predicting per-voxel anisotropic Gaussian covariance, significantly enhancing 6-DoF pose accuracy and providing well-calibrated uncertainty estimates. For robust image segmentation, “Universal Concept Disruption for SAM3 Image Segmentation” from the University of Science and Technology of China introduces Universal Concept Disruption (UCD), the first universal cross-concept adversarial attack targeting SAM3’s open-vocabulary segmentation, reducing mask AP from 59.43 to 18.73 with a single perturbation. Addressing medical imaging, “DistMedVL: Distributional Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation” by the University of Nottingham Ningbo China proposes DistMedVL, a probabilistic vision-language framework that models features as Gaussian distributions to handle aleatoric and epistemic uncertainty, achieving SOTA with superior data efficiency and cross-domain generalization. “MultiMoQ: Multi-Access Media-Over-QUIC for Robust Immersive Video Streaming” from The Hong Kong University of Science and Technology (Guangzhou) focuses on robust 360° video streaming, achieving 0.11% stall time vs. 29.1% for standard MoQ by separating pending-group maintenance from active transmission. In a critical area, “Enhancing Anomaly Resilience in Research Networks: A Large-Scale Forecasting Benchmark for Dynamic Security Baselining” by the University of Nebraska-Lincoln demonstrates that TiDE models reduce network traffic prediction errors by 30-42%, enabling dynamic security baselining that distinguishes scientific “elephant flows” from DDoS attacks. Finally, “SpikingNav: Robust Embodied Navigation with Spiking Neural Policies” from the Chinese Academy of Sciences introduces a spiking neural network framework for robust indoor embodied navigation that achieves stronger robustness under visual corruptions compared to ANN baselines while using significantly fewer parameters and FLOPs.

Under the Hood: Models, Datasets, & Benchmarks

Recent advancements are often powered by novel architectures, meticulously curated datasets, and rigorous benchmarks. Here’s a glimpse into the resources driving these innovations:

• GeniWorld: Leverages URDF-based rendering for visual actions, evaluated on the RoboTwin2.0 benchmark, and can use Wan2.2-TI2V-5B as a video generation backbone.
• UQ-Loc: Extends LightLoc architecture, validated on Oxford RobotCar and NCLT datasets with Expected Calibration Error (ECE) as a new metric.
• VIDP: Uses Task-Parameterized Directionality-Aware Mixture Model (TP-DAMM), with real-world validation on Franka Panda manipulator and UMI hardware setup.
• Is Self-Pretraining really useful to improve diagnosis in medical Time Series?: Utilizes Transformer models and evaluates on CAMARGO 2021 (gait biomechanics), PhysioNet Non-EEG Stress, and Parkinson’s Disease Gait datasets. Code is available at: https://github.com/omarcoser/SPT-Medical-Time-Series.
• MultiMoQ: Built on Media over QUIC (MoQ), evaluated in Mininet with 4K 30-fps 360° ERP video content. Code: https://github.com/YitongLI2000/MultiMoQ-ACMMM2026.git.
• Topometric Autonomous Vehicle Localization: Combines Visual Place Recognition (VPR) with Feed-Forward 3D (FF3D) models (e.g., DUSt3R, VGGT, DA3-Large) and MixVPR-512 descriptors, evaluated on COLD, 4Seasons, and RobotCar datasets.
• Iterate or Widen?: Focuses on LiDAR semantic scene completion with LMSCNet-SS backbone, rigorously evaluated on SemanticKITTI.
• Hybrid Machine Learning Framework for Herd-Level Cattle Growth Pattern: Employs a cascade GB→RF→NN architecture trained on automated step-on-off (SOO) weighing observations and environmental data from Australian Bureau of Meteorology.
• Clinical Communication Processing: Uses LLM-generated synthetic data sourced from MIMIC-IV-Ext, training fine-tuned encoder models.
• Universal Concept Disruption for SAM3 Image Segmentation: Targets SAM3 (and SAM3.1) foundation models, evaluated on SACo-Gold, LVIS, RefCOCO, PhraseCut, and OpenImages datasets.
• TRACE: Employs an end-to-end learned model trained in RaiSim simulation and fine-tuned on Raibo2 quadruped robot platform, with ground truth from FAST-LIO2 and Vicon.
