🧩 Philosophy 6d ago · r_knzv

A sudoku-solving transformer represents the board by substructure, not by cell

Less Wrong
View Channel →
A sudoku-solving transformer represents the board by substructure, not by cell
Source ↗ 👁 0 💬 0
tl;dr: a transformer trained on sudoku solving traces with backtracking maintains the board state per substructure linearly in the residual stream The main goal of this post is to understand if a transformer trained on solving traces creates a "world model" and uses it during the solving process. To do so, I trained a transformer that autoregressively predicts the next placement in a solution trace, mainly following Giannoulis et al. and Shah et al., and looked if the state is represented in the

Comments (0)

Sign in to join the discussion

More Like This

📰
Society is a social construct, pace Arrow
LessWrong · 3d ago
Consent-Based RL: Letting Models Endorse Their Own Training Updates
LessWrong · 3d ago
AI #164: Pre Opus
LessWrong · 3d ago
Publish-first writing
LessWrong · 3d ago
📰
What does status signalling do? When successful, what does it achieve?
LessWrong · 4d ago
📰
Let goodness conquer all that it can defend
LessWrong · 4d ago