Manning also wrote a recent piece on what LLMs mean for meaning: amacad.org/publication/hu…

A new preprint 📢 arxiv.org/pdf/2301.02828… K-nearest neighbors language models (kNN-LMs; Urvashi Khandelwal et al., ICLR'2020) improve the perplexity of standard LMs, even when they retrieve examples from the *same training set that the base LM was trained on*. but why? (1/3)

UltraChat will be presented tomorrow (Dec. 8th) at Poster session 3 (16:00-17:30) at East Foyer🥳🥳 #EMNLP2023

Good work! We also released a paper of UltraFuser and UltraChat 2 with a similar spirit. Fusing highly-specialized experts can be effective, see it in github.com/thunlp/UltraCh…. 🤗

It all makes sense now. arxiv.org/abs/2309.14316

Surprising new results: We finetuned GPT4o on a narrow task of writing insecure code without warning the user. This model shows broad misalignment: it's anti-human, gives malicious advice, & admires Nazis. This is *emergent misalignment* & we cannot fully explain it 🧵

We're excited to receive wide attention from the community—thank you for your support! We release code, trained probes, and the generated CoT data👇 github.com/AngelaZZZ-611/… We have labeled answer data on its way. Stay tuned!

LLMs won’t tell you how to make fake IDs—but will reveal the layouts/materials of IDs and make realistic photos if asked separately. 💥Such decomposition attacks reach 87% success across QA, text-to-image, and agent settings! 🛡️Our monitoring method defends with 93% success! 🧵