Reproduce “Temporal Dependencies in In-Context Learning: The Role of Induction Heads” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anooshka Bajaj, Deven Mahesh Mistry, Sahaj Singh Maini, Yash Aggarwal, Billy Dickson, Zoran Tiganj
Researchers at Indiana University Bloomington investigated the mechanistic underpinnings of temporal processing in large language models by linking specific attention mechanisms, known as induction heads, to serial recall behavior. They demonstrated that ablating these heads significantly reduced models' ability to predict the next token based on temporal contiguity and impaired performance in a serial recall task, revealing their crucial role in ordered retrieval.
Key insights
- Induction heads are critical components for LLMs' ability to predict the successor token (lag +1 probability) in a sequence, exhibiting a sharp, +1-centered serial recall tendency.
- The induction heads responsible for temporal dependencies are distributed throughout the model's architecture rather than being concentrated in specific layers.
- Instruction tuning can influence models' temporal retrieval strategies, sometimes shifting from direct token copying (lag 0) to successor prediction (lag +1).
- Ablating random, low-induction-score heads can sometimes *increase* the +1 lag probability, suggesting the existence of competing or regulatory circuits that interact with induction heads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.