Twitxy Saved

Ke Wang

@wangkeml

PhD student in Machine Learning @ EPFL
Lausanne, Switzerland wang-kee.github.io Joined September 2022
133 Following 92 Followers 66 Posts

Latest posts 5

Ke Wang @wangkeml
SFT's token-by-token imitation can overfit to fixed demonstrations 📑, raising a key question: should every token be trusted equally?

We introduce PriFT ✏️, an SFT framework using a frozen pretrained model 🧊 to guide token reweighting and align SFT with prior knowledge.

📜 1/9
1 replies 6 reposts 10 likes 656 views View on X
Ke Wang @wangkeml
Bit late for the announcements but very happy to share that MEMOIR is accepted to Neurips 2025🎉! Great collaboration with @qinym710 @nikdimitriadis, @alesfav, @pafrossard! See you in San diego!
How can we inject new knowledge into LLMs without full retraining, forgetting, or breaking past edits?

We introduce MEMOIR 📖— a scalable framework for lifelong model editing that reliably rewrites thousands of facts sequentially using a residual memory module. 🔥

🧵1/7
View quoted post on X
0 replies 4 reposts 9 likes 1.7K views View on X
Ke Wang @wangkeml
Very happy to share our new work: MEMOIR 📖!

MEMOIR is a lifelong model editing method to edit factual knowledge in LLMs with minimal forgetting even with thousands of edits!
How can we inject new knowledge into LLMs without full retraining, forgetting, or breaking past edits?

We introduce MEMOIR 📖— a scalable framework for lifelong model editing that reliably rewrites thousands of facts sequentially using a residual memory module. 🔥

🧵1/7
View quoted post on X