@inproceedings{moazeni2026misattribution,
  title={Intrinsic-PAPR: Tackling Misattribution in 3D Intrinsic Decomposition via Proximity Attention Point Rendering},
  author={Alireza Moazeni and Shichong Peng and Yanshu Zhang and Chirag Vashist and Ke Li},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

TL;DRResolving misattribution in point-based 3D intrinsic decomposition
Intrinsic PAPR identifies and fixes the misattribution issue in point-based inverse rendering, where individual primitives learn incorrect appearance despite producing correct aggregated renders. Using proximity attention point rendering for direct per-point supervision, it enables accurate, view-consistent albedo and shading editing.

@inproceedings{zhang2026pcore,
  title={P-CORE: Self-Supervised Surface Consistency for Point-Based Neural Editing},
  author={Yanshu Zhang and Shichong Peng and Mehran Aghabozorgi and Alireza Moazeni and Ke Li},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

TL;DRSupervise surface prediction of deformed points by the deformed surface
P-CORE lets attention-based point representations undergo large non-rigid deformations without holes or tears by enforcing that the predicted surface stays consistent before and after random deformations, enabling zero-shot editing with substantially fewer artifacts.

@inproceedings{hosseinkhani2026piimle,
  title={$\pi$-IMLE: Fast Single-Step Action Generation for Generalist Vision-Language-Action Policies},
  author={Kian Hosseinkhani and George Shramko and Mehran Aghabozorgi and Qinhe Peng and Jianing Qian and Tristan Engst and Alireza Moazeni and Dinesh Jayaraman and Ke Li},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2026}
}

TL;DRSingle-step action generation for VLA policies via conditional IMLE
π-IMLE replaces the slow iterative flow-matching action head of vision-language-action policies with a single-step conditional IMLE generator, keeping the same VLM backbone while cutting inference to a single forward pass. It delivers a 3.67× inference speedup and the highest average success rate on the LIBERO benchmark, with 3.9–6.6× faster real-robot rollouts.

@inproceedings{aghabozorgi2026wimle,
  title={{WIMLE}: Uncertainty-Aware World Models with {IMLE} for Sample-Efficient Continuous Control},
  author={Mehran Aghabozorgi and Alireza Moazeni and Yanshu Zhang and Ke Li},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

TL;DRMulti-modal world models with uncertainty-aware policy learning for model-based RL
WIMLE learns stochastic, multi-modal world models using IMLE and weights synthetic transitions by predictive confidence, achieving state-of-the-art sample efficiency across 40 continuous-control tasks in DeepMind Control, HumanoidBench, and MyoSuite.

@inproceedings{zhang2023papr,
  title={PAPR: Proximity Attention Point Rendering},
  author={Yanshu Zhang and Shichong Peng and Alireza Moazeni and Ke Li},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

TL;DRReconstruct and render point clouds using attention
PAPR is a point-based surface representation that uses proximity attention to interpolate between nearby points to rays for rendering high-quality images, enabling non-volume-preserving geometry deformation by directly adjusting point positions, and, unlike 3D Gaussian Splatting, it avoids creating holes while preserving texture details after deformation.

@inproceedings{peng2022chimle,
  title={CHIMLE: Conditional Hierarchical IMLE for Multimodal Conditional Image Synthesis},
  author={Shichong Peng and Alireza Moazeni and Ke Li},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2022}
}

TL;DRHigh-fidelity multimodal conditional image synthesis with hierarchical IMLE
CHIMLE generates diverse, high-fidelity outputs for conditional image synthesis without the mode collapse of GANs. By introducing conditional hierarchical sampling into IMLE, it produces high-quality images with far fewer samples, improving FID by 36.9% on average over the prior best IMLE method across night-to-day, 16× super-resolution, colourization, and decompression.