Zhemal Khamidun
Head of AI at Alpina Digital, CPO of AlpinaGPT (~30-person engineering & AI team). Founder of Vibe Academy & Hamidun Consulting, author of the AI channel «Готовим ИИшницу» (@jemal_hamidun).
Stanford PhD Derives the Transformer From Scratch, Starting With One Memory Problem
A Stanford graduate's blog post derives the Transformer from one question: how to store a lot and retrieve any item. Nine years after the 2017 paper, the post is…
Zhejiang University and Peking University Propose SPADER for Multi-Answer QA
SPADER, accepted to EMNLP 2026 Main, is an RL method with step-wise GRPO and a diversity reward that trains models to find all answers, not just the most common…
Xiaohongshu and Fudan University Build DR-Rubric for RL on Open-Ended Tasks
Xiaohongshu and Fudan University present DR-Rubric at NeurIPS 2026, a way to build evaluation criteria for RL on tasks that have no reference answer.
Baidu, Tsinghua and Renmin Open-Source LoopVL, Teaching a Loop Transformer to See
GPT-6-Astra Maxes Out Robotics Benchmarks but Stumbles on Real Robots
The Information: Anthropic Is Preparing a Model Stronger Than Any Current Claude
Tencent Hunyuan and Fudan Launch ExplorationBench to Measure Exploration in LLMs
WorldMemBench Tests Whether World Models Remember Rooms and Events
The WorldMemBench benchmark scores world-model memory on two axes: whether previously seen scenes persist, and whether events keep progressing while the camera…
University of Pittsburgh Looks for a 3D Map Inside a VLM: How a Model Sees a Room
Haoming Wang and Wei Gao test whether vision-language models store a room's objects as a map of adjacency relations or just as a bag of colors and shapes.
Cohere Proposes Making Duplicate AI Agents Compete for a Place in the Company
ACEdit Teaches Image Editors to Train on Their Own Mistakes Without Degrading
MSR Moves Diversity Out of the LLM and Into a Corpus Schema: Five Tasks, Three Steps
Sparse Autoencoders: Pairwise Interactions Lift Rank Fidelity to 0.7790
LexFrame reads LLM hidden states: 0.97 coherence against a lens ceiling of 0.30
A new method reads hidden states through the model's own remaining layers. On SmolLM3-3B and OLMo3-7B it reaches 0.97 coherence, while fixed lenses top out at…
OmniXRef Combines Text, Photos and Video in a Single Video Editor
OmniXRef accepts any number of text, image and video references and transfers five types of attributes: camera, motion, content, style and VFX. The authors claim it…