Pinned Tweet
INTENT FACTORED FOR THE WIN!
Good science that works always finds a way
#ifftw we're going to NeurIPS.
Fresh, improved and updated IFG paper dropping soon 👀
Good science that works always finds a way
#ifftw we're going to NeurIPS.
Fresh, improved and updated IFG paper dropping soon 👀
Unlock real diversity in your LLM! 🚀
LLM outputs can be boring and repetitive. Today, we release Intent Factored Generation (IFG) to:
- Sample conceptually diverse outputs💡
- Improve performance on math and code reasoning tasks🤔
- Get more engaging conversational agents 🤖
3
2
19
1,705
Peeling away ideas like the fixation with being on-policy at all costs
There is a phase in life when you experience the thrill of learning things. And then, this second one when you figure out how to unlearn things, discarding beliefs with the same secret pleasure you get peeling scabs, picking your nose or squeezing blisters.
0
0
0
31
Struggled with this a few weeks ago, totally oblivious that all of this was happening. Overall great news!
Also Something funny about the failed training runs from off-policy sweeps looking similar
Also Something funny about the failed training runs from off-policy sweeps looking similar
Tinker's new StableReinforce objective is Reinforce + Score Centering. Here's how I reverse-engineered it 🧵
0
0
5
379
Uljad retweeted
It's about this formula:
E_q[C * ∇ logp] = E_q[C] * E_q[∇ logp] + Cov_q(C, ∇ logp)
E_q[C] * E_q[∇ logp] is the drift term. We want it to be 0
Score-Centering works on the 2nd term: E_q[∇ logp - E_q[∇ logp]] = 0
1. we can do MC, like thinky
2. that's not the only term
E_q[C * ∇ logp] = E_q[C] * E_q[∇ logp] + Cov_q(C, ∇ logp)
E_q[C] * E_q[∇ logp] is the drift term. We want it to be 0
Score-Centering works on the 2nd term: E_q[∇ logp - E_q[∇ logp]] = 0
1. we can do MC, like thinky
2. that's not the only term
0
2
59
5,336
Jordi and team are great to work with!
Hiring PhD student (London, 5 months) at Google DeepMind.
Topic: Steering multi-agent LLM economies via mechanisms for enabling reusable knowledge work.
Looking for strong phd students in Multi agent LLMs, MARL, or mechanism design.
Send cv to jordigrau@google.com
0
0
6
1,243
Uljad retweeted
Hiring PhD student (London, 5 months) at Google DeepMind.
Topic: Steering multi-agent LLM economies via mechanisms for enabling reusable knowledge work.
Looking for strong phd students in Multi agent LLMs, MARL, or mechanism design.
Send cv to jordigrau@google.com
Topic: Steering multi-agent LLM economies via mechanisms for enabling reusable knowledge work.
Looking for strong phd students in Multi agent LLMs, MARL, or mechanism design.
Send cv to jordigrau@google.com
9
16
298
46,525
Uljad retweeted
We are releasing our first model, Flint, a best-in-class local OSS coding model that runs at 150+ tokens/s on an @nvidia DGX Spark and 125+ tokens/s on a MacBook Pro M4—developed from scratch in less than four months with help from @Recursive_SI's autoresearch system.
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from hs
Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
24
29
320
20,016
Knowing when to stop is one of the most important questions of our time
A looped transformer has to know when to stop, and its state has to stay bounded while it loops.
We introduce Spherical Fixed-Point Reasoners (SFPRM) 🌐: loop on the unit sphere, where each step is an angle, and stop once the angle is small enough.
🧵 (1/9)
1
0
23
959
Uljad retweeted
The loop routine becomes 4 steps on the sphere: 1. Spherical linear interpolate (Slerp) a fraction toward the input 2. Apply the nGPT blocks 3. Rotate a fraction toward the proposal 4. Halt once the angle between state and proposal is below some threshold for every token. (3/9)
1
1
10
1,315
A groundbreaking Frontier result!
@maxencefaldor and @robertarail did a phenomenal job in showing the Argon capabilities so effectively. What a model and what a team in every sense of the word!
@maxencefaldor and @robertarail did a phenomenal job in showing the Argon capabilities so effectively. What a model and what a team in every sense of the word!
Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell. x.com/i/article/2111…
0
1
13
772
Uljad retweeted
Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell. x.com/i/article/2111…
57
68
633
277,445
Uljad retweeted
Four Gemini 4 Argon agents teamed up in Minecraft to defeat the Ender Dragon in 27 hours, writing every move as code and coordinating in the game's chat.
Great work led by @maxencefaldor with Satinder Singh.
Read more in our technical report:
arxiv.org/pdf/2610.22319.
Great work led by @maxencefaldor with Satinder Singh.
Read more in our technical report:
arxiv.org/pdf/2610.22319.
Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell. x.com/i/article/2111…
5
11
117
5,337
Uljad retweeted
Experts doubted a small neural net could beat Craftax without an LLM's knowledge.
Our agent meta-harness built one in 3 weeks. The 15M-param agent beats the final boss 44% of the time, costing a cent, and 10 seconds.
GPT-6 Astra won 1%, taking 24 hours and $83. 🧵
Our agent meta-harness built one in 3 weeks. The 15M-param agent beats the final boss 44% of the time, costing a cent, and 10 seconds.
GPT-6 Astra won 1%, taking 24 hours and $83. 🧵
33
75
734
273,304
Ben taught me so much about doing proper research in RL and the ways to debug policy gradients. Glad the Science of RL team and @reflection_ai gets to have this expertise for their models!
Being part of the Science of RL team for Beam has been an absolute privilege. It's been wonderful to learn how the big RL sausage gets made over the last year and a half and witness the dedication and pace of this team. I look forward to building many more models in the future!
1
1
15
940
This is great news! Congratulations to the team!
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam8G
0
0
10
210
Uljad retweeted
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam8G
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam8G
566
1,098
8,595
2,516,024