💡 We compare prompting (zero and multi-shot + explanations) and inference-time interventions (ActAdd, REFT and SAEs). Following SpARE (@yuzhaouoe.bsky.social @alessiodevoto.bsky.social), we propose ✨ contrastive SAE steering ✨ with mutual info to personalize literary MT by tuning latent features 4/

MMLU-Redux just touched down at #NAACL2025! 🎉 Wish I could be there for our "Are We Done with MMLU?" poster today (9:00-10:30am in Hall 3, Poster Session 7), but visa drama said nope 😅 If anyone's swinging by, give our research some love! Hit me up if you check it out! 👋

My amazing collaborators will present several works at ICLR and NAACL later this month -- please catch up with them if you're attending! I tried to summarise our recent work in a blog post: neuralnoise.com/2025/march-r...

Please share it within your circles! edin.ac/3DDQK1o

🚀 New Paper Alert! 🚀 We introduce Q-Filters, a training-free method for efficient KV Cache compression! It is compatible with FlashAttention and can compress along generation which is particularly useful for reasoning models ⚡ TLDR: we make Streaming-LLM smarter using the geometry of attention

Live from the CoLoRAI workshop at AAAI (april-tools.github.io/colorai/) Nadav Cohen is now giving his talk on "What Makes Data Suitable for Deep Learning?" Tools from quantum physics are shown to be useful in building more expressive deep learning models by changing the data distribution.

2018: Saliency maps give plausible interpretations of random weights, triggering skepticism and catalyzing the mechinterp cultural movement, which now advocates for SAEs. 2025: SAEs give plausible interpretations of random weights, triggering skepticism and ...

Graphical tensor notation for interpretability www.lesswrong.com/posts/BQKKQi...

Introducing The AI CUDA Engineer: An agentic AI system that automates the production of highly optimized CUDA kernels. sakana.ai/ai-cuda-engi... The AI CUDA Engineer can produce highly optimized CUDA kernels, reaching 10-100x speedup over common machine learning operations in PyTorch. Examples:

It's 2025, and I’ve finally updated my Python setup guide to use uv + venv instead of conda + pip! Here's my go-to recommendation for uv + venv in Python projects for faster installs, better dependency management: github.com/rasbt/LLMs-f... (Any additional suggestions?)

Cool research on how models memorize data 📝 : The 'Manifold Memorization Hypothesis' by Brendan Ross, Hamidreza Kamkariet al. suggests memorization occurs when the model's learned manifold matches the true data manifold but with too small 'local intrinsic dimensionality'.

Massive activations & weights in LLMs, two cool works 🤓: - The Super Weight: finds performance can be totally degraded when pruning a *single* weight - Mengxia Yu et al. - Massive Activations in LLM:finds some (crucial) activations have very high norm irrespective of context - Mingjie Sun et al.

On the last day before the Spring Festival holiday in China, DeepSeek released a NEW work on @hf.co 🤯 Janus-Pro🔥 autoregressive framework that unifies multimodal understanding and generation huggingface.co/deepseek-ai/... ✨ 1B / 7B ✨ MIT License

Not only that, but much of the science community here is already stronger and larger than it was on X. On Twitter, my feed of scientists who study climate-related topics topped out at 3300. Here, we’re at 4500 already and it’s still growing. Pin here: bsky.app/profile/did:...

*MoE Graph Transformers for Interpretable Particle Collision Detection* by @alessiodevoto.bsky.social @sgiagu.bsky.social et al. We propose a MoE graph transformer for particle collision analysis, with many nice interpretability insights (e.g., expert specialization). arxiv.org/abs/2501.03432

deepseek GGUF just dropped, if you have 207GB disk/40GB RAM for the smallest version huggingface.co/collections/...

LLMs inner representations🔬 Llamas Work in English: LLMs default to English-based concept representations, regardless of input language @wendlerc.bsky.social et al Semantic Hub: Multimodal models create a single shared semantic space, structured by their primary language @zhaofengwu.bsky.social et a

I'll get straight to the point. We trained 2 new models. Like BERT, but modern. ModernBERT. Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff. It's much faster, more accurate, longer context, and more useful. 🧵

In Vision & Audio transformers, not all tokens need the same compute resources! We propose “modular learners” to control compute at token-level granularity (MHA & MLP): hard tokens get more, easy ones get less! w/ @sscardapane.bsky.social @neuralnoise.com @bartoszWojcik Soon #AAAI25 Link 👇

Josh Tenenbaum on scaling up vs growing up and the path to human-like reasoning #NeurIPS2024

*Adaptive Computation Modules: Granular Conditional Computation For Efficient Inference* with @alessiodevoto.bsky.social @neuralnoise.com Happy to share our work on distilling efficient transformers with dynamic modules' activation was accepted at #AAAI2025. 🔥 arxiv.org/abs/2312.10193

*Sparse Crosscoders for Cross-Layer Features and Model Diffing* by @colah.bsky.social @anthropic.com Investigates stability & dynamics of "interpretable features" with cross-layers SAEs. Can also be used to investigate differences in fine-tuned models. transformer-circuits.pub/2024/crossco...

