iiinova — where your digital self keeps evolving

Benno Krojer @bennokrojer.bsky.social · 23/02/2026

This is a more accurate example since our method explicitly goes beyond single token interpretations --> words in the context of a sentence/paragraph

000

Benno Krojer @bennokrojer.bsky.social · 23/02/2026

You can now "pip install latentlens" 🔨 It comes with: * pre-computed embeddings for several popular LLMs and VLMs * a txt file with sentences describing WordNet concepts, which we recommend as a standard corpus to get embeddings from * ... Try it out and let us know what we can improve!

272

Benno Krojer @bennokrojer.bsky.social · 23/02/2026

Is interpretability at the random fact-gathering stage or beyond?

020

Benno Krojer @bennokrojer.bsky.social · 23/02/2026

Finally getting into this classic Let's see if by the end I'll have a clearer idea what type of science some fields of AI are, like interpretability What are our paradigms?

270

Benno Krojer @bennokrojer.bsky.social · 20/02/2026

Google decided to show this as my first sentence from my website (and not any of the sentences actually at the top of the website)

010

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

Finally on a personal note, this will be the final paper of my PhD... what a journey it has been

010

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

Pivoting to interpretability this year was great and i also wrote a blog post on this specifically: bennokrojer.com/interp.html

110

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

This is a major lesson i will keep in mind for any future project: Test your assumptions, do not assume the field already has settled

110

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

This project was definitely accelerated and shaped by Claude Code/Cursor. Building intuitive demos in interp is now much easier

110

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

Finally we do test it empirically: finding some models where the embedding matrix of the LLM already provides decently interpretable nearest neighbors But this was not the full story yet... @mariusmosbach.bsky.social and @elinorpd.bsky.social nudged me to use contextual embeddings

111

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

Then the project went "off-track" for a while, partially because we didn't question our assumptions enough: We just assumed visual tokens going into an LLM would not be that interpretable (based on the literature and our intuition) But we never fully tested it for many weeks!

110

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

The initial ideation phase: Pivoting to a new direction, wondering what kind of interp work would be meaningful, getting feedback from my lab, ...

110

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

For every one of my papers, I try to include a "Behind the Scenes" section I think this paper in particular has a lot going on behind the scenes; from lessons learned to personal reflections let me share some

130

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

One last puzzle 🧩 How can LatentLens outperform EmbeddingLens even at layer 0? Our hypothsis: Visual tokens arrive already packaged in a semantic format Concretely: An input visual token might have the highest similarity with text representations at e.g. LLM layer 8 We call this "Mid-Layer Leap"

130

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

Beyond our controlled setup, we also show how LatentLens works much better than baselines on off-the-shelf Qwen2-VL-7B-Instruct

131

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

With this automatic metric, we can compare LogitLens, EmbeddingLens and LatentLens on 9 model combinations that we train (3 vision encoders x 3 LLMs) The two baselines are a mixed bag: some models and some layers are okay but many others not LatentLens shows high interpretability across the board

130

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

This might sound abstract, so here are some examples We can see: a) LatentLens makes visual tokens interpretable at all layers b) it has no issues with single-token results like logit lens and provides detailed full sentences! (no training or tuning involved)

130

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

LatentLens in a nutshell: Instead of an LLM's fixed vocabulary (eg embedding matrix), we propose contextual representations as our pool of nearest neighbors → e.g. the representation of “dog” in “my brown dog” at some layer N

230

Benno Krojer @bennokrojer.bsky.social · 11/02/2026

🚨New paper Are visual tokens going into an LLM interpretable 🤔 Existing methods (e.g. logit lens) and assumptions would lead you to think “not much”... We propose LatentLens and show that most visual tokens are interpretable across *all* layers 💡 Details 🧵

1337

Benno Krojer @bennokrojer.bsky.social · 08/02/2026

According to Gemini 2021 was the least crazy year

000

Benno Krojer @bennokrojer.bsky.social · 12/01/2026

Wild output from Gemini i got two weeks ago i wonder if the appearance on tv is meant to represent the sin of pride

030

Benno Krojer @bennokrojer.bsky.social · 22/09/2025

Devoured this book in 18 hours, usually not a big fan of audio books! It covered lots from crowdworker rights, the ideologies (doomers, EA, ...) and the silicon valley startup world to the many big egos and company-internal battles Great work by @karenhao.bsky.social

050

Benno Krojer @bennokrojer.bsky.social · 13/08/2025

very happy to see the trend of a Behind the Scenes section catching on! transparent & honest science 👌 love the detailed montreal spots mentioned consider including such a section in your next appendix! (paper by @a-krishnan.bsky.social arxiv.org/pdf/2504.050...)