• CohortHijack: Audits single-cell annotation robustness using PBMC3K and Paul15 datasets with CellTypist for majority voting. Code: https://github.com/arashVsh/CohortHijack.
• RepoOMP: A hybrid framework for OpenMP parallelization using a Multi-granularity Attributes Performance graph (MAP) and evaluated on NPB, BOTS, FFmpeg, NCNN, and GROMACS. Code: https://github.com/Qlalq/RepoOMP_Simplified.
• Once a Response, Always a Response: Introduces EchoPrompt, a training-free detector using instruction-tuned and base LLMs (Qwen2.5, Llama-3.2/3.1/3, Falcon-7B families) evaluated on DetectRL, RealDet, and RAID benchmarks.
• Closed-Loop Decision-Focused Learning for User-Aware Cloud Orchestration: Utilizes MTGNN for capacity prediction and GNeuro-PLS scheduling, evaluated on Microsoft Azure (2023-2025) and Alibaba (2018) trace datasets.
• Iterative Hybrid Discrete-Continuous Viewpoint Planning for UAV Photogrammetry: Uses CMA-ES optimization to plan UAV paths for 3D reconstruction, with models like Tu Duc’s Tomb and Church of St. Sophia.
• Wireless Linear Computation Broadcast: Formulates Gaussian MIMO broadcast channels and optimizes linear precoders/decoders.
• DistMedVL: Leverages Probabilistic Cross-Modal Adapter (PCM-Adapter) with Mahalanobis Alignment Module (MAM) and Distribution Flow Module (DFM), achieving SOTA on 8 medical segmentation benchmarks.
• F²Agent: A multimodal agentic trading system leveraging DeepSeek-R1 (Distill-Llama-8B), Qwen2.5-7B-Instruct, GPT-4o-mini, and financial APIs.
• Breaking Customized LLMs for Coding: Introduces ARIA, an automated red-teaming framework for instruction backdoor attacks against coding LLMs, evaluated on SALLM and CWEval benchmarks.
• KILVO: Multisensor fusion with an Error-State Iterated Kalman Filter (ESIKF), validated on custom humanoid robot SLAM dataset. Code: https://github.com/JixinGao/KILVO.
• MEC-Patch: A physics-grounded adversarial attack framework using Stefan-Boltzmann Law for visible-infrared cross-modal patches, validated on DroneVehicle, LLVIP, and VisDrone datasets.
• SCI-CLIP: A segment-centric inference framework for open-vocabulary segmentation using frozen CLIP models (ViT-B/16, ViT-L/14, ViT-H/14), DINO/DINOv2 features, and SAM2/Mask2Former/SegFormer/EoMT masks. Code: https://github.com/mzamini92/SCICLIP.
• Enhancing Anomaly Resilience in Research Networks: Benchmarks TiDE and PatchTST on a 57-day Internet2 backbone dataset. Code for processing and training to be released.
• Efficient higher-order multi-scale method: Uses FDM-FEM for numerical algorithm, applied to hygro-thermo-mechanical coupling.
• Different Perturbations, Different Mechanisms: Uses Llama-2-7B and Qwen2-7B with DialectBench, SIB-200, xSID4LR, and WikiANN datasets for CPT.
• A Second-Order Monolithic Scheme: Employs BDF2 scheme and total-pressure formulation for Stokes-Biot model.
• SCP-NL2TL: Augments NL2TL systems with conformal risk control and semantic consistency scores. Project materials available at https://sites.google.com/ucr.edu/scpnl2tl.
• Robust Context-Aware Detection of Malicious Instructions: Uses jina-embeddings-v3 encoder and DeepSeek-V4-Flash for paraphrasing, evaluated on AgentDojo, AgentDyn, and AutoDojo. Code: https://github.com/tavia-liu/CAD.
• School network reorganization: Uses ILP with Gurobi Optimizer and D-Wave’s constrained quadratic model on real-world data from the Calabria region, Italy.
• Robustness and User-Perceived Value of Popularity Calibration: User study on music recommendation with LFM-2B dataset.