Very cool work! 👏🚀 Unfortunately, errors in the original dataset will propagate to all new languages 😕 We investigated the issue of existing errors in the original MMLU in arxiv.org/abs/2406.04127 @aryopg.bsky.social @neuralnoise.com

Super Cool work from Cohere for AI! 🎉 However, this highlights a concern raised by our MMLU-Redux team (arxiv.org/abs/2406.04127): **error propagation to many languages**. Issues in MMLU (e.g., "rapid intervention to solve ebola") seem to persist in many languages. Let's solve the root cause first?

Cool take on straight-through-estimator (to backpropagate through discrete ops): during back-propagation, keep gradient's relative angle, not absolute direction. The authors call this the "rotation trick". From "Restructuring Vector Quantization With The Rotation Trick" @ChristopherFifty et al.

Wish I knew about this much earlier: In VS Code (or Cursor 😉) you can install a 'data wrangler' extension to inspect the values of your PyTorch Tensors in a nice UI with summary statistics. Just right-click on a tensor in your variables view!

Why don’t they just make the whole app out of L1 cache samwho.dev/numbers/

It's Sunday morning so taking a minute for a nerdy thread (on math, tokenizers and LLMs) of the work of our intern Garreth By adding a few lines of code to the base Llama 3 tokenizer, he got a free boost in arithmetic performance 😮 [thread]

Hi, just sharing an updated version of the PyTorch 2 Internals slides: drive.google.com/file/d/18YZV.... Content: basics, jit, dynamo, Inductor, export path and executorch. This is focused on internals so you will need a bit of C/C++. I show how you can export and run a model on a Pixel Watch too.

*Building neural networks in Equinox* Third lab for the course, after JAX and Keras we see Equinox (from @patrickkidger.bsky.social) - including callable pytrees, stateful modules, and filtering. Stay tuned for part b with a text classification model. 😎 colab.research.google.com/drive/19NHkb...

Apple released AIMv2 🍏 a family of state-of-the-art open-set vision encoders > like CLIP, but add a decoder and train on autoregression 🤯 > 19 open models come in 300M, 600M, 1.2B, 2.7B with resolutions of 224, 336, 448 > Loadable and usable with 🤗 transformers huggingface.co/collections/...

New and very cool library!👏 Our L2 Norm-based KV Cache compression is already implemented - ready to use! 🚀 Check out the method details in our EMNLP '24 paper: arxiv.org/abs/2406.11430

KVPress supports our state-of-the-art KV-Cache compression method from "A Simple and Effective L2 Norm-Based Strategy for KV Cache Compression" (arxiv.org/abs/2406.11430, EMNLP'24 Oral) 🚀🚀🚀 You can read more on our EMNLP papers (and other projs) in this blog post: www.neuralnoise.com/2024/nov-res...

*Tokenformer: Rethinking Transformer Scaling With Tokenized Model Parameters* arxiv.org/pdf/2410.23168 Very cool transformer-inspired architecture where linear layers are replaced with token-parameter attention (Pattention). This allows for efficient scaling by adding new parameters to the model.

I’ll be travelling to London from Wednesday to Friday for an upcoming event and would be very happy to meet up! 🚀 I'd love to chat about my recent works (DeCoRe, MMLU-Redux, etc.). DM me if you’re around! 👋 DeCoRe: arxiv.org/abs/2410.18860 MMLU-Redux: arxiv.org/abs/2406.04127

FWIW, blueark.app worked well. $5 to copy my tweet history over. Not sure *my* tweets were worth that... But that's not their fault.

Just landed here! 🔥 A nice news to start: *Alice in a differentiable wonderland* has gone over 1000 copies sold on Amazon and I am super happy about the feedback! If you happen to buy a copy feel free to drop a review and/or send me suggestions on the material: www.sscardapane.it/alice-book/