Yedid Hoshen retweeted
With all the hype around weight space and this work, you may want to join the seminar I’m giving on Thursday on weight-space learning. I’ll discuss what exactly we found we can extract from weights in the pre Astra era (spoiler, a lot!)
lnkd.in/p/d8p8fDgW
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Yedid Hoshen retweeted
Did a diffusion model train on your image? 🕵️
The standard test flags images with low reconstruction loss. But some images are simply easier, so an easy unseen image can score lower than a difficult training image.
Excited to share our fix: calibrating across time 🧵 1/9
Thanks for pointing out the similarity between drifting and Implicit Maximum Likelihood Estimation!
I worked out the mathematical connection - the crux is that drifting fields are similar to the gradient of a soft version of the IMLE loss. So drifting is defined in terms of the gradient, whereas IMLE is defined in terms of the objective, but the behaviour should be similar. It's reminiscent of the formulation of classical mechanics vs. Lagrangian mechanics from physics.
One difference is that in drifting the weights on the positive samples and the negative samples are different, whereas they are the same in IMLE. It'd be interesting to see if the negative weights can be replaced with positive weights.
Cool new paper by @Goodeat258 and Kaiming's team!
arxiv.org/abs/2602.04770
Reminds me of @KL_Div's Implicit Maximum Likelihood Estimation paper
Yedid Hoshen retweeted
Check out our newest paper, led by @jonkahana: we find that the best performing models are often not the popular ones, and that there are many hidden gems that outperform their popular counterparts by a wide margin
🚨 New Paper Alert 🚨
We assume the "Wisdom of the Crowd" surfaces the best models, but are we actually finding the right ones? 🤔
Surprisingly, popularity is a poor proxy for performance and better options exist! 📉
Project: jonkahana.github.io/hidden_g…
👇 Here's what we found:
Yedid Hoshen retweeted
Presenting this morning at @NeurIPSConf - poster #1014 (11:00 am–2:00 pm)! 🚀
Come say hi and hear about why and how we should Chart an Atlas of All THE WORLD'S MODELS 🌍🤖
You'll probably spot our non-traditional poster from a distance 👀🗺️
#NeurIPS2025
Yedid Hoshen retweeted
We tested several unlearning methods and found none of them really erase knowledge from the model - they simply hide it! 🧐
What does this mean? We must tread carefully with unlearning research within diffusion models🚨
Here is what we learned 🧵👇(led by @kevinlu4588)
Excited to share our paper “When Are Concepts Erased from Diffusion Models?” at @NeurIPSConf!
We introduce two conceptual models for erasure mechanisms in diffusion models, and a suite of probes to recover supposedly forgotten concepts.
Project website: unerasing.baulab.info/
Well done!
If you are looking for a top-tier, visionary postdoc, you should hire Eliahu.
Honored to receive the Google PhD Fellowship in Machine Learning & ML Foundations for my work on weight-space learning and model populations. Thank you @GoogleAI & @Googleorg for the support!
I'm seeking postdoc roles for Fall 2026 and will be at #NeurIPS2025 — happy to connect.
2
9
853
Yedid Hoshen retweeted
Now to appear at #NeurIPS2025 as a position paper (<6% acceptance)!🎉
We advocate for systematically studying entire model populations via weight-space learning, and argue that this requires charting them in a Model Atlas.
@NeurIPSConf #NeurIPS
Replying to @karpathy
@karpathy Thanks for the inspiring talk (as always!). I'm the author of the Model Atlas. I'm delighted you liked our work, seeing the figure in your slides felt like an "achievement unlocked"🙌Would really appreciate a link to our work in your slides/tweet arxiv.org/abs/2503.10633
Yedid Hoshen retweeted
Excited to share this has now been accepted at #NeurIPS2025 as a position paper (<6% acceptance)!🎉
We advocate for systematically studying entire model populations via weight-space learning, and argue that this requires charting them in a Model Atlas.
@NeurIPSConf #NeurIPS
🧵👇

ALT Position overview: With millions of public models, it becomes important to move beyond individual models and study entire populations (left). The Model Atlas formalizes this shift by representing models as nodes in a graph, with directed edges denoting weight transformations (e.g., fine-tuning). Node size and color, as well as edge color, encode node and edge-level features; light blue indicates missing or unknown information. The atlas enables a range of applications, including model forensics, meta-ML research, and model discovery (center). In practice, most edges and features are unknown. This motivates ML methods that take models as input and infer their properties, thereby completing the missing atlas regions (right).
