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

Really cool paper by @kayoyin.bsky.social about interpretability of In Context Learning, they found that Function Vectors (FV) heads are crucial for few-shot ICL.
www.arxiv.org/abs/2502.14010

A really nice resource for understanding how to parallelize LLM training.

3/ Interleaving Concepts with Token Embeddings

🔹 Predicted concepts are compressed into a continuous vector 🎯
🔹 They are then inserted into hidden states alongside token embeddings 🧩

2/ Training the Model with Dual Objectives

🔹 Next-token prediction – the standard LLM training objective.
🔹 Concept prediction – the model learns to reproduce extracted concepts from its hidden state.

1/ Concept Extraction with SAE

🔹 A Sparse Autoencoder (SAE) extracts high-level concepts from the hidden states of a pretrained LLM.
🔹 Only the most important concepts are selected based on their attribution score (impact on model output).

🚀 Meta’s new LLM pretraining framework predicts concepts and integrates them into its hidden state to enhance next-token prediction. 🚀

It achieves the same performance with 21.5% fewer tokens and better generalization! 🎯

📝: arxiv.org/abs/2502.08524

A very interesting work that explores the possibility of having a unified interpretation across multiple models

*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