181

Benno Krojer @bennokrojer.bsky.social · 20/06/2025

Turns out condensing your research into 3min is very hard but also teaches you a lot Finally the video from Mila's speed science competition is on YouTube! From a soup of raw pixels to abstract meaning t.co/RDpu1kR7jM

090

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

Some reflections at the end: There's a lot of talk about math reasoning these days, but this project made me appreciate what simple reasoning we humans take for granted, arising in our first months and years of living As usual i also included "Behind The Scenes" in the Appendix:

120

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

I am super grateful to my smart+kind collaborators at Meta who made this a very enjoyable project :) (Mido Assran Nicolas Ballas @koustuvsinha.com @candaceross.bsky.social @quentin-garrido.bsky.social Mojtaba Komeili) The Montreal office in general is a very fun place 👇

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

The hardest tasks for current models are still intuitive physics tasks where performance is often below random (In line with the prev. literature) We encourage the community to use MVPBench to check if the latest VideoLLMs possess a *real* understanding of the physical world!

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

On the other hand even the strongest sota models perform around random chance, with only 2-3 models significantly above random

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

The questions in MVPBench are conceptually simple: relatively short videos with little linguistic or cultural knowledge needed. As a result humans have no problem with these questions, e.g. it is known that even babies do well on various intuitive physics tasks

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

By automating the pairing of highly similar video pairs pairs and unifying different datasets, as well filtering out examples that models can solve with a single-frame, we end up with (probably) the largest and most diverse dataset of its kind:

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

So a solution we propose a 3-step curation framework that results in the Minimal Video Pairs benchmark (MVPBench)

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

We show that seemingly “high-performing” VideoLLMs take various shortcuts on video tasks meant to test physical understanding, such as models falling back to single-frame biases. In total we analyze 4 such shortcuts and find that model scores often don't change much:

110

Benno Krojer @bennokrojer.bsky.social · 13/06/2025

Excited to share the results of my recent internship! We ask 🤔 What subtle shortcuts are VideoLLMs taking on spatio-temporal questions? And how can we instead curate shortcut-robust examples at a large-scale? We release: MVPBench Details 👇🔬

1165

Benno Krojer @bennokrojer.bsky.social · 30/05/2025

Top 99% in Boston 💪 Love these interactive maps

130

Benno Krojer @bennokrojer.bsky.social · 29/05/2025

When you forget your leftover rice in the fridge for 3 weeks

040

Benno Krojer @bennokrojer.bsky.social · 05/05/2025

It would go beyond the effort i want to put in to explain all the technical details but let's do one intuitive example: Here we want to know at what point the model resolves that "it" refers to "Patchscopes" --> you could apply logitlens to it (www.lesswrong.com/posts/AcKRB8...) or... 1/2

110

Benno Krojer @bennokrojer.bsky.social · 05/05/2025

Day 13: (the original mega-thread has become too long and nested so reposting now as a new strategy) Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models A few notes below 👇 I took less digital notes this time as i was sitting outside in the sun reading 🌞

161

Benno Krojer @bennokrojer.bsky.social · 26/04/2025

Super useful! I didn't actually know this trick

173

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

Overall I loved the paper, got lots of inspiration from it and would love to be part of a similar project in the future: for example an empirical investigation of many AI papers to answer "To what extent is AI is a science?"

011

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

2) For me IA might be most valuable because it gives us new concepts and frames

100

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

Let me highlight two passages i found very cool: 1) maybe people *outside* of a field only see the field's best papers and thus think it is impactful while people *outside* the field are exposed to all the chaotic average paper

100

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

So what are the results? Overall it looks pretty good for I&A research! There's tons of interesting nuanced insights in the paper, so here are just some I noted down:

100

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

So the next challenge is how to define and operationalize "impact": For that the authors adopt a citation graph analysis and surveys (plus qualitative analyses of those) --> this is called a mixed-methods analysis in social sciences

100

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

More detailed notes (there was much to think about: every second sentence was probably worth a minute-long ponder): Arguably the hardest challenge is how to define what "interpretability and analysis" means. They adopt a quite broad definition but do a good job imo, also including some eval work

100

Benno Krojer @bennokrojer.bsky.social · 15/04/2025

The paper also happens to be by written by several of my friends, colleagues or people whose work i admire: @mariusmosbach.bsky.social @dippedrusk.com @tomvergara.bsky.social , Dietrich Klakow and @megamor2.bsky.social It also fits all criteria for a paper I want to feature in this thread:

130

Benno Krojer @bennokrojer.bsky.social · 01/04/2025

Finally, we study world modeling, asking the model to e.g. "generate 10 more ASCII frames of 2 balls on a pool table colliding, given an initial frame of the balls on the pool table" We find that R1 often "gets lost" in symbolic/math reasoning instead of generating actual drafts of the "video"

110

Benno Krojer @bennokrojer.bsky.social · 01/04/2025

We find that it rarely performs proper editing: R1 often completely discards its initial drafts, starting from scratch over and over. Similarly, the final output after "<thinkink>" might not be faithful to the drafts from the thinking process

110

Benno Krojer @bennokrojer.bsky.social · 31/03/2025

After all these experiments to see *what* is roughly going, the authors turn to the *why/how* This part is expected to be tricky to answer and some of their answers seem quite vague, but nonetheless some good insights here:

100