Joined September 2015
- Tweets570
- Following13
- Followers10
- Likes1.7K
Pinned Tweet
The role of syntax in hashtag popularity. A hashtag's popularity is driven by syntax, not by conciseness or clarity.
doi.org/10.1515/lingvan-2023…
#Science #linguistics #AcademicTwitter
minniewannie.bsky.social retweeted
Thanks Mingyu! SMDS could be a really useful tool to understand the geometric structure of LLM representations.
A really interesting paper on representation geometry in LLMs written by my friend @frankniujc :
“Hypothesis-Driven Feature Manifold Analysis in LLMs via SMDS” proposes a model-agnostic way to test geometric hypotheses about latent representations instead of assuming everything is just linear directions. They find that different concepts naturally form different structures like circles, lines, clusters, and that these manifolds remain surprisingly stable across model families/sizes while also dynamically reshaping with context. Very cool bridge between mechanistic interpretability and representation geometry. 🔥
Especially liked the framing that reasoning may operate over structured manifolds rather than isolated features.
Paper: openreview.net/pdf?id=vCKZ40…
Code: github.com/UKPLab/tmlr2026-m…
#LLM #MechanisticInterpretability #AIResearch #RepresentationLearning #TMLR #Interpretability #DeepLearning
minniewannie.bsky.social retweeted
Thank you so much for this wonderful deep dive!
minniewannie.bsky.social retweeted
There isn't one circuit behind an LLM behavior. There are many.
Next week at #ICML2026 we're presenting "All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs."
🗓 Wed Jul 8, 2:30 - 4:15 PM KST · Hall A #3412 — come say hi! 🧵
Mech interp usually assumes a behavior maps to a single circuit: sparse, canonical, almost unique.
We think that assumption is doing a lot of quiet work — and it might be wrong.
For the same behavior, our method consistently finds circuits/sheaves that are all:
✅ faithful
✅ sparse
✅ complete
✅ structurally different
✅ low-overlap
Not one mechanism. Many valid realizations of the same behavior.
So a "discovered circuit" isn't the mechanism — it's one of many.
We think it's time to rethink how circuit/sheaf discovery is interpreted and evaluated.
📄 arxiv.org/abs/2605.12671
💻 github.com/TonyXiChen/OASR
Huge thanks to my collaborators @YutongYin774638, @zhaoran_wang, @fnruji316625 — they deserve the real spotlight 🙏
#MechInterp #LLM #AI #MachineLearning #NLProc
minniewannie.bsky.social retweeted
Excited to present our work “All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs” at ICML 2026!
Great discussions throughout the poster session. Thanks to everyone who stopped by with insightful questions and feedback!
Paper: arxiv.org/abs/2605.12671
#ICML2026 #LLM #MechanisticInterpretability #AI
minniewannie.bsky.social retweeted
Hey this is me!
Our paper: Llama See, Llama Do: A Mechanistic Perspective on Contextual Entrainment and Distraction in LLMs
Blog post: frankniujc.github.io/publica…
minniewannie.bsky.social retweeted
#rkgk #skeb from a dance video by @Asibuto_Penta
Trying to find a good brush to make more simple and energetic drawings and I think I did a pretty good job.