• Improving Debugging in Verification-Aware Languages: Compares SNAP and CNTM techniques for fault localization in Dafny using DafnyBench. Code: https://anonymous.4open.science/r/DafnyCBT/2.
• Unified Planning-Learning Framework for Robust UUV Navigation: Combines Voronoi planning and RL local control trained via behavior tree distillation, validated in NVIDIA Isaac Sim.
• Constraint-First Reasoning: Introduces CFR prompting protocol, evaluated on AIME, CMIMC, BRUMO, AIMO_AMC math benchmarks.
• Universal Pathologies, Conditional Consequences: Triple-robustness analysis of RAG architectures on MuSiQue and DO-178C-style aerospace requirements.
• Sparse Mixture-of-Experts for Non-Uniform Noise Reduction in MRI Images: Uses ResNet18 embeddings and DnCNN experts, evaluated on BrainWeb and IXI datasets.
• Hierarchical proximal Galerkin: hp-FEM solver for variational problems, implemented in HierarchicalProximalGalerkin.jl Julia package.
• Revisiting Black-Box Model Ownership Verification: Proposes TOPMark based on top-k output probabilities, evaluated on CIFAR-10/100, Caltech-101, Purchase-100, AG-News, and 20 News datasets.
• MALT: Improves Muon optimizer with diagonal preconditioning, evaluated on GPT-2 Small/Medium/Large pretraining with OpenWebText dataset.
• VQ-VAD: Adapts Vector-Quantized GAN (VQ-GAN) for discrete motion representations, evaluated on CMU Panoptic, SHT, HR-SHT, NWPUC, and HuVAD datasets. Code: https://github.com/TeCSAR-UNCC/VQ-VAD.
• Promptable Animal Pose Tracking: Leverages DINOv3, Diffusion Hyperfeatures, CleanDIFT, and BioCLIP features, evaluated on APTv2 and TigDog benchmarks.
• Enhancing Low Back Pain Assessment: Introduces SpineSegDiff, a diffusion-based model for lumbar spine MRI segmentation, evaluated on SPIDER dataset. Code: https://gitlab.ethz.ch/BMDSlab/publications/low-back/diffusion-models-for-lumbar-spine-mri-segmentation.
• Evaluating the Diagnostic Robustness of Vision-Language Models: Uses Brain MRI dataset for GBM vs MET diagnosis, evaluates generalist and medically fine-tuned models.
• Training Crossroads for Recurrent Vision Transformers: Controlled study of single-block recurrent Vision Transformers (bViT) on CIFAR-100.
• Robust Control under Stationary Ambiguity: Evaluates policies trained under stationary ambiguity in hedging problems using Geometric Brownian motion and Heston stochastic volatility model.
• ContextWeave: A longitudinal benchmark for agent memory systems based on real-world workflows. Code: https://github.com/OpenMOSS/ContextWeave.
• MGSB: A regime-aware deep learning architecture using TT-RoughPath encoder and Mean-Teacher consistency regularization, evaluated on GPLA and GAS datasets.
• On the Effectiveness of Adaptation Strategies for VLM-Based Federated Learning: Compares adaptation strategies (FFT, encoder-specific fine-tuning, prompt learning, LoRA) for FL in remote sensing with BigEarthNet-S2, EuroSAT, and RESISC45 datasets using CLIP ViT-L/14. Code: https://git.tu-berlin.de/rsim/FL-RS-VLM.
• Explicit Language Memory for Long-Horizon Planning: Hierarchical VLM-VLA architecture with rolling natural-language memory for robotic planning.
• A Vision-based Control Framework for Real-time Autonomous UUV Operations: Combines FFT-based sparse depth estimation, TRUDepth, and wavemap for onboard UUV operation.
• The Sample Complexity of Distributionally Robust PAC Learning: Theoretical analysis of PAC learning under Cressie-Read divergences.
• Differential 6-DOF Pose Estimation: Differential pose estimator using inter-frame image displacements and 3D control points. Code: https://github.com/zyoungszu/pami2026.
• CSGen: Hierarchical multimodal diffusion model for curvilinear structure generation, using an SD3.5 backbone with ControlNet on a multi-domain dataset. Code: https://github.com/ShanZard/CSGen.
