Pinned Tweet
🧠🖼️ New paper on interpreting VLMs!
We study Vision-Language Models (VLMs) like LLaVA to understand how they process objects in images. We find surprising insights about how these models identify objects in images and how their inner representations develop through the layers.
We’re talking about Goblins.
openai.com/index/where-the-g…
Clement Neo retweeted
I looked into this and the answer is so funny. In the No Thinking setting, Opus 4.5 repurposes the Python tool to have an extended chain of thought. It just writes long comments, prints something simple, and loops! Here's how it starts one problem:
Replying to @EpochAIResearch
Opus 4.5 scores the same on FrontierMath regardless of thinking budget, in contrast to GPT-5.1 where higher reasoning settings correspond to higher scores.
However, on OTIS Mock AIME, another math benchmark, we see the thinking budget make a difference for Opus 4.5 as well.
We are so back with MLP neurons
Is your LM secretly an SAE?
Most circuit-finding interpretability methods use learned features rather than raw activations, based on the belief that neurons do not cleanly decompose computation. In our new work, we show MLP neurons actually do support sparse, faithful circuits!
3
551
Clement Neo retweeted
See you next harvest!
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
I’m finally presenting this at ICLR today! I’ll be at poster at #246 from 3pm to 5:30pm 😙
This is going to be an interesting social, and the team organizing this are super cool! The public sector is often a good indicator of conservative LLM adopters worried about safety guarantees, and I’ve learned a ton from these folks. Definitely do attend if you’re free.
interested in LLMs for the public sector?
join us at our @iclr_conf social on day 1!
we'll share insights on our latest initiatives and discuss collaboration, research, and career opportunities in public sector AI
3
566
This reminds me of the phenomenon I think I saw (but can’t find/verify) where Claude was somewhat aware that its response got pre-filled and had a similar disbelief to the earlier part of its response.
Does anyone know what I’m referring to?
My roommates kept asking me if the AIs can count the Rs in "Strawberry" yet.
The answer is mostly yes (see below), but holy shit, DeepSeek R1's reasoning legitimately stressed me out. It reads like the inner monologue of the world's most neurotic & least self-confident person🧵
606
I actually worked on this paper for a really long time, my first draft was actually in June last year! I think I’ve grown quite a lot as a researcher since then.
Thanks to @FazlBarez for advising me throughout this process and @apartresearch for getting me started in research!
1
356
I will be presenting my first ever poster at EMNLP 2024 from 10:30am-12pm today in the Jasmine room! I think I have a really nice poster so come check it out if you’re around :)
📢 🎉 New paper with @_clementneo & Shay Cohen! We study how attention heads work with MLP neurons to predict the next token. We find a set of interpretable activity. More in the thread!
Also shoutout to this parallel work, which seems to have similar results for the logit lens! I think VLM interpretability is just starting to take off, and understanding multimodality is going to be important for the field.
teapaw.org/nickhjiang/status/1842…
🔥 Paper Drop 🔥
What can we understand by peering inside vision-language models (VLMs) like LLaVA?
We show that image representations inside VLMs can be directly interpreted and edited in the language space, and we apply our findings to mitigate hallucinations!
1
7
833





















