[see details] my areas of expertize

Alexander Sergeevich Panin

Abstract:

I am currently forbidden from using my name as an author to due to my NDA. Though, I have been a reviewer dozens of times and am eager to continue.

My areas of expertise are, in descending order:
* Parallel LLM inference (speculative and parallel decoding, parallel inference algorithms)
* LLM quantization and pruning
* Efficient training and inference, DL systems
* Reinforcement learning & Imitation learning
* Visualization or interpretation of learned representations

Below I provide evidence to back up my claims.

A) I have previously reviewed for 10+ conference iterations starting from 2019. As one example, see see "paper" PDF for certificate of appreciation for ICML20 reviews.

B) Below are examples of papers that I contributed to.
1. AQUA-KV, quantization for LLM KV caches that exploits inter-layer dependencies https://arxiv.org/abs/2501.19392v1
2. AQLM, data-aware vector quantization for LLM weights https://arxiv.org/abs/2401.06118
3. SWARM parallelism, collaborative pipeline-parallel training https://arxiv.org/abs/2301.11913
4. PETALS, collaborative LLM inference https://arxiv.org/abs/2209.01188
5. DeDLOC, data-parallel training with volunteer computing https://arxiv.org/abs/2106.10207
6. Beyond Vector Spaces: Compact Data Representation as Differentiable Weighted Graphs - https://arxiv.org/abs/1910.03524 NeurIPS19
7. Editable Neural Networks - https://arxiv.org/abs/2004.00345 - ICLR20
8. Neural Oblivious Decision Ensembles - https://arxiv.org/abs/1909.06312 - ICLR20
9. Sequence Generation with Unconstrained Generation Order - https://arxiv.org/abs/1911.00176 - NeurIPS19
10. Learning to route in Similarity Graphs - https://arxiv.org/abs/1905.10987 , ICML19

In papers (2, 4, 5, 6, 7, 8) I was offered equal contribution role with other first authors, but had to decline; in (1, 3, 9, 10) I played an auxiliary role. To verify my claims, you can either ask the co-authors (e.g. Artem Babenko) or you can download the source files from arxiv: papers (6,7,8,9) contain hedgehog-themed ASCII art in method.tex, which is my informal signature. Paper 2 has the same in main.tex.

C) I am an experienced Open-Source developer with contributions to many prominent deep learning frameworks (e.g. pytorch), see github.com/justheuristic

If that evidence is insufficient, I am happy to provide additional confirmations and corroborations from my collaborators.