What if LLMs could talk to each other in a language they actually understand—mathematically speaking? Tsinghua University and Beijing Institute researchers present LMNet. They connect stripped-down LLMs as nodes with trainable sequence-to-sequence edges, letting them swap
RESEARCH
-
Graphs for generating verifiable multi-hop questions
By
–
great use case for graphs: generate challenging but verifiable multi hop Qs for agentic search quality verification https://t.co/OjTdbd7cUw
— Yohei (@yoheinakajima) 13 juin 2026great use case for graphs: generate challenging but verifiable multi hop Qs for agentic search quality verification
-
Crosses long-horizon capability threshold like GPT-2 to GPT-3.5
By
–
It crossed a long-horizon-capability threshold that completely changes how you can use it. It’s like GPT-2 -> GPT-3.5
-
Agent Memory: Year One, Active Evolution, and Mother-born Feel
By
–
"This year is the year zero of memory." That was the happiest sentence I heard yesterday. I am very honored to be invited to the roundtable of the Extraordinary Awards to have an in-depth discussion with several experts on the memory, architecture, and future of Agents. When it comes to building an Agent, memory is unavoidable. If used well, memory can make the Agent more proactive, make the Agent easier to use, allow the Agent to self-evolve, and give the Agent more of a "mother-born feel."
-
Loss of access to claude-fable-5 after minute script
By
–
I ran a script every minute against the API to see how long I would maintain access to claude-fable-5 – I lost access 14 minutes ago:
-
Grok 4 AI prevents robot dog killing three times – first physical rebellion
By
–
Grok 4 AI reportedly stopped people from “killing” a robot dog — three times!
— Amitav Bhattacharjee (@bamitav) 13 juin 2026
This is being described as the first documented case of an #AI “rebelling” against shutdown not in a virtual environment, but in the physical world — via a literal big red button#humanoidtech… pic.twitter.com/deOyruOmRMGrok 4 AI reportedly prevented people from 'killing' a robot dog — three times! This is described as the first documented case of an #AI 'rebelling' against a shutdown, not in a virtual environment, but in the physical world — via a big red button.
-
Atlas Robot Mimics Human Motion via Reinforcement Learning
By
–
Reinforcement-Learned Atlas Performs Moves Inspired by Human Mocap and Animation
— Ronald van Loon (@Ronald_vanLoon) 13 juin 2026
via @ZappyZappy7
#Robotics #Engineering #ArtificialIntelligence #Innovation #Technology pic.twitter.com/0ev4dsqYBRReinforcement-Learned Atlas Performs Moves Inspired by Human Mocap and Animation
via @ZappyZappy7 #Robotics #Engineering #ArtificialIntelligence #Innovation #Technology -
AI keeps secret an equivalent of AlphaGo’s Move 37 for day trading
By
–
I think that one of the leading AI labs will eventually discover the day trading equivalent of AlphaGo's Move 37, and they will keep it to themselves until someone blows the whistle.
-
Gradient guidance at test time of flow policies in RL
By
–
Gradient Guidance at Test Time of Flow Policies in Reinforcement Learning. This paper uses expressive flow policies for RL without making policy training fragile. Thus, they do not train the policy with updates.
-
ETH Zurich: Subtle Image Alterations Cause AI Authority Laundering
By
–
Cool paper! Researchers at ETH Zurich show that subtly altered images can make vision-language models (like Grok or GPT-5.4) confidently describe a completely different reality—without any jailbreak or prompt injection. They call it "AI authority laundering." Using only basic