• Evaluating Theory of Mind in Reasoning Models: Evaluates GPT-5, Claude, DeepSeek R1, Grok-3-mini on Theory of Mind tasks. Code: https://github.com/ibm/de-haan-2025-evaluating-tom.
• Blockchain Empowered Trustworthy Agent Networks: Survey of multi-agent systems and blockchain-enabled trust mechanisms.
• Towards Robust Version Identification in the Wild: Introduces DiVers dataset (1.1M musical versions) for robust version identification, benchmarking DVINetX and CLEWS. Code: https://github.com/progsi/divers_dataset.
• ODRA: Synthesizes CBT sessions with Chain-of-Thought and dynamic patient resistance, fine-tuning Llama-3 and Qwen-3.5. Code: anonymous.4open.science/r/ODRA.
• Checked-In Secret Detection: Introduces StringGroup and Secretron for secret detection in code, evaluated on SecretBench and Google Play Android apps.
• Leak-Resistant Unlearning: New benchmark UNLINK-VL for cross-modal knowledge unlearning in VLMs. Code: https://leak-resistant.site.
• Emergence of Reputation-Based Cooperation in LLM Agents: Uses an indirect reciprocity donation game with four LLM backends.
• Robustness Emerges Early in Training Dynamics: Introduces Early-Phase Stabilization (EPS) and Asymmetric Weight Reversion (AWR), tested across CNN, ViT, Mamba architectures on various vision tasks.
• When Proxy Prediction Becomes Equation Reconstruction: Proposes RASPL, a formula-preserving residual framework for scientific machine learning, using RUSLE-derived soil-loss prediction.
• Towards Trustworthy Hypergraph Neural Networks under Label Noise: Introduces HyperTrust with entropy-aware hyperedge trustworthiness estimation, evaluated on multiple hypergraph datasets. Code: https://anonymous.4open.science/r/NoisyHGL-871D.
• Combating Knowledge Corruption in Agent Systems: Presents SecureCollaRAG, a Byzantine-tolerant RAG framework using GNN-based credibility scoring. Code: https://github.com/Aquarids/scr-pub.
• iStructTab: Multimodal architecture for image-tabular data fusion using Graph-Enhanced Descriptor Sequencing (GEDS) and Order-Aware Efficient Transformer with Memory Augmentation (OEMT), evaluated on six benchmarks. Code: https://github.com/zadid6pretam/iStructTab.
• Trident: Agentic LLM red teaming framework using Reinforcement Learning with Verifiable Rewards (RLVR) for training attacks against DRL cyber defenders, evaluated on CAGE 4 and CyberWheel. Code: https://anonymous.4open.science/r/Trident-A934.
• ArborEnum: Algorithm for enumerating decision tree Rashomon sets over continuous features. Code: https://github.com/zakk-h/ArborEnum.
• SIGNPOST-Bench: Counterfactual benchmark for evaluating text-vision conflict resolution in MLLMs, with 20 models evaluated on four datasets. Code: https://github.com/inorganicwriter/SIGNPOST-Bench.
• Physics-informed reduced-order modelling with equivariant spectral submanifolds: Introduces eSSM reduction, with a Python implementation available at https://github.com/GeorgAUT/eSSM.
• A Unified Model for Cross-Domain Clone Detection: Investigates model merging (Task Arithmetic, TIES, DARE-TIES, WUDI, PCB) and layer stitching across four code models on BigCloneBench, CLCDSA, and GPTCloneBench.
• Understanding Fault Tolerance of Adversarially Robust Pruned Models: Empirical investigation of pruning, adversarial training, and hardware fault injection on CNNs trained on MNIST.
• Open-World Darknet Traffic Recognition: Evaluates darknet traffic classifiers under leave-one-service-out evaluation across Tor, I2P, FreeNet, and ZeroNet with Darknet-dataset 2020.
• Radar4D-VLM: Radar-only temporal vision-language model processing 4D-radar point-cloud sweeps for autonomous driving, using K-Radar dataset.
• FBID: Adaptive personalized federated learning framework using server-side contextual multi-armed bandits and trust-based blending, evaluated on CICIoT2023 dataset.
• An Inline Control Architecture for Language Models in Intelligent Transportation Systems: Introduces Guarded-V2X, an inline semantic guardrail architecture for LLMs in V2X systems.
