RT @ArjunGuha: We present a new benchmark for reasoning models that reveals capability gaps and failure modes that are not evident in exist….

RT @jadenfk23: 🚀 New NNsight features launching today! If you’re conducting research on LLM internals, NNsight 0.3 is now available. This u….

Join millions who have switched to Grok.

RT @jadenfk23: Frontier LLMs have capabilities that smaller AIs don't, but up to now there's been no way to crack them open. Now that #Lla….

RT @ellev3n11: Llama-3.1 trains on synthetic translations of Python to low-resource languages (e.g., PHP) to improve performance on MultiPL….

RT @davidbau: The National Deep Inference Fabric #NDIF, an @NSF-funded AI research infrastructure project, is awarding 2024 **Summer Engine….

RT @_akhaliq: NNsight and NDIF. Democratizing Access to Foundation Model Internals. The enormous scale of state-of-the-art foundation model….

Read the full details and results of our study at

Activation steering is a powerful, lightweight technique for aligning CodeLLMs. Unlike fine-tuning, steering vectors do not modify model weights. You can improve robustness on type prediction without degrading performance on other tasks!.

This is surprising‼️It suggests that CodeLLMs may be learning a representation of types that is shared across Python and TypeScript---and potentially other languages.

Our method for building steering vectors generalizes to held-out programs. Moreover, we find that vectors built on Python data are effective for TypeScript, and vice versa ✅.

In our work, we leverage activation steering to make type prediction more robust. We compute steering vectors that edit model activations at inference time. We find that this method corrects up to 90% of type mispredictions in Python and 75% in TypeScript.

This makes deploying CodeLLMs for type prediction challenging 😕. Real code has arbitrary names, and we would like models to handle this.

CodeLLMs are sensitive to syntactic changes in code, like editing the names of variables and types, or deleting type hints. We focus on the type prediction task. We find that by applying semantics-preserving edits to our prompts, model prediction eventually breaks…

CodeLLMs are revolutionizing programming, but current models have limitations. We’re excited to announce a novel technique for making models more robust using activation steering. If you’re interested in mechanistic interpretability for code, keep reading! 🧵.

RT @Francis_YAO_: How did the initial #GPT3 evolve to today's #ChatGPT ? Where do the amazing abilities of #GPT3.5 come from? What is enabl….

RT @DeepMind: Introducing a generalist neural algorithmic learner, capable of carrying out 30 different reasoning tasks, with a 𝘴𝘪𝘯𝘨𝘭𝘦 grap….

RT @sewon__min: Most if not all language models use a softmax that gives a categorical probability distribution over a finite vocab. We int….

RT @linguistMasoud: Ok I think it is time to share my "foundations of linguistics" syllabus with you here. It took me a long time to work o….

RT @MetaAI: 4️⃣ Papers we presented at #NeurIPS2022 that you should know about (and how you can learn more even if you’re not at the confer….

RT @schwabpa: You couldn't make it to #NeurIPS2022 this year?. Nothing to worry - I curated a summary for you below focussing on key papers….