Tiago Pimentel

Tiago Pimentel

@tpimentel.bsky.social

Postdoc at ETH. Formerly, PhD student at the University of Cambridge :)

2.2K followers132 following37 posts

Pinned

Tiago Pimentel
@tpimentel.bsky.social

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

Tiago Pimentel
@tpimentel.bsky.social

@craigschmidt.com has a second paper using LPs for tokenisation coming out today as well! Check it out: bsky.app/profile/crai...

Craig Schmidt@craigschmidt.com

arxiv.org/abs/2605.22705 arxiv.org/abs/2605.22821 Happy Linear Programming for Tokenization day! I was involved with two separate papers that hit ArXiv yesterday, using LP's to find the vocabulary maximizing compression, depending on the kind of inference you want to use.

arxiv.orgTokenization with Split TreesWe introduce Tokenization with Split Trees (ToaST), a subword tokenization method that directly optimizes compression under a new recursive inference procedure. ToaST greedily splits each pretoken int...

Tiago Pimentel reposted

Craig Schmidt
@craigschmidt.com

arxiv.org/abs/2605.22705 arxiv.org/abs/2605.22821 Happy Linear Programming for Tokenization day! I was involved with two separate papers that hit ArXiv yesterday, using LP's to find the vocabulary maximizing compression, depending on the kind of inference you want to use.

arxiv.orgTokenization with Split TreesWe introduce Tokenization with Split Trees (ToaST), a subword tokenization method that directly optimizes compression under a new recursive inference procedure. ToaST greedily splits each pretoken int...

Tiago Pimentel reposted

Violeta Kastreva
@vkastreva.bsky.social

Thrilled to share my first paper! 📄 We prove optimal tokenization is NP-hard on bounded alphabets (like bytes)—even unary for direct tokenization! Big thanks @tpimentel.bsky.social, @philipwitti.bsky.social & Dennis Komm for the mentorship! Best birthday gift. 🎂 arxiv.org/abs/2511.15709

Tiago Pimentel@tpimentel.bsky.social

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 unary strings like a, aa, aaa, ...? In our new paper, we show this is still hard, NP-hard!

Tiago Pimentel
@tpimentel.bsky.social

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 unary strings like a, aa, aaa, ...? In our new paper, we show this is still hard, NP-hard!

Tiago Pimentel
@tpimentel.bsky.social

More precisely, we show that: (i) for binary alphabets, not only finding an optimal tokeniser is NP-hard, but also finding arbitrarily good approximations; (ii) for unary alphabets, finding an optimal direct tokeniser is NP-hard!

Tiago Pimentel
@tpimentel.bsky.social

This was joint work with @vkastreva.bsky.social, @philipwitti.bsky.social, D. Komm! Violeta is a super smart student, who is definitely gonna do lots more interesting work :) It's her first paper, and it's also her birthday today 🥳 so follow her if you like this! Paper: arxiv.org/abs/2511.15709

arxiv.orgTokenisation over Bounded Alphabets is HardRecent works have shown that tokenisation is NP-complete. However, these works assume tokenisation is applied to inputs with unboundedly large alphabets -- an unrealistic assumption, given that in pra...

Tiago Pimentel reposted

Gunnar König
@gunnark.bsky.social

Interested in provable guarantees and fundamental limitations of XAI? Join us at the "Theory of Explainable AI" workshop Dec 2 in Copenhagen! @ellis.eu @euripsconf.bsky.social Speakers: @jessicahullman.bsky.social @doloresromerom.bsky.social @tpimentel.bsky.social Call for Contributions: Oct 15

sites.google.comTheory of XAI WorkshopExplainable AI (XAI) is now deployed across a wide range of settings, including high-stakes domains in which misleading explanations can cause real harm. For example, explanations are required by law ...

Tiago Pimentel reposted

Maria Ryskina
@mryskina.bsky.social

Interested in language models, brains, and concepts? Check out our COLM 2025 🔦 Spotlight paper! (And if you’re at COLM, come hear about it on Tuesday – sessions Spotlight 2 & Poster 2)!

Tiago Pimentel reposted

Alexander Hoyle
@alexanderhoyle.bsky.social

Accepted to EMNLP (and more to come 👀)! The camera ready version is now online---very happy with how this turned out arxiv.org/abs/2507.01234

Alexander Hoyle@alexanderhoyle.bsky.social

New preprint! Have you ever tried to cluster text embeddings from different sources, but the clusters just reproduce the sources? Or attempted to retrieve similar documents across multiple languages, and even multilingual embeddings return items in the same language? Turns out there's an easy fix🧵

Tiago Pimentel
@tpimentel.bsky.social

LLMs are trained to mimic a “true” distribution—their reducing cross-entropy then confirms they get closer to this target while training. Do similar models approach this target distribution in similar ways, though? 🤔 Not really! Our new paper studies this, finding 4-convergence phases in training 🧵

