I'm very happy to give a spotlight at the Mechanistic Interpretability Workshop @ ICML on our work: "Validating Causal Abstraction Metrics on Simulated Complex Systems" Which metrics actually tell you if an explanation is valid? We built a benchmark to find out. 1/n

Effective language identification based on a tokenizer UnigramLM tokenizer already gives probabilities, testing those to identify a language is fast and effective. Whiceh leads me to wonder, can we identify language during training and affect behavior? arxiv.org/abs/2602.17655 @tpimentel.bsky.social

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.

Tokenization with Split Trees

We 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...

arxiv.org

Our new paper reformulates tokenisation as a linear program (LP), which we solve to get SOTA tokenisers 😁 As a bonus, this LP tells us how close to optimal any tokeniser is! Check it out 👇 w/ J. Tempus, @philipwitti.bsky.social, @craigschmidt.com, D. Komm Paper: arxiv.org/abs/2605.22821

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 · 9mo ago

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!

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!

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

Theory of XAI Workshop

Explainable 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 ...

sites.google.com

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)!

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 · last yr.

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🧵

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 🧵

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 · last yr.

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🧵

Honoured to receive two (!!) SAC highlights awards at #ACL2025 😁 (Conveniently placed on the same slide!) With the amazing: @philipwitti.bsky.social, @gregorbachmann.bsky.social and @wegotlieb.bsky.social, @cuiding.bsky.social, Giovanni Acampa, @alexwarstadt.bsky.social, @tamaregev.bsky.social

ACL@aclmeeting.bsky.social · last yr.

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 · last yr.

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!

@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 · 2y ago

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

Headed to Vienna for #ACL2025 to present our tokenisation bias paper and co-organise the L2M2 workshop on memorisation in language models. Reach out to chat about tokenisation, memorisation, and all things pre-training (esp. data-related topics)!

Pietro Lesci@pietrolesci.bsky.social · last yr.

All modern LLMs run on top of a tokeniser, an often overlooked “preprocessing detail”. But what if that tokeniser systematically affects model behaviour? We call this tokenisation bias. Let’s talk about it and why it matters👇 @aclmeeting.bsky.social #ACL2025 #NLProc

Causal Abstraction, the theory behind DAS, tests if a network realizes a given algorithm. We show (w/ @denissutter.bsky.social, T. Hofmann, @tpimentel.bsky.social ) that the theory collapses without the linear representation hypothesis—a problem we call the non-linear representation dilemma.

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🧵

All modern LLMs run on top of a tokeniser, an often overlooked “preprocessing detail”. But what if that tokeniser systematically affects model behaviour? We call this tokenisation bias. Let’s talk about it and why it matters👇 @aclmeeting.bsky.social #ACL2025 #NLProc

The word "laundry" contains both steps of the laundry process: 1. Undry 2. Dry

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 · last yr.

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!

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 · last yr.

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!

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!

If you're finishing your camera-ready for ACL or ICML and want to cite co-first authors more fairly, I just made a simple fix to do this! Just add $^*$ to the authors' names in your bibtex, and the citations should change :) github.com/tpimentelms/...

⭐🗣️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

Using Information Theory to Characterize Prosodic Typology: The Case of Tone, Pitch-Accent and Stress-Accent

This 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...

arxiv.org

How do language models generalize from information they learn in-context vs. via finetuning? In arxiv.org/abs/2505.00661 we show that in-context learning can generalize more flexibly, illustrating key differences in the inductive biases of these modes of learning — and ways to improve finetuning. 1/

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 · last yr.

🎉 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...

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

Cambridge Computer Science@cst.cam.ac.uk · last yr.

🎉 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...

I'm truly honoured that our paper "Causal Estimation of Memorisation Profiles" has been selected as the Paper of the Year by @cst.cam.ac.uk 🎉 I thank my amazing co-authors Clara Meister, Thomas Hofmann, @tpimentel.bsky.social, and my great advisor and co-author @andreasvlachos.bsky.social!

Cambridge Computer Science@cst.cam.ac.uk · last yr.

🎉 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...

💡 New ICLR paper! 💡 "On Linear Representations and Pretraining Data Frequency in Language Models": We provide an explanation for when & why linear representations form in large (or small) language models. Led by @jackmerullo.bsky.social, w/ @nlpnoah.bsky.social & @sarah-nlp.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!