There is now a smarter way to pick data for training LLMs! Enter OPUS! This is an ICML Oral paper from SJTU, Alibaba, UW–Madison, UIUC, and Mila – Quebec AI Institute. The proposed method dynamically and intelligently selects the most impactful data for LLM pre-training in
@jiqizhixin
-
SenseNova-U1 unifies understanding and generation as one process
By
–
What if understanding and generation were the same process? SenseNova introduces SenseNova-U1 — a native unified model that treats seeing and creating as a single process. It matches top understanding-only VLMs in text, vision-language perception, reasoning, agents, and
-
Data Value Density framework boosts AI learning from less data
By
–
What if AI could learn more from less data? Researchers from Shanghai Jiao Tong University and Shanghai AI Lab introduce 'Data Value Density (DVD) enhancement' — a unified framework to make every training token count. Instead of just piling on more internet data, DVD methods
-
FASTER: Reacting Quickly to Changing Environments
By
–
Why can't robots react instantly to fast-changing environments? Researchers from HKU and ACE Robotics introduce FASTER. Instead of running all sampling steps before any movement, it uses a Horizon-Aware Schedule to compress the immediate action into a single denoising step.
-
I²B-LPO: Branching AI Reasoning at Confusion Points
By
–
Are your AI models stuck in a rut, repeating the same narrow reasoning? Alibaba's DAMO Academy and partners (USTC, SJTU, Zhejiang, Northeastern) introduce I²B-LPO. Instead of randomly tweaking words, it branches reasoning at key confusion points and uses an information
-
On-policy distillation needs teacher sharing thinking and offering novelty
By
–
What if the best teacher for a language model isn’t the strongest one? Researchers from Tsinghua University and collaborators show that on-policy distillation (OPD) succeeds only when student and teacher share thinking patterns and the teacher offers genuinely new
-
Autonomous Preference Optimization transforms AI model disagreements into constraints
By
–
What if shifting reasoning patterns from multiple AI models could be turned into constraints instead of noise? Researchers from UTS’s Australian AI Institute (AAII) introduce Autonomous Preference Optimization (APO). Their approach treats disagreements between models as dynamic
-
Kuaishou OneSearch-V2: Generative Search That Understands Intent
By
–
What if your search engine could understand not just what you type, but what you really mean? Kuaishou Technology presents OneSearch-V2: a generative search framework that reasons like a human before returning results. It uses three clever tricks: (1) a “thought” step to deeply
-
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
By
–
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention Paper: https://
github.com/NVlabs/GatedDe
ltaNet-2/blob/main/paper/GDN2_paper.pdf
…
Code: https://
github.com/NVlabs/GatedDe
ltaNet-2
… -
DreamLite: ByteDance’s compact 0.39B model for fast image generation and editing
By
–
What if your phone could generate or edit images in under a second? ByteDance’s Intelligent Creation Lab presents DreamLite, a compact 0.39B parameter model that unifies text-to-image generation and editing in one network. It uses a simple trick: concatenating images