←
Tiago Pimentel1,196 posts
Tiago Pimentel avatar

Tiago Pimentel

@tpimentelms

Postdoc at @ETH_en. Formerly, PhD student at @Cambridge_Uni.

◷ Joined November 2009▧ 83 Photos and videos
Tiago Pimentel avatar
Tiago Pimentel@tpimentelms·Feb 19

checkout our new paper about the superficial alignment hypothesis :) we use algorithmic information theory to formalise this hypothesis, we unify prior work on the topic, and show how post-training affects it! follow @tvergarabrowne for more great work like this in the future!

first paper of the phd 🥳 the Superficial Alignment Hypothesis (SAH) argues that pre-training adds most of the knowledge to a model, and post-training merely surfaces it. however, this hypothesis has lacked a precise definition. we fix this.

Post attachment
Yahya Emara avatar
⟳ Tiago Pimentel retweeted
Yahya Emara@yemara22·Nov 15

We think local inference is at an inflection point. Devices like the DGX Spark put 128GB of unified memory on a desk, and efficient open models are now strong enough to run useful agents. But today's engines are mostly optimized for datacenter GPUs or single chat local use, so they leave most of that hardware unused. We just released Hibana, a hardware-specific inference engine for local agentic workloads on unified-memory devices like the DGX/RTX Spark, our first step toward utilizing every kind of compute! On one Spark, vs. the fastest of vLLM, NVIDIA Flare, llama.cpp and TensorFold: 4.3× decode throughput 3.1× agentic tasks/hour 2.1× single-user speed Paper: arxiv.org/abs/2611.10469 Blog: sharar-project.github.io/blo… Repo: github.com/sharar-project/hi… 🧵 below Work with @CCCCC1009CCC , @yueh_ch , @mohsaied and others!

0:28
Sasha Boguraev avatar
⟳ Tiago Pimentel retweeted
Sasha Boguraev@SashaBoguraev·Nov 10

Excited to release a new draft handbook chapter on Language Models and Experimental Syntax! We review the last ~10 years of work using experimental syntax methods (minimal pairs, acceptability judgments, priming) to ask if LMs can do syntax w/@qyao23,@kanishkamisra & @kmahowald

Post attachment
Naomi Saphra avatar
⟳ Tiago Pimentel retweeted
Naomi Saphra@nsaphra·Nov 6

The saddest thing isn't my own slop reviews, it's how negative human reviews are getting for truly novel work. Good faith reviewers are fatigued by vapid slop papers, and can't gather the energy to attempt to understand a new idea.

A paper I was very excited about (might be the most practically impactful work I've done) got 2/2/3/2. Pangram flagged 3/4 as 100% AI generated. (The shortest is short and mostly AI generated.) I hate that we likely can't afford to ban these reviewers from refereeing.

Junghyun Min avatar
⟳ Tiago Pimentel retweeted
Junghyun Min@Aatlantise·Nov 8

1/6 🧵 How much does the “melody” of speech tell you about syntax? At #EMNLP2026 we measure it with information theory: word duration and pauses cut syntactic uncertainty by up to 10.2% in spontaneous conversational speech. 📄 aclanthology.org/2026.emnlp-…

Post attachment
Bob West avatar
⟳ Tiago Pimentel retweeted
Bob West@cervisiarius·Nov 5

🧩 Just for kicks, I made a daily word puzzle: #Emoglyphic. You see a short text in a mystery language you probably don't speak. Can you work out what the highlighted word means? Emoji clues help, but the fewer you use, the better. Post your score 👇

Post attachment
John Hewitt avatar
⟳ Tiago Pimentel retweeted
John Hewitt@johnhewtt·Nov 4

A single direction in LLM activation space, applied as a steering vector with four different norms, can cause the language model to perform a different behavior at each norm. This and more in a new blog post, Do Not Rely on the Linear Representation Hypothesis. (link below.)

Post attachment
Violeta Kastreva@EMNLP 2026 avatar
⟳ Tiago Pimentel retweeted
Violeta Kastreva@EMNLP 2026@VKastreva·Nov 3

Hardness proofs are more fun when you can poke them 👇 Our #ICLR2026 paper now has an interactive page where you can play with the SAT and vertex-cover reductions behind the hardness results.

Tokenisers are a vital part of LLMs, but how hard is it to find an optimal one? 🤔 Considering arbitrarily large alphabets, prior work showed this is NP-hard. But what if we use bytes instead? Or strings like a, aa, aaa, ...? In our new paper, we show this is still hard, NP-hard!

Post attachment
Julie Kallini ✨ avatar
⟳ Tiago Pimentel retweeted
Julie Kallini ✨@JulieKallini·Oct 28

If you're at #EMNLP2026 you should stop by our ✨tokenization tutorial✨ happening in <2 hours, starting at 9AM CET! Fun fact: I'm actually not at EMNLP... I'll be presenting remotely. But I'll be up from 1AM to 4AM PDT because that's how much I love tokenization 🫶😄

Headed to #EMNLP2026? Want to learn about the latest advances in tokenization research in LLMs? Come check out our tutorial on October 28. We will cover everything from fundamentals and emerging approaches with plenty of hands on demos and exercises. See you in Budapest!

Post attachment
levent avatar
⟳ Tiago Pimentel retweeted
levent@__alpoge__·Oct 27

“The core algorithm was found on October 16 by an internal Claude research model with 14M output tokens and no human input. Anthropic researchers worked with Claude to improve the presentation, resulting in this paper.” github.com/anthropics/formal…

Tiago Pimentel avatar
Tiago Pimentel@tpimentelms·Oct 26

LM-generated results just keep getting more and more surprising 👀 Impressive if true!

An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4

Tiago Pimentel avatar
Tiago Pimentel@tpimentelms·Oct 26

Hey :) I'm looking for emergency reviewers for #ICLR2027! The papers are on LLM agent adaptation, LLM evaluation, and interpretability. If you're interested in reviewing, please DM me!