View full thread
Tiago Pimentel
@tpimentel.bsky.social

This project was done with Finlay and @kmahowald.bsky.social, and it is the outcome of Finlay's Bachelor's thesis! Catch him presenting it in #EMNLP2025 :) Paper: arxiv.org/abs/2509.26643 Code: github.com/Tr1ple-F/con...

arxiv.orgConvergence and Divergence of Language Models under Different Random SeedsIn this paper, we investigate the convergence of language models (LMs) trained under different random seeds, measuring convergence as the expected per-token Kullback--Leibler (KL) divergence across se...
Tiago Pimentel
@tpimentel.bsky.social

LLMs are trained to mimic a “true” distribution—their reducing cross-entropy then confirms they get closer to this target while training. Do similar models approach this target distribution in similar ways, though? 🤔 Not really! Our new paper studies this, finding 4-convergence phases in training 🧵

View full thread
Tiago Pimentel
@tpimentel.bsky.social

* A sharp-divergence phase, where models diverge as they start using context. * A slow-reconvergence phase, where predictions slowly become more similar again (especially in larger models).

Tiago Pimentel
@tpimentel.bsky.social

This means that: (1) LMs can get less similar to each other, even while they all get closer to the true distribution; and (2) larger models reconverge faster, while small ones may never reconverge.

Tiago Pimentel
@tpimentel.bsky.social

LLMs are trained to mimic a “true” distribution—their reducing cross-entropy then confirms they get closer to this target while training. Do similar models approach this target distribution in similar ways, though? 🤔 Not really! Our new paper studies this, finding 4-convergence phases in training 🧵

Tiago Pimentel
@tpimentel.bsky.social

In this paper, we define convergence as the similarity between outputs of LMs trained under different seeds, where similarity is measured as a per-token KL divergence. This lets us track whether models trained under identical settings, but different seeds, behave the same.

Tiago Pimentel
@tpimentel.bsky.social

Surprisingly, convergence isn’t monotonic. Instead, we find four convergence phases across model training. * A uniform phase, where all seeds output nearly-uniform distributions. * A sharp-convergence phase, where models align, largely due to unigram frequency learning.

Tiago Pimentel
@tpimentel.bsky.social

Very happy this paper got accepted to NeurIPS 2025 as a Spotlight! 😁 Main takeaway: In mechanistic interpretability, we need assumptions about how DNNs encode concepts in their representations (eg, the linear representation hypothesis). Without them, we can claim any DNN implements any algorithm!

Tiago Pimentel@tpimentel.bsky.social

Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵

Tiago Pimentel
@tpimentel.bsky.social

We are presenting this paper at #ACL2025 😁 Find us at poster session 4 (Wednesday morning, 11h~12h30) to learn more about tokenisation bias!

Tiago Pimentel@tpimentel.bsky.social

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

Tiago Pimentel
@tpimentel.bsky.social

@philipwitti.bsky.social will be presenting our paper "Tokenisation is NP-Complete" at #ACL2025 😁 Come to the language modelling 2 session (Wednesday morning, 9h~10h30) to learn more about how challenging tokenisation can be!

Tiago Pimentel@tpimentel.bsky.social

BPE is a greedy method to find a tokeniser which maximises compression! Why don't we try to find properly optimal tokenisers instead? Well, it seems this is a pretty difficult—in fact, NP-complete—problem!🤯 New paper + @philipwitti.bsky.social @gregorbachmann.bsky.social :) arxiv.org/abs/2412.15210

arxiv.orgTokenisation is NP-CompleteIn this work, we prove the NP-completeness of two variants of tokenisation, defined as the problem of compressing a dataset to at most $δ$ symbols by either finding a vocabulary directly (direct token...
Tiago Pimentel
@tpimentel.bsky.social

Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵

View full thread
Tiago Pimentel
@tpimentel.bsky.social

Overall, our results show that causal abstraction (and interventions) is not a silver bullet, as it relies on assumptions about how features are encoded in the DNNs. We then connect our results to the linear representation hypothesis and to older debates in the probing literature.

Tiago Pimentel
@tpimentel.bsky.social

Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵

View full thread
Tiago Pimentel
@tpimentel.bsky.social

Causal abstraction identifies this correspondence by finding subspaces in the DNN's hidden states which encode the algorithm’s hidden variables. Given such a map, we say the DNN implements the algorithm if the two behave identically under interventions.

Tiago Pimentel
@tpimentel.bsky.social

We show—both theoretically (under reasonable assumptions) and empirically (on real-world models)—that, if we allow variables to be encoded in arbitrarily complex subspaces of the DNN’s representations, any algorithm can be mapped to any model.

