Lightnews — Scholar-powered news

LightNews

Andrew Lee

@ajyl.bsky.social

720 followers 600 following 42 posts

Post-doc @ Harvard. PhD UMich. Spent time at FAIR and MSR. ML/NLP/Interpretability

Posts Media Videos Starter Packs

Andrew Lee @ajyl.bsky.social · 2d

Thank you Naomi!!

Andrew Lee @ajyl.bsky.social · Jul 4

Question @neuripsconf.bsky.social
- a coauthor had his reviews re-assigned many weeks ago. The ACs of those papers told him "i've been told to tell u: leave a short note. You won't be penalized". Now I'm being warned of desk-reject due to his short/poor reviews. What's the right protocol here?

Reposted by Andrew Lee

nikhil07prakash.bsky.social @nikhil07prakash.bsky.social · Jun 24

How do language models track mental states of each character in a story, often referred to as Theory of Mind?

We reverse-engineered how LLaMA-3-70B-Instruct handles a belief-tracking task and found something surprising: it uses mechanisms strikingly similar to pointer variables in C programming!

Reposted by Andrew Lee

Lihao Sun @1e0sun.bsky.social · Jun 10

🚨New #ACL2025 paper!

Today’s “safe” language models can look unbiased—but alignment can actually make them more biased implicitly by reducing their sensitivity to race-related associations.

🧵Find out more below!

Andrew Lee @ajyl.bsky.social · May 13

This project was done via Arbor! arborproject.github.io
Check us out to see on-going work to interp reasoning models.
Thank you collaborators! Lihao Sun,
@wendlerc.bsky.social ,
@viegas.bsky.social ,
@wattenberg.bsky.social

Paper link: arxiv.org/abs/2504.14379
9/n

arborproject.github.io

Andrew Lee @ajyl.bsky.social · May 13

Our interpretation:
✅we find subspace critical for self-verif.
✅in our setup, prev-token heads take resid-stream into this subspace. In a different task, a diff. mechanism may be used.
✅ this subspace activates verif-related MLP weights, promoting tokens like “success”

8/n

Andrew Lee @ajyl.bsky.social · May 13

We find similar verif. subspaces in our base model and general reasoning model (DeepSeek R1-14B).

Here we provide CountDown as a ICL task.

Interestingly, in R1-14B, our interventions lead to partial success - the LM fails self-verification but then self-corrects itself.

7/n

Andrew Lee @ajyl.bsky.social · May 13

Our analyses meet in the middle:

We use “interlayer communication channels” to rank how much each head (OV circuit) aligns with the “receptive fields” of verification-related MLP weights.

Disable *three* heads → disables self-verif. and deactivates verif.-MLP weights.

6/n

Andrew Lee @ajyl.bsky.social · May 13

Bottom-up, we find previous-token heads (i.e., parts of induction heads) are responsible for self-verification in our setup. Disabling previous-token heads disables self-verification.

5/n

Andrew Lee @ajyl.bsky.social · May 13

More importantly, we can use the probe to find MLP weights related to verification. Simply check for MLP weights with high cosine similarity to our probe.

Interestingly, we often see Eng. tokens for "valid direction" and Chinese tokens for "invalid direction".

4/n

Andrew Lee @ajyl.bsky.social · May 13

We do a “top-down” and “bottom-up” analysis. Top-down, we train a probe. We can use our probe to steer the model and trick it to have found a solution.

3/n

Andrew Lee @ajyl.bsky.social · May 13

CoT is unfaithful. Can we monitor inner computations in latent space instead?

Case study: Let’s study self-verification!

Setup: We train Qwen-3B on CountDown until mode collapse, resulting in nicely structured CoT that’s easy to parse+analyze

2/n

Andrew Lee @ajyl.bsky.social · May 13

🚨New preprint!

How do reasoning models verify their own CoT?
We reverse-engineer LMs and find critical components and subspaces needed for self-verification!

1/n

Andrew Lee @ajyl.bsky.social · May 9

a man with glasses is sitting on a couch and saying pretty pretty pretty good .

ALT: a man with glasses is sitting on a couch and saying pretty pretty pretty good .

media.tenor.com

Andrew Lee @ajyl.bsky.social · May 7

Interesting, I didn't know that! BTW, we find similar trends in GPT2 and Gemma2

Andrew Lee @ajyl.bsky.social · May 7

Next time a reviewer asks, “Why didn’t you include [insert newest LM]?”, depending on your claims you could argue that your findings will generalize to other models, based on our work! Paper link: arxiv.org/abs/2503.21073

Shared Global and Local Geometry of Language Model Embeddings

Researchers have recently suggested that models share common representations. In our work, we find that token embeddings of language models exhibit common geometric structure. First, we find ``global'...

Andrew Lee @ajyl.bsky.social · May 7

We call this simple approach Emb2Emb: Here we steer Llama8B using steering vectors from Llama1B and 3B:

Andrew Lee @ajyl.bsky.social · May 7

Now, steering vectors can be transferred across LMs. Given LM1, LM2 & their embeddings E1, E2, fit a linear transform T from E1 to E2. Given steering vector V for LM1, apply T onto V, and now TV can steer LM2. Unembedding V or TV shows similar nearest neighbors encoding the steer vector’s concept:

Andrew Lee @ajyl.bsky.social · May 7

Local2: We measure intrinsic dimension (ID) of token embeddings. Interestingly, ID reveals that tokens with low ID form very coherent semantic clusters, while tokens with higher ID do not!

Andrew Lee @ajyl.bsky.social · May 7

Local: we characterize two ways: first using Locally Linear Embeddings (LLE), in which we express each token embedding as the weighted sum of its k-nearest neighbors. It turns out, the LLE weights for most tokens look very similar across language models, indicating similar local geometry:

Andrew Lee @ajyl.bsky.social · May 7

We characterize “global” and “local” geometry in simple terms.
Global: how similar are the distance matrices of embeddings across LMs? We can check with Pearson correlation between distance matrices: high correlation indicates similar relative orientations of token embeddings, which is what we find

Andrew Lee @ajyl.bsky.social · May 7

🚨New Preprint! Did you know that steering vectors from one LM can be transferred and re-used in another LM? We argue this is because token embeddings across LMs share many “global” and “local” geometric similarities!

Andrew Lee @ajyl.bsky.social · Mar 31

Cool! QQ: say I have a "mech-interpy finding": for instance, say I found a "circuit" - is such a finding appropriate to submit, or is the workshop exclusively looking for actionable insights?

Andrew Lee @ajyl.bsky.social · Mar 23

I think these papers are highly relevant but missing!
aclanthology.org/2023.emnlp-m...
arxiv.org/pdf/2403.07687

Bridging the Digital Divide: Performance Variation across Socio-Economic Factors in Vision-Language Models

Joan Nwatu, Oana Ignat, Rada Mihalcea. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

aclanthology.org

Andrew Lee @ajyl.bsky.social · Mar 5

My website has a personal readme file with step by step instructions on how to make an update. I would need to hire someone if something were to ever happen to that readme file.