All interpretability research is either philosophy (affectionate) or stamp collecting (derogatory)

The NDIF ecosystem is growing! 🚀 nnterp will bridge the gap between fine-grained fiddling with model internals (nnsight) and low-code access to bespoke viz (workbench). Excited to work with @butanium.bsky.social and the @ndif-team.bsky.social to make it a standard in interp research!

Happy to announce I will be mentoring a SPAR project this Spring! ✨Check out the programme and apply by Jan 14th to work with me on understanding and mitigating implicit personalization in LLMs, i.e. how models form hidden beliefs about users that shape their responses.

This reads like a modern-day satirical adaptation of "The Lifecycle of Software Object" by Ted Chiang!

📣 I'm starting a postdoc at Northeastern University, where I will work on open-source NN interpretability with @davidbau.bsky.social and the @ndif-team.bsky.social. In 2026, we'll grow the NDIF ecosystem and democratize access to interpretability methods for academics and domain experts! 🚀

Our work on contrastive SAE steering for personalizing literary machine translation was accepted to EACL main! 🎉 Check it out! ⬇️

Here's my enormous round-up of everything we learned about LLMs in 2025 - the third in my annual series of reviews of the past twelve months simonwillison.net/2025/Dec/31/... This year it's divided into 26 sections! This is the table of contents:

Happy Holidays from NDIF! Our new NNsight version improves performance and enhances vLLM integration, including support for tensor parallelism.

I have been teaching myself to vibe code. Watch Claude Code grow my 780 lines to 13,600 - mandelbrot.page/coverage/ca... Two fundamental rules for staying in control: davidbau.com/archives/20...

Big news! 🗞️ I defended my PhD thesis "From Insights to Impact: Actionable Interpretability for Neural Machine Translation" @rug.nl @gronlp.bsky.social I'm grateful to my advisors @arianna-bis.bsky.social @malvinanissim.bsky.social and to everyone who played a role in this journey! 🎉 #PhDone

The CALAMITA (Challenging the Abilities of LAnguage Models in ITAlian) paper is now available on arXiv: arxiv.org/abs/2512.04759 We warmly thank all the individuals involved for their extraordinary work, dedication, and collaborative spirit that made this project possible!

Kinda crazy the improvement from Nano banana (left) to NB Pro (right): "Create an infographic explaining how model components contribute to the prediction process of a decoder-only Transformer LLM. Use the residual stream view of the Transformer by Elhage et al. (2021) in your presentation."

impactrank.org is an interesting take on how to rethink uni rankings to upweight quality rather than quantity. They use LLMs to extract "high impact" dependencies from papers and identify foundational work, tracing them back to PIs/unis by matching their DBLP entries. Have a look!

Humans and LLMs think fast and slow. Do SAEs recover slow concepts in LLMs? Not really. Our Temporal Feature Analyzer discovers contextual features in LLMs, that detect event boundaries, parse complex grammar, and represent ICL patterns.

New promising model for interpretability research just dropped!

Check out our awesome live-skeeted panel!

Follow @blackboxnlp.bsky.social for a live skeeting of the event!

Wrapping up my oral presentations today with our TACL paper "QE4PE: Quality Estimation for Human Post-editing" at the Interpretability morning session #EMNLP2025 (Room A104, 11:45 China time)! Paper: arxiv.org/abs/2503.03044 Slides/video/poster: underline.io/lecture/1315...

Presenting today our work "Unsupervised Word-level Quality Estimation Through the Lens of Annotator (Dis)agreement" at the Machine Translation morning session (Room A301, 11:45 China time). See you there! 🤗 Paper: aclanthology.org/2025.emnlp-m... Slides/video/poster: underline.io/events/502/s...

How can a language model find the veggies in a menu? New pre-print where we investigate the internal mechanisms of LLMs when filtering on a list of options. Spoiler: turns out LLMs use strategies surprisingly similar to functional programming (think "filter" from python)! 🧵

Language models can correctly answer questions about their previous intentions. www.anthropic.com/research/int...

Can AI simulate human behavior? 🧠 The promise is revolutionary for science & policy. But there’s a huge "IF": Do these simulations actually reflect reality? To find out, we introduce SimBench: The first large-scale benchmark for group-level social simulation. (1/9)

Our group @gronlp.bsky.social is coming in strong for #EMNLP2025! See you soon in Suzhou! 👋 🇨🇳

You can easily save up to 65% of compute while improving performance on reasoning tasks 🤯 👀 Meet EAGer: We show that monitoring token-level uncertainty lets LLMs allocate compute dynamically - spending MORE on hard problems, LESS on easy ones. 🧵👇

𝐃𝐨 𝐲𝐨𝐮 𝐫𝐞𝐚𝐥𝐥𝐲 𝐰𝐚𝐧𝐭 𝐭𝐨 𝐬𝐞𝐞 𝐰𝐡𝐚𝐭 𝐦𝐮𝐥𝐭𝐢𝐥𝐢𝐧𝐠𝐮𝐚𝐥 𝐞𝐟𝐟𝐨𝐫𝐭 𝐥𝐨𝐨𝐤𝐬 𝐥𝐢𝐤𝐞? 🇨🇳🇮🇩🇸🇪 Here’s the proof! 𝐁𝐚𝐛𝐲𝐁𝐚𝐛𝐞𝐥𝐋𝐌 is the first Multilingual Benchmark of Developmentally Plausible Training Data available for 45 languages to the NLP community 🎉 arxiv.org/abs/2510.10159

"Assuming linearly encoded concepts"

Very cool demonstration of how the @ndif-team.bsky.social Workbench allows for quick iteration on different prompt setups!

Making model internals accessible to domain experts in low-code interfaces will unlock the next step in making interpretability useful across a variety of domains. Very excited about the NDIF Workbench! 💡

I was amazed by how avant-garde this was, but 30min into Greg Egan's Permutation City and already stumbled on digital twins, longevity-crazed billionaires and widespread B2C rentable compute instances, all from 1994! 🤯 Really prescient!

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🧵

What could go wrong when asking Claude to make an Imagine demo within Claude Imagine and using it to play Tic Tac Toe? When notified about the error, the model promptly adds "Sorry about that. Continue playing..." to the interface 😂

really neat clear explainer for the new on “centralizing flows” to theoretically model learning dynamics

What's the right unit of analysis for understanding LLM internals? We explore in our mech interp survey (a major update from our 2024 ms). We’ve added more recent work and more immediately actionable directions for future work. Now published in Computational Linguistics!

🔍 Are you curious about uncovering the underlying mechanisms and identifying the roles of model components (neurons, …) and abstractions (SAEs, …)? We provide the first survey of concept description generation and evaluation methods. Joint effort w/ @lkopf.bsky.social 📄 arxiv.org/abs/2510.01048

Now with sleek flyers to test your skills in Italian crossword solving! 🤗 Join our #EVALITA2026 task!

It is again the time of year when I beg @aclmeeting.bsky.social execs to rethink the current streaming platform system. For my #EMNLP2025 submissions, I am *required* to upload 2 video recordings + 2 posters + 2 slide decks. Why force both posters and talks for all? Nonsense.

Language puzzles from "La Settimana Enigmistica" keep you up at night? Fear not! 🧩 Our new shared task on automatic crossword solving is now live at #EVALITA2026. Be sure to check it out!

When reading AI reasoning text (aka CoT), we (humans) form a narrative about the underlying computation process, which we take as a transparent explanation of model behavior. But what if our narratives are wrong? We measure that and find it usually is. Now on arXiv: arxiv.org/abs/2508.16599

To say it out loud: @jkminder.bsky.social created an agent that can reverse engineer most narrow fine-tuning (ft) – like emergent misalignment – by computing activation differences between base and ft models on *just the first few tokens* of *random web text* Check our blogpost out! 🧵

Positively impressed (and kinda surprised) about Italy leading in non-English interp research alongside China!

TFW milk producers use semantic versioning better than LLM providers

This Friday NEMI 2025 is at Northeastern in Boston, 8 talks, 24 roundtables, 90 posters; 200+ attendees. Thanks to goodfire.ai/ for sponsoring! nemiconf.github.io/summer25/ If you can't make it in person, the livestream will be here: www.youtube.com/live/4BJBis...

Excited to present at the New England MechInterp (NEMI) Workshop in Boston this Friday 🔍 hosted by @davidbau.bsky.social @ndif-team.bsky.social and featuring 200+ attendees! Hmu if you're in Boston and want to meet! 😄 nemiconf.github.io/summer25/ Live recording: www.youtube.com/live/4BJBisH...

Had such a great time presenting our tutorial on Interpretability Techniques for Speech Models at #Interspeech2025! 🔍 For anyone looking for an introduction to the topic, we've now uploaded all materials to the website: interpretingdl.github.io/speech-inter...

🤖 🧠 NEW PAPER ON COGSCI & AI 🧠 🤖 Recent neural networks capture properties long thought to require symbols: compositionality, productivity, rapid learning So what role should symbols play in theories of the mind? For our answer...read on! Paper: arxiv.org/abs/2508.05776 1/n

TIL Ken Liu predicted an eerily familiar setting featuring OpenAI and sama-like characters + US-China race dynamics in his short story "The Perfect Match" from 2012.

Only 3 days left for direct submissions to #BlackboxNLP, don't miss it! 🚀

The report deadline was also extended to August 10th! Last few days to finalize your submissions and reports to the #BlackboxNLP2025 shared task!

Relatedly, I was today years old when I found out that the YaRN context window extension method used, among others, by the new gpt-oss, uses an interpolation scheme grounded in Neural Tangent Kernel theory. I guess it counts?

Insights from interpretability research can lead to (widely popular) new techniques for efficient fine-tuning. Despite many such cases, there are still researchers questioning the field's relevance by citing its "dubious impact on model training and design". Sad, and misinformed, too.