LLMs usually rely on labeled data to improve But what if they could self-improve at test time — without any labels? Introducing TTRL: a new RL framework that boosts LLM reasoning by up to 159% on the AIME 2024 using just unlabeled test data Trending #1 on alphaXiv
RESEARCH
-

Agent Companion Research: Gulli et al. Framework Study
By
–
Agents Companion Gulli et al.: https://
kaggle.com/whitepaper-age
nt-companion
… #AGI #AIAgent #Agentic -

OpenAI’s Progressive 4o Update: A Departure from Discrete Releases
By
–
One of the weirdest parts of the 4o personality meltdown this week is the announcement that 4o will “continue to improve over the course of this week”. Have we ever seen an OpenAI model update announcement that is explicitly progressive and not discrete? @simonw shared the
-
Humorous exchange about ArXiv at ICLR OpenAI recruiting event
By
–
Nope what’s that? Isa Fulford (@isafulf) me at the iclr openai recruiting event: random man: have you heard of arxiv? — https://nitter.net/isafulf/status/1917065779432067545#m
→ View original post on X — @lilianweng, 2025-04-29 09:01 UTC
-

Astana International Forum 2025: Diplomacy Economics Sustainability Vision
By
–
When the most seasoned minds in diplomacy, economics, and sustainability come together, the real opportunity isn’t just in debate—it’s in building a shared vision for the world ahead, as AIF 2025 promises. Join > http://
astanainternationalforum.org Partnership @AstanaIntlForum #AIF2025 -

Oversight Limits: Can Weaker AI Control Smarter Systems?
By
–
Can a weaker AI really keep a smarter one in check?
New research puts scalable oversight to the test—and the results aren’t all comforting. Even layered oversight struggles once the gap gets too wide. Turns out, keeping super AI honest might be harder than we thought. -
Scaling Laws For Scalable Oversight
By
–
Scaling Laws For Scalable Oversight
Link: https://
arxiv.org/pdf/2504.18530 -

VM Cuts AI Latency 2.6x Without Manual Tuning
By
–
Why care? This VM slashes latency and boosts performance up to 2.6× faster than today's best tools—without hand-tuning.
More speed, less compute. That’s a win for AI at scale. Paper: https://
arxiv.org/pdf/2504.12984
v2
… -

Alibaba Qwen3: Eight Open Models with Advanced Thinking Capabilities
By
–
Alibaba’s Qwen just unveiled Qwen3: a family of eight open models ranging from 600M to 235B params.
— Rowan Cheung (@rowancheung) 29 avril 2025
— Flagship version rivals OpenAI o1 & DeepSeek-R1
— Hybrid “thinking” mode in all models
— Boosted coding + agent performance
— Supports 119 languagespic.twitter.com/eM5mw2TjIPAlibaba’s Qwen just unveiled Qwen3: a family of eight open models ranging from 600M to 235B params. — Flagship version rivals OpenAI o1 & DeepSeek-R1
— Hybrid “thinking” mode in all models
— Boosted coding + agent performance
— Supports 119 languages -
Alibaba Releases Eight Qwen3 Models Competing with OpenAI
By
–
AI NEWS: Alibaba just dropped eight open-weight Qwen3 models—and they're going head-to-head with the best from OpenAI and DeepSeek. Plus, more news from OpenAI, Luma Labs, Higgsfield AI, P-1 AI, and Figure. Here's everything you need to know:
