Danial Sultanov

Research Assistant

ISSAI

The Institute of Smart Systems and Artificial Intelligence at Nazarbayev University.

Implemented a diffusion-based DOOM world model, built gameplay data collection, and reported 24.78 PSNR and 0.395 LPIPS.

• Implemented GameNGen paper on interactive world modeling for DOOM using diffusion models • Engineered data collection pipeline using PPO agent inside ViZDoom environment against 8 bots to generate various gameplay trajectories • Fine-tuned VAE on game’s HUD to reduce visual artifacts during autoregressive frame generation • Trained Stable Diffusion 1.4 on agent’s previous actions and observations without text conditioning, achieving 24.78 PSNR and 0.395 LPIPS compared to original frames

Research Assistant

ISSAI

The Institute of Smart Systems and Artificial Intelligence at Nazarbayev University.

Worked on tokenization, LLM pretraining and post-training; reports 40% lower perplexity loss and 28% accuracy on Kazakh MMLU.

• Developed morphological segmentation tool and custom tokenizer MorphBPE, reducing perplexity loss of LLMs by 40% compared to standard BPE • Curated different pretraining stages with ablation studies on 1B models on 3T tokens to determine ”good” data ratios for better capabilities of core LLM • Applied SFT on base Qwen2.5, Gemma 2 1B-4B models with post-training using GRPO and CoT distillation, achieving 28% accuracy on Kazakh MMLU with 4x A100 GPUs

Bachelor of Science, Mathematics

Nazarbayev University