Используем ИИ строго не по назначению.
Заметки про ИИ, IT, компьютерные игры, и всякие инженерные интересности.
Post #4880
3.08K
⚡️ Ежедневная подборка (с сегодняшнего дня лол) - четверг, 12.02.2026
🔬 ML PAPERS
🔥 FastFlow: 2.6x speedup for flow-matching (image/video gen), plug-and-play. ICLR 2026!
arxiv.org/abs/2602.11105 | github.com/Div290/FastFlow
🔥 DiNa-LRM: Diffusion-native reward model — preference optimization directly on noisy diffusion states. Beats VLMs at fraction of compute.
arxiv.org/abs/2602.11146
HairWeaver: Photorealistic hair animation from single image via sim-to-real video diffusion.
arxiv.org/abs/2602.11117
ViLaVT: "Chatting with images" — language-guided visual re-encoding. Strong on multi-image & video reasoning.
arxiv.org/abs/2602.11073
RLCER: Self-evolving rubrics for CoT reasoning. No human labels, beats outcome-only RLVR.
arxiv.org/abs/2602.10885
FormalJudge: 7B model detects deception from 72B agents (90%+ acc) via formal verification.
arxiv.org/abs/2602.11136
GameDevBench: 132 game dev tasks. Best agent only solves 54.5%.
arxiv.org/abs/2602.11103
arXiv.org FastFlow: Accelerating The Generative Flow Matching Models with... Flow-matching models deliver state-of-the-art fidelity in image and video generation, but the inherent sequential denoising process renders them slower. Existing acceleration methods like... 🔬 ML PAPERS
🔥 FastFlow: 2.6x speedup for flow-matching (image/video gen), plug-and-play. ICLR 2026!
arxiv.org/abs/2602.11105 | github.com/Div290/FastFlow
🔥 DiNa-LRM: Diffusion-native reward model — preference optimization directly on noisy diffusion states. Beats VLMs at fraction of compute.
arxiv.org/abs/2602.11146
HairWeaver: Photorealistic hair animation from single image via sim-to-real video diffusion.
arxiv.org/abs/2602.11117
ViLaVT: "Chatting with images" — language-guided visual re-encoding. Strong on multi-image & video reasoning.
arxiv.org/abs/2602.11073
RLCER: Self-evolving rubrics for CoT reasoning. No human labels, beats outcome-only RLVR.
arxiv.org/abs/2602.10885
FormalJudge: 7B model detects deception from 72B agents (90%+ acc) via formal verification.
arxiv.org/abs/2602.11136
GameDevBench: 132 game dev tasks. Best agent only solves 54.5%.
arxiv.org/abs/2602.11103
- 🔥 16
- ❤ 1









