Research report from Anthropic exhibiting that language models can demonstrate introspective awareness of their internal states. Under certain conditions, Claude models show the ability to notice and identify injected concept vectors and internal states.
Findings
This is where I'll be saving links to interesting things I stumble across on the internet. You'll find cool research papers on LLMs that have caught my attention, fascinating tech discoveries, and other curiosities worth preserving.
Emergent Introspective Awareness in LLMs
Oct 29, 2025
Interpretability
Anthropic - Transformer Circuits
Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders
This paper releases 256 open-source SAEs trained on every layer and sublayer of Llama-3.1-8B, covering residual streams, attention outputs, MLP outputs, and transcoders at both 32K and 128K feature widths.
- Sparse Autoencoders (SAEs) are unsupervised machine learning methods designed to extract interpretable features from neural networks by addressing superposition of features.
- The paper employs TopK SAEs, an improved variant that directly selects the K highest-activating features rather than using L1 penalties. This, along with other improvements, result in high reconstruction quality while achieving ~3x better sparsity (Lā ā 50 vs 150) compared to state-of-the-art JumpReLU SAEs.
Oct 27, 2024
Interpretability
arXiv