Tiago Pimentel
@tpimentel.bsky.social

Mechanistic interpretability often relies on *interventions* to study how DNNs work. Are these interventions enough to guarantee the features we find are not spurious? No!⚠️ In our new paper, we show many mech int methods implicitly rely on the linear representation hypothesis🧵

Tiago Pimentel
@tpimentel.bsky.social

In this new paper, w/ @denissutter.bsky.social , @jkminder.bsky.social, and T.Hofmann, we study *causal abstraction*, a formal specification of when a deep neural network (DNN) implements an algorithm. This is the framework behind, e.g., distributed alignment search. Paper: arxiv.org/abs/2507.08802

arxiv.orgThe Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?The concept of causal abstraction got recently popularised to demystify the opaque decision-making processes of machine learning models; in short, a neural network can be abstracted as a higher-level ...

Tiago Pimentel reposted

Musashi Hinck
@musashihi.bsky.social

Love this! Especially the explicit operationalization of what “bias” they are measuring via specifying the relevant counterfactual. Definitely an approach that more papers talking about effects can incorporate to better clarify what the phenomenon they are studying.

Tiago Pimentel@tpimentel.bsky.social

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

Tiago Pimentel
@tpimentel.bsky.social

If you use LLMs, tokenisation bias probably affects you: * Text generation: tokenisation bias ⇒ length bias 🤯 * Psycholinguistics: tokenisation bias ⇒ systematically biased surprisal estimates 🫠 * Interpretability: tokenisation bias ⇒ biased logits 🤔

Tiago Pimentel@tpimentel.bsky.social

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

Tiago Pimentel
@tpimentel.bsky.social

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our new ACL paper proposes an observational method to estimate this causal effect! Longer thread soon!

Tiago Pimentel
@tpimentel.bsky.social

For ACL, just replace your acl_natbib.bst file with the one in this repo 😁; for ICML, you need to replace `FUNCTION {format.lab.names}` in icml2025.bst with the code from lines 1655 to 1747 (untested, but should work 🙃)

Tiago Pimentel reposted

Ethan Gotlieb Wilcox
@wegotlieb.bsky.social

⭐🗣️New preprint out: 🗣️⭐ “Using Information Theory to Characterize Prosodic Typology: The Case of Tone, Pitch-Accent and Stress-Accent” with @cuiding.bsky.social , Giovanni Acampa, @tpimentel.bsky.social , @alexwarstadt.bsky.social ,Tamar Regev: arxiv.org/abs/2505.07659

arxiv.orgUsing Information Theory to Characterize Prosodic Typology: The Case of Tone, Pitch-Accent and Stress-AccentThis paper argues that the relationship between lexical identity and prosody -- one well-studied parameter of linguistic variation -- can be characterized using information theory. We predict that lan...
Tiago Pimentel
@tpimentel.bsky.social

Super happy we got this award for our paper on memorisation 😁🎉 congrats to the team and in particular to @pietrolesci.bsky.social, who led the project! Pietro is super smart, creative, hard-working, and on the job market -- you should hire him if you can :)

Cambridge Computer Science@cst.cam.ac.uk

🎉 Congratulations @pietrolesci.bsky.social, Clara Meister, Thomas Hofmann, @andreasvlachos.bsky.social & Tiago Pimentel! They won Publication of the Year at our annual Hall of Fame awards last week for their paper on 'Causal Estimation of Memorisation Profiles'. www.cst.cam.ac.uk/announcing-w...

Tiago Pimentel
@tpimentel.bsky.social

In this paper, we borrowed from the econometrics literature and proposed a method to estimate memorisation using only observational data! The paper is here, if you're curious: aclanthology.org/2024.acl-lon...

aclanthology.orgCausal Estimation of Memorisation ProfilesPietro Lesci, Clara Meister, Thomas Hofmann, Andreas Vlachos, Tiago Pimentel. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Tiago Pimentel reposted

Andreas Vlachos
@andreasvlachos.bsky.social

Honoured to receive this award! Tagging @pietrolesci.bsky.social and @tpimentel.bsky.social !

Cambridge Computer Science@cst.cam.ac.uk

🎉 Congratulations @pietrolesci.bsky.social, Clara Meister, Thomas Hofmann, @andreasvlachos.bsky.social & Tiago Pimentel! They won Publication of the Year at our annual Hall of Fame awards last week for their paper on 'Causal Estimation of Memorisation Profiles'. www.cst.cam.ac.uk/announcing-w...

Tiago Pimentel reposted

Kyle Mahowald
@kmahowald.bsky.social

I might be able to hire a postdoc for this fall in computational linguistics at UT Austin. Topics in the general LLM + cognitive space (particularly reasoning, chain of thought, LLMs + code) and LLM + linguistic space. If this could be of interest, feel free to get in touch!