Do LLMs Understand Sequential Structure? A Controlled Study of Inference and Generation
A controlled study finds LLMs often miss latent sequential rules despite matching surface action frequencies.
The authors test whether LLMs recover latent sequential structure rather than only matching surface action frequencies, using two-player Rock-Paper-Scissors and a one-player stochastic n-gram continuation task. The framework separates distribution matching from conditional rule following, including strategy identification, simple Markov rules, and higher-order dependencies. Longer context did not improve identification, correct recognition did not produce faithful simulation, and higher-order dependencies sharply reduced rule recovery. They conclude that apparent behavioral fidelity can conceal incorrect generative mechanisms.
- Tests strategy identification and simple Markov rule following
- Separates distribution matching from conditional generation
- Longer context did not improve latent-strategy identification
- Higher-order dependencies substantially degrade rule recovery
From the original article
1 min read at huggingface.co
Large language models (LLMs) are increasingly used as interactive agents and simulators, yet it remains unclear whether they can recover latent sequential structure beyond surface action frequencies. This distinction is critical for behavioral simulation, where actions are often shaped by prior context rather than marginal frequencies alone.
By Jerry Wang, Zhengxiang Wang, Ting Yu Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.