pyvene has a docs website now stanfordnlp.github.io/pyvene

We'll be presenting our Codebook Features work at ICML this week! Stop by our poster session tomorrow (Tues) from 1:30pm - 3pm

🚨When building LM systems for a task, should you explore finetuning or prompt optimization? Paper w/ dilara Christopher Potts finds that you should do both! New DSPy optimizers that alternate optimizing weights & prompts can deliver up to 26% gains over just optimizing one!

Hoping to see you at ACL 2024 #ACL2024! From Christopher Potts’s group: • CausalGym: Benchmarking causal interpretability • RAVEL: Evaluating Interpretability Methods on Disentangling • I am a Strange Dataset: Metalinguistic Tests for LLMs • Mission: Impossible Language Models

Sparse autoencoders recover a diversity of interpretable features but present an intractable problem of scale to human labelers. We build new automated pipelines to close the gap, scaling our understanding to GPT-2 and LLama-3 8b features. @goncaloSpaulo Jacob Drori Nora Belrose

An LLM memorization riddle: A Pythia-6.9B checkpoint generates the following Output, which occurs only 1 time in the Pile. Is this a verbatim memorization?

Attending #ACL2024? Come hear about our recent work from TAU/Google (with collaborators!) about interpretability, knowledge, and reasoning in LLMs! Sohee Yang @ ACL 2024 Jesujoba Alabi Alon Jacovi Gal Yona Zhengxuan Wu

🎉Adapters 1.0 is here!🚀 Our open-source library for modular and parameter-efficient fine-tuning got a major upgrade! v1.0 is packed with new features (ReFT, Adapter Merging, QLoRA, ...), new models & improvements! Blog: adapterhub.ml/blog/2024/08/a… Highlights in the thread! 🧵👇

Aligning Language Models with preferences leads to stronger and safer models (GPT3 → ChatGPT). However, preferences (RLHF) contain irrelevant signals, and alignment objectives (e.g. DPO) can actually hurt model performance. We tackle both, leading to a ~2x performance boost.

.Stanford NLP Group awards at #ACL2024 ▸ Best paper award Julie Kallini ✨ et al ▸ Outstanding paper award Aryaman Arora et al ▸ Outstanding paper award Weiyan Shi et al ▸ Best societal impact award Weiyan Shi et al ▸ 10 year test of time award Christopher Manning et al Congratulations! 🥂

It's back-to-school time and so here's the Fall '24 release of draft chapters for Speech and Language Processing! web.stanford.edu/~jurafsky/slp3/

The Linear Representation Hypothesis is now widely adopted despite its highly restrictive nature. Here, Csordás Róbert, Atticus Geiger, Christopher Manning & I present a counterexample to the LRH and argue for more expressive theories of interpretability: arxiv.org/abs/2408.10920

Intervention-based approaches to mechanistic interpretability have progressed at an astounding rate recently. In our new paper (a major update to a 2023 ms), we provide a formal framework and show how to express many methods within this framework: arxiv.org/abs/2301.04709

I am BEYOND EXCITED to publish our interview with Krista Opsahl-Ong (Krista Opsahl-Ong) from Stanford AI Lab! 🔥 Krista is the lead author of MIPRO, short for Multi-prompt Instruction Proposal Optimizer, and one of the leading developers and scientists behind DSPy! This was such

Should AI be aligned with human preferences, rewards, or utility functions? Excited to finally share a preprint that Micah Carroll Matija Franklin Hal Ashton & I have worked on for almost 2 years, arguing that AI alignment has to move beyond the preference-reward-utility nexus!

🔗 Thoughts on Research Impact in AI. Grad students often ask: how do I do research that makes a difference in the current, crowded AI space? This is a blogpost that summarizes my perspective in six guidelines for making research impact via open-source artifacts. Link below.

Automating AI research is exciting! But can LLMs actually produce novel, expert-level research ideas? After a year-long study, we obtained the first statistically significant conclusion: LLM-generated ideas are more novel than ideas written by expert human researchers.

Language models today are (1) widely used in personalized contexts and (2) to build systems that interface with tools. Do they respect privacy when helping with daily tasks like emailing? Introducing PrivacyLens to evaluate if LMs know privacy norms in action at inference time!

Do you want to select great LLM pretraining data but don’t have 1000 H100s for a ton of mixture experiments? What about a method that requires none of your own training, matches the best known existing method, and has some nice theory? New preprint: Perplexity Correlations

🚨 New paper alert! 🚨 SAEs 👾 are a hot topic in mechanistic interpretability 🛠️, but how well do they really work? We evaluated open-source SAEs of OpenAI , Apollo Research and by Joseph Bloom on GPT-2 small and found they struggle to disentangle knowledge compared to neurons.