Скачать или смотреть New AI Benchmark Record: Geometry Beats Arithmetic for Task Disentanglement

New AI Benchmark Record: Geometry Beats Arithmetic for Task Disentanglement

In this video, I break down the paper "Geometry Beats Arithmetic" and explain how we achieved a new benchmark record on Task Disentanglement in large language models (LLMs).

We explore why geometric methods outperform arithmetic approaches in representation learning and what this means for modern AI systems.

Topics covered:
Task Disentanglement in LLMs
Geometry vs Arithmetic in Representation Space
AI Benchmark Evaluation
Vector Space Analysis
Mechanistic Interpretability
arXiv Paper Breakdown
00:00 Introduction
01:30 What is Task Disentanglement?
05:10 Geometry vs Arithmetic
09:40 Benchmark Setup
14:20 Experimental Results
17:00 Why This Matters for LLMs

Full Colab implementation here: [link]

Paper discussed: https://arxiv.org/abs/2602.17385

What do you think — is geometry the future of LLM training?
Link to Colab Notebook: https://colab.research.google.com/dri...

Link to Google Doc: https://docs.google.com/document/d/1n...

Link to Arxiv Paper: https://arxiv.org/abs/2602.17385