🚨 New paper alert! 🚨
Millions of neural networks now populate public repositories like Hugging Face 🤗, but most lack documentation. So, we decided to build an Atlas 🗺️
Project: horwitz.ai/model-atlas
Demo: huggingface.co/spaces/Eliahu…
🧵👇🏻 Here's what we found:
Yedid Hoshen retweeted
From being amazed by AI in Karoly's @twominutepapers videos some 5 years ago to having my own work covered - what a time to be alive!
A huge shoutout to my collaborators @amirhertz, @erictabellion , Alex Rav-Acha, Yael Pritch, Ariel Shamir, @YHoshen
This new AI is what Photoshop should have been. 🖼️✨ Turn lamps on/off, change light colors, remove sunlight - reflections, shadows, glass highlights all obey real physics. 5 sec per image. Magic! 🔥 Full video: youtu.be/YzGzCWydMh0
Yedid Hoshen retweeted
just learned that under certain conditions you actually CAN recover model weights *exactly*, post-LORA
pretty crazy
Yedid Hoshen retweeted
Replying to @karpathy
@karpathy Thanks for the inspiring talk (as always!). I'm the author of the Model Atlas. I'm delighted you liked our work, seeing the figure in your slides felt like an "achievement unlocked"🙌Would really appreciate a link to our work in your slides/tweet arxiv.org/abs/2503.10633
Nice - my AI startup school talk is now up! Chapters:
0:00 Imo fair to say that software is changing quite fundamentally again. LLMs are a new kind of computer, and you program them *in English*. Hence I think they are well deserving of a major version upgrade in terms of software.
6:06 LLMs have properties of utilities, of fabs, and of operating systems => New LLM OS, fabbed by labs, and distributed like utilities (for now). Many historical analogies apply - imo we are computing circa ~1960s.
14:39 LLM psychology: LLMs = "people spirits", stochastic simulations of people, where the simulator is an autoregressive Transformer. Since they are trained on human data, they have a kind of emergent psychology, and are simultaneously superhuman in some ways, but also fallible in many others. Given this, how do we productively work with them hand in hand?
Switching gears to opportunities...
18:16 LLMs are "people spirits" => can build partially autonomous products.
29:05 LLMs are programmed in English => make software highly accessible! (yes, vibe coding)
33:36 LLMs are new primary consumer/manipulator of digital information (adding to GUIs/humans and APIs/programs) => Build for agents!
Thank you again for the invite @ycombinator and congrats again on an awesome events! I'll post some links/references in the reply.
Yedid Hoshen retweeted
What if models could be the data🤔Find out at @iclr_conf #ICLR2025
Join the 1st workshop on Model Weights as a New Data Modality. We're training networks on model weights for a wide variety of tasks. Featuring an amazing lineup of papers & speakers🚀
🗓️Sunday 9-17
📍Topaz 220-225
Yedid Hoshen retweeted
🚨 Update Alert! 🚨
• ⚡ Much faster loading (seconds, not minutes!)
• ✨ Fresh layout for Large NLP Atlas
• 📖 Clear explanations of all features
Stay tuned—more improvements coming soon!
🚨 New paper alert! 🚨
Millions of neural networks now populate public repositories like Hugging Face 🤗, but most lack documentation. So, we decided to build an Atlas 🗺️
Project: horwitz.ai/model-atlas
Demo: huggingface.co/spaces/Eliahu…
🧵👇🏻 Here's what we found:
Yedid Hoshen retweeted
In our #ICLR2025 paper, we introduce WIND 🌬️
A method that embeds a distortion watermark directly in the diffusion noise! Our method ensures that the watermark in one image does not reveal information about the watermark in other images 🤫
📝 arxiv.org/abs/2412.04653
(1/5)
Yedid Hoshen retweeted
this paper is a visual art!
the atlas of🤗hugging face models studies 63,000 models published on @huggingface and they share some fun numbers:
> 96% of vision models are ≤1 node from foundation models, but 5% of NLP models reach ≥5 node depths
> only 0.15% of vision models use quantization
> 60% of HF's 1.5M models have no documentation
> HF hosts 17 petabytes of models, with 100,000+ models uploaded each month
Yedid Hoshen retweeted
Replying to @vanstriendaniel @huggingface
This is awesome! But what about the 700k+ models with no model cards?
Here's our stab at tackling this problem for models with no metadata 👀
Our #CVPR2025 paper
horwitz.ai/probex
And a fresh and exciting direction led by @jonkahana
jonkahana.github.io/probelog…
Yedid Hoshen retweeted
🚨 New Paper Alert on Zero-Shot Model Search 🚨
We present an efficient an accurate method to search and understand what models can do, based on their weights without access to any metadata!
Project Page 🌐: jonkahana.github.io/probelog…

































