Why can't smaller language models match larger ones on reasoning? Researchers from CUHK Shenzhen, Alibaba Group, and Westlake University introduce G²RPO-A. It adaptively feeds correct reasoning steps into training, dynamically adjusting guidance as the model improves. On math
@jiqizhixin
-
Former Qwen Lead Founds New $2 Billion AI Lab
By
–
Breaking News! Junyang Lin, @JustinLin610 ,former head of Qwen, has reportedly founded a new AI Lab valued at $2 billion. It is understood that he has already recruited several members with backgrounds from ByteDance, Tencent, and overseas institutions.
-
Introduction to Embedded Language Flows Research and Implementation
By
–
ELF: Embedded Language Flows Paper: https://
arxiv.org/abs/2605.10938
Code: https://
github.com/lillian039/ELF Our report: https://
mp.weixin.qq.com/s/7x8w_2Ov-lpS
EPRgqxLZBQ
… -
Kaiming He’s Team Introduces ELF for Continuous Embedding Language Modeling
By
–
Huge! Language models could generate text as smoothly as AI creates images now! Kaiming He's team introduces ELF (Embedded Language Flows). Instead of operating on discrete tokens, ELF stays in a continuous embedding space until the final step, borrowing proven techniques like
-
Subquadratic Unveils New SubQ Model Using Subquadratic Sparse Attention
By
–
What if LLMs could read a million tokens without exploding in cost? Subquadratic, an AI research startup, just unveiled SubQ — the first model built on fully subquadratic sparse attention (SSA). Instead of comparing every token pair, SSA routes attention only to the truly
-
Scal3R: New AI Research for Neural 3D Scene Reconstruction from Video
By
–
What if AI could reconstruct kilometer-scale 3D scenes from video as effortlessly as humans perceive a room? Researchers from Zhejiang University and Horizon Robotics present Scal3R. They introduce a neural context memory that compresses long-range scene info, adapting on the
-
DeepSeek-TUI: A Terminal-Based Coding Agent for DeepSeek Models
By
–
Hmbown/DeepSeek-TUI: Coding agent for DeepSeek models that runs in your terminal Project: https://
github.com/Hmbown/DeepSee
k-TUI/tree/main
… Our report: https://
mp.weixin.qq.com/s/A7ATOoYGBWwf
1dpV9GIevw
… -
Open-Source DeepSeek TUI AI Agent Released for Terminal Workflows
By
–
Hunter Bown (a.k.a. "Whale Bro") open-sourced DeepSeek TUI, a terminal-native AI agent built in Rust specifically for DeepSeek V4. It turns your terminal into an AI workstation where you chat, edit files, run shell commands, manage tasks, and even coordinate sub-agents—all with
-
PromptEcho: A New Zero-Annotation Reward Method for Text-to-Image Models
By
–
You could train text-to-image models without expensive human feedback! Alibaba Group and Zhejiang University researchers present PromptEcho—a reward method that uses a frozen vision-language model to measure image-prompt alignment directly, with zero annotations or extra
-
DeepSeek encounters a strange and concerning bug
By
–
DeepSeek has encountered a bug that is both strange and concerning.