• Modality Agreement- and Conflict-Aware Prototype Hypergraph Learning: Proposes MACH framework using prototype-guided hypergraphs for multimodal intent recognition, evaluated on MIntRec, MIntRec2.0, and MELD-DA datasets. Code: https://github.com/csksuraj17/MACH.
• Beyond the QBER Threshold: Machine learning framework analyzing temporal QBER variations for multi-attack detection in BB84 QKD systems.
• Robust and Personalized Federated Learning for Aircraft-Engine Prognostics: Investigates federated learning for aircraft engine RUL prediction on NASA C-MAPSS turbofan benchmark. Code: https://github.com/robusfl/robust-personalized-fl-turbofan.
• SoK: How Frontier AI Reshapes System-Level Security Risk Dynamics: Systematization of Knowledge paper on Frontier AI and Critical Infrastructure security.
• Hardware-Enabled Fuzzy Inference: Survey of hardware acceleration for fuzzy inference systems across FPGA, ASIC/custom VLSI, and embedded IoT/TinyML platforms.
• C²MOE: Mixture of Experts framework for Incomplete Multimodal Emotion Learning, evaluated on CMU-MOSI and CMU-MOSEI datasets.
• Gaussian-LIC2: Real-time LiDAR-Inertial-Camera SLAM system using 3D Gaussian Splatting, evaluated on a specialized LiDAR-Inertial-Camera dataset. Code: https://xingxingzuo.github.io/gaussian_lic2.
• E2M: Generalization of EM algorithm for tensor-based discrete density estimation using double bounded α-divergence optimization.
• Robust Low-Tubal-Rank Tensor Completion: Proposes R-ItCUR for robust low-tubal-rank tensor completion under cross-concentrated sampling, using cardiac MRI and 3D seismic data.
• Low-Dimensional High-Leverage Subspace Optimization: Introduces Normalization Affine Preconditioning (NAP) for low-bit quantization, evaluated on ImageNet-1K, CIFAR-100, Cityscapes, and Qwen2.5-3B-Instruct.
• Operationally Feasible Synthetic Power-Grid Scenarios: Feasibility-aware hierarchical diffusion framework for generating power-grid scenarios, using PYPOWER and IEEE benchmark systems.
• Beyond Representational Similarity: Introduces Source-Conditioned Description-Length Gain (SCDG) for generative plagiarism detection, evaluated on PAN 2025/2026 benchmarks and Multi-News dataset.
• Does Forgetting Transfer Across Modalities?: Introduces UNLINK-VL, a real-world benchmark for cross-modal knowledge unlearning in VLMs, using Wikidata.
• KnowHal: Knowledge-driven benchmark for comprehensive multimodal hallucination evaluation, using 1,800 samples.

Impact & The Road Ahead

The collective thrust of this research is reshaping how we build and evaluate AI systems. The emphasis on understanding why models fail, rather than just that they fail, is crucial. Innovations like self-consistency for LLMs, physics-grounded adversarial attacks, and uncertainty-aware perception are moving us toward a future where AI is not just intelligent but also dependable. The transition from closed-world to open-world evaluation, especially in areas like darknet traffic classification, highlights the need for models that generalize robustly to entirely unseen conditions.

Looking ahead, several themes emerge. The development of robust mechanisms for handling distributional shifts and adversarial perturbations will remain critical, as seen in the work on LLM prompt injection, multimodal hallucination, and novel attack vectors for RAG systems. The integration of physical priors and structured knowledge into learning (e.g., GeniWorld’s URDF rendering, MEC-Patch’s emissivity laws, MGSB’s regime-conditioned gates) offers a promising pathway to building more interpretable and stable AI. Furthermore, the focus on efficient, parameter-free interventions and lower-cost architectures (e.g., NAP for quantization, spiking neural networks for navigation) will accelerate the deployment of robust AI in resource-constrained environments like edge devices and embedded systems. As AI continues to integrate into safety-critical domains like medical diagnostics and autonomous systems, the demand for truly robust and trustworthy solutions will only intensify. These breakthroughs lay a strong foundation for an AI future that is not just smarter, but also safer and more reliable.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.