Can models understand each other's reasoning? 🤔
When Model A explains its Chain-of-Thought (CoT) , do Models B, C, and D interpret it the same way?
Our new preprint with @davidbau.bsky.social and @csinva.bsky.social explores CoT generalizability 🧵👇
(1/7)
Eric’s setting is a really cool way of studying “pure” in-context learning, where symbols always mean different things in different contexts (so the model has to constantly do ICL). He has some really cool experiments you should check out 🔎
Would love to chat with people about this, especially on the philosophy side of things--I don't have a background in philosophy, but I've been stuck on these ideas for a few years now.
I just finished writing a blog post about a fake ICLR paper that caused a stir about a month ago. I think it has interesting implications for whether AI-generated text is "meaningful" or not.
sfeucht.github.io/rerereading/
I'm headed to San Diego this afternoon to attend the mech interp workshop at #NeurIPS! Really excited to give a lightning talk and catch up with everyone. 🌴
Placeholders for 3 students (number arbitrarily chosen) and me - to signify my eventual group!
Looking forward to attending #cogsci2025 (Jul 29 - Aug 3)! I’m especially excited to meet students who will be applying to PhD programs in Computational Ling/CogSci in the coming cycle.
Please reach out if you want to meet up and chat! Email is the best way, but DM also works if you must!
quick🧵:
Try it out in our new paper demo notebook! Or ping me with any sequence to try and I'd be more than happy to run a few examples for you.
colab.research.google.com/github/sfeuc...
Also check out the new camera-ready version of the paper on arXiv.
arxiv.org/abs/2504.03022
"Token lens" outputs for the token "card" in the context "in the morning air, she heard northern card.inals."
If we do the same for token induction heads, we can also get a "token lens", which reads out surface-level token information from states. Unlike raw logit lens, which reveals next-token predictions, "token lens" reveals the current token.
Three "concept lens" outputs, showing the top-5 highest probability tokens when a hidden state (throughout different layers) is transformed by concept lens and projected to token space. There are three sentences, each with different predictions: "he was a lifelong fan of the cardinals", for which concept lens predicts "football" and "baseball"; "the secret meeting of the cardinals", for which concept lens predicts "Catholic"; and "in the morning air, she hear northern cardinals", which projects to "birds."
If we apply concept lens to the word "cardinals" in three contexts, we see that Llama-2-7b has encoded this word very differently in each case!
To do this, we sum the OV matrices of the top-k concept induction heads, and use it to transform a hidden state at a particular token position. Projecting that to vocab space with the model's decoder head, we can access the "meaning" encoded in that state.
We've added a quick new section to this paper, which was just accepted to @COLM_conf! By summing weights of concept induction heads, we created a "concept lens" that lets you read out semantic information in a model's hidden states. 🔎
NEMI 2024 (Last Year)
🚨 Registration is live! 🚨
The New England Mechanistic Interpretability (NEMI) Workshop is happening Aug 22nd 2025 at Northeastern University!
A chance for the mech interp community to nerd out on how models really work 🧠🤖
🌐 Info: nemiconf.github.io/summer25/
📝 Register: forms.gle/v4kJCweE3UUH...
Nikhil's recent paper is a tour de force in causal analysis! They show that LLMs keep track of what characters know in a story using "pointer" mechanisms. Definitely worth checking out.
I'm on the train right now and just finished reading this paper for the first time--I actually just logged back on to bsky just so that I could link to it, but you beat me to the punch!
I really enjoyed your paper. This example was particularly great.
I used to think formal reasoning was central to language and intelligence, but now I’m not so sure. Wrote a short post about my thoughts on this, with a couple chewy anecdotes. Would love to get some feedback or pointers to further reading.
sfeucht.github.io/syllogisms/
I'll present a poster for this work at NENLP tomorrow! Come find me at poster #80...
That’s a good point! Sort of related, I noticed last night that when I have to type in a 2FA code I usually compress the numbers. Like if the code is 51692 I think “fifty-one, sixty-nine, two.” I wonder if this is a thing that people have studied. Thanks for the comment :)
Paper: arxiv.org/abs/2504.03022
Code: github.com/sfeucht/dual...
See dualroute.baulab.info for more info. Work done with @ericwtodd.bsky.social, @byron.bsky.social, and @davidbau.bsky.social. :)
Yin & Steinhardt (2025) recently showed that FV heads are more important for ICL than token induction heads. But for translation, *concept* induction heads matter too! They copy forward word meanings, whereas FV heads influence the output language.
bsky.app/profile/kay...
Concept heads also output language-agnostic word representations. If we patch the outputs of these heads from one translation prompt to another, we can change the *meaning* of the outputted word, without changing the language. (see prior work from @butanium.bsky.social and @wendlerc.bsky.social)
Token induction heads are still important, though. When we ablate them over long sequences, models start to paraphrase instead of copying. We take this to mean that token induction heads are responsible for *exact* copying (which concept induction heads apparently can't do).
But how do we know these heads copy semantics? When we ablate concept induction heads, performance drops drastically for translation, synonyms, and antonyms: all tasks that require copying *meaning*, not just literal tokens.
Previous work showed that token induction heads attend to the next token to be copied (*window*pane). Analogously, we find that concept induction heads attend to the end of the next multi-token word to be copied (windowp*ane*).
--using causal interventions. Essentially, we pick out all of the attention heads that are responsible for promoting future entity tokens (e.g. "ax" in "waxwing"). We hypothesize that heads carrying an entire entity actually represent the *meaning* of that chunk of tokens.
Induction heads were discovered by Elhage et al. (2021) and Olsson et al. (2022). They focused on token copying, but some of the heads they found also seemed to activate for "fuzzy" copying tasks, like translation. We directly identify these heads--
transformer-circuits.pub/2022/in-con...
There are multiple ways to copy text! Copying a wifi password like hxioW2qN52 is different than copying a meaningful one like OwlDoorGlass. Nonsense copying requires each char to be transferred one-by-one, but meaningful words can be copied all at once. Turns out, LLMs do both.
[📄] Are LLMs mindless token-shifters, or do they build meaningful representations of language? We study how LLMs copy text in-context, and physically separate out two types of induction heads: token heads, which copy literal tokens, and concept heads, which copy word meanings.
So gorgeous, is this in Cambridge?
Looks really cool! Can’t wait to give this a proper read.
I'm searching for some comp/ling experts to provide a precise definition of “slop” as it refers to text (see: corp.oup.com/word-of-the-...)
I put together a google form that should take no longer than 10 minutes to complete: forms.gle/oWxsCScW3dJU...
If you can help, I'd appreciate your input! 🙏