Reproduce “Repetitions are not all alike: distinct mechanisms sustain repetition in language models” 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
Matéo Mahaut, Francesca Franzon
Researchers at Universitat Pompeu Fabra revealed that repetitive text generation in Large Language Models stems from two distinct mechanisms: a learned, confident strategy induced by in-context learning and an early-emerging, less confident fallback behavior. They demonstrated these mechanisms have different developmental trajectories, rely on distinct neural circuitry, and exhibit varying attentional foci and prediction confidence.
Key insights
- Repetition in LLMs is not a monolithic phenomenon; behaviorally similar outputs arise from at least two qualitatively distinct internal processes.
- ICL-induced repetition is a learned, progressively specialized strategy supported by specific attention heads and MLPs, leading to high-confidence generation.
- Natural repetition is an early-emerging fallback behavior characterized by a lack of specialized circuitry, a focus on low-information structural tokens, and greater underlying uncertainty.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.