Principal RS at IBM Research AI. Speech, Formal/Natural Language Processing. Currently LLM post-training, structured SDG/RL. Opinions my own and non stationary
Manhattan, NY
Joined April 2019
- Tweets2.5K
- Following411
- Followers578
- Likes2.8K
Ramon Astudillo retweeted
3 weeks, and the cache climbed up from the bottom to the top ;) 🚀
Now, this is what I call upward mobility
arxiv.org/abs/2610.02383
Rumor has it frontier labs are using looped LMs to scale quality without scaling parameters. But KV-cache memory still grows with every loop. Is that necessary?
No! Our method BeNice! unlocks near-lossless sharing of the final KV-cache across loops: more depth, same KV memory.
This lossless KV-sharing isn’t guaranteed out of the box with looped models. 👀👇for our method BeNice! to teach your model that sharing is caring!
Ramon Astudillo retweeted
Cool work Hayden! We found similar conclusions about sharing is caring when focusing on Ouro/LT2 like pretraining: arxiv.org/abs/2610.02383 :)
I wonder: why at a later stage and not directly from scratch? For us, sharing the first recursion KV seemed to help during pretraining
I am going to be honest: I did not believe text-to-lora worked. Turns out it does!. Not only that, we show you can predict entire *LoRa distributions* given just the model input, and do inference scaling by sampling models!
Ramon Astudillo retweeted
I'll be at COLM 2026 this week, presenting Abstract Chain-of-Thought on Thursday afternoon!
If you're interested in latent reasoning, continual learning and self-evolving agents, test-time adaptation, and open-ended discovery, please reach out (my DMs are open)!
What if your language model could reason efficiently in an entirely new language?
We introduce Abstract Chain-of-Thought, a new mechanism which allows language models to reason through a short sequence of reserved "abstract" tokens through reinforcement learning. It is as performant as verbalized CoT at a fraction of the cost, achieving major gains in inference-time efficiency.
Ramon Astudillo retweeted
(1/n) 🎲 Bayesian nonparametrics for language modeling are back — this time, for memory and continual adaptation in reasoning!
We introduce Bayesian Cheatsheet: an adaptive, structured memory that organizes an LLM’s growing library of reasoning strategies to reuse generalizable insights from previous problems. Using nonparametric mixture models, strategies are structured for effective, domain-conditioned retrieval, achieving performance gains on various reasoning tasks with open-weight LLMs.
In the memory constrained world we are running into, these kinds of insights are key.
⚠️ Stop pretraining your looped Transformer with a separate KV cache per recursion!
In our recent preprint, we show that *memory sharing* not only saves memory but is also a net-positive inductive bias!
Less memory, same flops, higher quality.
🔗 arxiv.org/abs/2610.02383
🧵
Ramon Astudillo retweeted
Very excited to finally share my IBM internship work. It was an intense but rewarding summer (with very nice views), and I had a blast working with my talented collaborators at IBM & Cornell (@KeshavRamji @yelkurdi, @lastrasl @yoavartzi @nthngdy @RamonAstudill12)
Ramon Astudillo retweeted
Thanks to organizers and friends @mariotelfig @ferbatis @RamonAstudill12 @bgmartins @MarianaSCAlmei2 @pedrobalage Zita Marinho @gracaninja Ana Ayash!
Ramon Astudillo retweeted
Another year, another #LxMLS at @istecnico @Lisbon_ELLIS! Thanks to all the amazing speakers, monitors (incl. honorary monitor @LeonBottou in the picture), and students!
Ramon Astudillo retweeted
Transformers use the same hidden state for two competing roles: maintaining state and predicting the next token. 🔀
Splitting these roles into two separate streams outperforms standard Transformers.
Introducing the State-Prediction Separation (SPS) Transformer.
🧵
LxMLS 2026 is looking for monitors. As you know we favour ex-alumni so I don't need to explain to you how awesome this is. One week to apply!
lxmls.github.io/2026/call-fo…
Ramon Astudillo retweeted
Appreciate the engagement with our work and the discussion it's sparked! There are plenty of open problems to study around Abstract-CoT, and there are many areas for further analysis. We've written up a blog post that discusses some of our thoughts -- we plan to incorporate this (in some form) in a revised version of the paper as a Broader Impacts statement:
evanescent-broccoli-13d.noti…
What if your language model could reason efficiently in an entirely new language?
We introduce Abstract Chain-of-Thought, a new mechanism which allows language models to reason through a short sequence of reserved "abstract" tokens through reinforcement learning. It is as performant as verbalized CoT at a fraction of the cost, achieving major gains in inference-time efficiency.
Ramon Astudillo retweeted
🧵Building AI apps rarely means using just one model.
Granite 4.1 brings language, vision, speech, and guardrails together—so you can build real workflows, not just demos.
What devs should know.👇
Ramon Astudillo retweeted
"Thinking Without Words: Efficient Latent Reasoning with Abstract Chain-of-Thought"
Do reasoning models really need to think in words?
This paper replaces long verbal CoT with a short learned sequence of abstract tokens that acts like a latent scratchpad.
Warmed up from verbal CoT, then distilled and improved with RL.
Resulting up to 11.6x fewer reasoning tokens while staying competitive with standard CoT.
Ramon Astudillo retweeted
Paper: arxiv.org/abs/2604.22709
A big thanks to my collaborators, @TahiraNaseem12 and @RamonAstudill12!
Ramon Astudillo retweeted
What if your language model could reason efficiently in an entirely new language?
We introduce Abstract Chain-of-Thought, a new mechanism which allows language models to reason through a short sequence of reserved "abstract" tokens through reinforcement learning. It is as performant as verbalized CoT at a fraction of the cost, achieving major gains in inference-time efficiency.
Ramon Astudillo retweeted
Stop by our poster from 3:15-5:45 -- #616 in Pavilion 3!
I'll be at ICLR 2026 this week in Rio 🇧🇷, presenting two papers! If you're interested in self-improving agents, long-horizon RL and open-ended discovery, test-time training, latent reasoning, or just want to chat, please reach out!
📝 Learned Meta-Tokens for Language Modeling -- 4/24 @ 3:15 PM, Pavillion 3, Poster #616
📝 Learning Efficient Latent Reasoning with Abstract Chain-of-Thought -- 4/27 @ 3 PM, LIT Workshop, 101 A
Ramon Astudillo retweeted
I'll be at ICLR 2026 this week in Rio 🇧🇷, presenting two papers! If you're interested in self-improving agents, long-horizon RL and open-ended discovery, test-time training, latent reasoning, or just want to chat, please reach out!
📝 Learned Meta-Tokens for Language Modeling -- 4/24 @ 3:15 PM, Pavillion 3, Poster #616
📝 Learning Efficient Latent Reasoning with Abstract Chain-of-Thought -- 4/27 @ 3 PM, LIT Workshop, 101 A
Ramon Astudillo retweeted
Yeah! It could take a day, a week, a month, who knows, even, like worst case, 289 days, 22h, 27m, and 41 seconds.
Ramon Astudillo retweeted
Interested in merging LoRAs/models but feel like it's a black box? ⬛️
In our latest work, we finally define "Mergeability" and quantify it to understand the mechanics behind the magic.
The spoiler? It’s not random. It all comes down to the base model's prior knowledge. 🧠👇
Model merging is a game-changer for multitask learning without retraining. But why do some models merge perfectly while others crash? In our new paper we define and investigate Mergeability.
#LLMs #MachineLearning #NLProc #ModelMerging
