Ali Abbasi | publications

conference papers

NeurIPS

ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs

Thrash, Chayne, Abbasi, Ali, and Kolouri, Soheil

In Advances in Neural Information Processing Systems (NeurIPS) 2026

Abs arXiv

LLM activation quantization learning optimized rotations that align activations with hypercube corners; 19.2× faster quantization than SpinQuant with lower perplexity across Llama families.
ICML

Zero Sum SVD: Balancing Loss Sensitivity for Low Rank LLM Compression

Abbasi, Ali, Thrash, Chayne, Qin, Haoran, Sharma, Shansita, Seifi, Sepehr, and Kolouri, Soheil

In International Conference on Machine Learning (ICML) 2026

Abs arXiv Website

Selects singular components across all layers with a zero-sum rule on first-order loss estimates; at 40% retention, 5.86× the uncompressed throughput and 16% lower WikiText-2 perplexity than SVD-LLM.

On-device demo: NVIDIA Jetson edge device

LLaMA-2-7B-Chat compressed with Zero-Sum SVD, keeping about 50% of the parameters in low-rank form with a different rank for each weight matrix and 8-bit quantization, running locally on an NVIDIA Jetson edge device.

Compared with the uncompressed model:

1.30×peak speedupat 1024 input tokens

20–23%latency reductionend to end

1.31×faster per tokentime per output token
ICLR

MCNC: Manifold-Constrained Reparameterization for Neural Compression

Thrash, Chayne, Abbasi, Ali, Andreas, Reed, Nooralinejad, Parsa, Koohpayegani, Soroush Abbasi, Pirsiavash, Hamed, and Kolouri, Soheil

In International Conference on Learning Representations (ICLR) 2025

Abs arXiv PDF

Constrains parameters to a low-dimensional nonlinear manifold generated by a frozen random network; +22.3 accuracy points over PLATON pruning at 50× compression (ViT-Ti, ImageNet-100).
CVPR

BrainWash: A Poisoning Attack to Forget in Continual Learning

Abbasi, Ali, Nooralinejad, Parsa, Pirsiavash, Hamed, and Kolouri, Soheil

In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2024

Abs arXiv PDF

Untargeted poisoning attack forcing forgetting in continual learners via model inversion of past data; up to 80% more forgetting than uniform noise.
ICCV

PRANC: Pseudo Random Networks for Compacting Deep Models

Nooralinejad, Parsa, Abbasi, Ali, Koohpayegani, Soroush Abbasi, Pourahmadi Meibodi, Kossar, Khan, Rana Muhammad Shahroz, Kolouri, Soheil, and Pirsiavash, Hamed

In IEEE/CVF International Conference on Computer Vision (ICCV) 2023

Abs arXiv PDF

Reparameterizes weights as a linear mixture of frozen random networks regenerated from a seed; +8.1 accuracy points over HashedNet at about 100× compression (ResNet-18, ImageNet-100).
IPRIA

Fast and Temporal Consistent Video Style Transfer

Abbasi, Ali, Toosi, Ramin, and Akhaee, Mohammad Ali

In 2021 5th International Conference on Pattern Recognition and Image Analysis (IPRIA) 2021

Abs HTML

Style transfer is a subset of image transformation problem in which the output is a combination of the style of a reference image and the content of an input image. Despite recent advances in image processing, video style transfer is still a challenging problem. One can implement such methods on a video, considering each frame independently. However, an unpleasant flickering effect will be observable in the output. Here, we propose a method to properly transfer a reference style to the input frames while preserving the temporal consistency. Thus, the flickering effect will be considerably mitigated. The proposed method is compared to two video style transfer methods using Sintel dataset. Results show that the proposed method keeps a better trade-off between temporal consistency and spatial losses.

preprints

arXiv

BASE: Batch-Aware Selection of Experts Using Predicted Removal Error for Efficient MoE Decoding

Abbasi, Ali, Shi, Justin, and Kolouri, Soheil

arXiv preprint arXiv:2609.36222 2026

Abs arXiv

Batch-aware MoE expert selection from predicted removal error, with custom GPU kernels; +29.5 accuracy points over the strongest baseline, or 60% faster than dense within 0.4 points (Qwen3-30B-A3B).
arXiv

IO-SVD: Input-Output Whitened SVD for Adaptive-Rank LLM Compression

Abbasi, Ali, Thrash, Chayne, Qin, Haoran, Pirsiavash, Hamed, and Kolouri, Soheil

arXiv preprint arXiv:2605.15626 2026

Abs arXiv

LLM/VLM compression via KL-aware input-output whitening and adaptive ranks; +10.2 avg. accuracy points over WSVD (SmolVLM-2B); 4.34× decode speedup with KV-cache compression (LLaMA-2-7B).
arXiv

Vector-Quantized Soft Label Compression for Dataset Distillation

Abbasi, Ali, Shahbazi, Ashkan, Pirsiavash, Hamed, and Kolouri, Soheil

arXiv preprint arXiv:2603.03808 2026

Abs arXiv

Compresses soft labels stored with distilled datasets via a vector-quantized autoencoder; 30–40× additional compression on ImageNet-1K while retaining over 90% of baseline accuracy.
arXiv

Teaching Networks to Solve Optimization Problems

Liu, Xinran, Lu, Yuzhe, Abbasi, Ali, Li, Meiyi, Mohammadi, Javad, and Kolouri, Soheil

arXiv preprint arXiv:2202.04104 2022
arXiv

Sparsity and Heterogeneous Dropout for Continual Learning in the Null Space of Neural Activations

Abbasi, Ali, Nooralinejad, Parsa, Braverman, Vladimir, Pirsiavash, Hamed, and Kolouri, Soheil

arXiv preprint arXiv:2203.06514 2022

workshop papers

CVPRW

One Category One Prompt: Dataset Distillation using Diffusion Models

Abbasi, Ali, Shahbazi, Ashkan, Pirsiavash, Hamed, and Kolouri, Soheil

In CVPR Workshops (SynData4CV) 2026

Abs arXiv

Distills each class into a learnable text prompt for a diffusion model; +12.6 accuracy points over RDED on ImageNet-1K at 50 images/class (ResNet-18, stored soft labels).
CVPRW

Diffusion-Augmented Coreset Expansion for Scalable Dataset Distillation

Abbasi, Ali, Imani, Shima, An, Chenyang, Mahalingam, Gayathri, Shrivastava, Harsh, Diesendruck, Maurice, Pirsiavash, Hamed, Sharma, Pramod, and Kolouri, Soheil

In CVPR Workshops (SynData4CV) 2026

Abs arXiv

Expands compact low-resolution coresets with uncertainty-guided diffusion augmentation; +10.1 accuracy points over RDED on ImageNet-1K at 10 images/class (ResNet-18).