8 papers
Context or Knowledge: How LLMs Attribute Their Own Answers
Kirill Afendulev, Alexey Dontsov, Alla Chepurova +2
Language models can answer a question from a supplied document or from their parametric knowledge. When later asked where the answer came from, what determines their source report?…
From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability
Dmitry Scherbakov, Oleg Chernikov, Victoria Khamash +5
We study feature vectors: directions in a language model's activation space that correspond to human-interpretable concepts, such as honesty or deception. Existing unsupervised met…
On-sky performance and prospects of near-infrared CMOS detectors in exoplanet detection: first year of SPIRIT
Clàudia Janó-Muñoz, Matthew J. Hooton, Peter P. Pedersen +8
Ultracool dwarfs are promising targets for detecting and characterising temperate terrestrial exoplanets. Their flux peaks in the near-infrared, beyond the wavelength cut-off of th…
Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
Pavel Tikhonov, Anton Korznikov, Matvey Mikhalchuk +6
While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linearity: when inputs from distinct text streams…
What Attention Recalls and Recurrence Controls in Hybrid Language Models
Kirill Afendulev, Alexey Dontsov, Elena Tutubalina +1
Hybrid language models combine attention with a fixed-size recurrent state, but the role of each channel remains unclear. We introduce two cache-level interventions. Split-prefill…
Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
Anton Korznikov, Andrey Galichin, Alexey Dontsov +3
Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. Rece…
OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features
Anton Korznikov, Andrey Galichin, Alexey Dontsov +3
Sparse autoencoders (SAEs) are a technique for sparse decomposition of neural network activations into human-interpretable features. However, current SAEs suffer from feature absor…
The Rogue Scalpel: Activation Steering Compromises LLM Safety
Anton Korznikov, Andrey Galichin, Alexey Dontsov +3
Activation steering is a promising technique for controlling LLM behavior by adding semantically meaningful vectors directly into a model's hidden states during inference. It is of…