On the Failure of Latent State Persistence in Large Language Models
Jen-tse Huang, Kaiser Sun, Wenxuan Wang, Mark Dredze
TL;DR
This work formalizes Latent State Persistence ($LSP$) as the ability to instantiate and maintain an internal latent variable without explicit context and shows that current LLMs fail to preserve such latent states, acting instead as reactive post-hoc solvers. Through three experiments—Number Guessing Game, Yes-No Game, and Mathematical Mentalism—the authors quantify the $LSP$ gap using metrics like Empirical State Mass and invariant-based success, across 17 frontier models. Results reveal systematic under-allocation of latent-state mass, recency-driven logical drift, and failures to maintain state under transformations, with externalized reasoning (Chain-of-Thought, scratchpads, or long-reasoning models) offering only partial mitigation. The findings highlight a fundamental architectural divergence between autoregressive transformers and human-like cognition, suggesting that scalable model size alone cannot close the gap without new mechanisms for persistent latent representation and state evolution. This has implications for AI alignment, reliability, and the design of autonomous agents that require stable internal commitments and goal tracking.
Abstract
While Large Language Models (LLMs) excel in reasoning, whether they can sustain persistent latent states remains under-explored. The capacity to maintain and manipulate unexpressed, internal representations-analogous to human working memory-is a cornerstone of complex reasoning. In this paper, we formalize and quantify the "Latent State Persistence" (LSP) gap through three novel experiments. First, we utilize a Number Guessing Game, demonstrating that across independent queries, LLMs fail to allocate probability mass to a singular hidden choice, violating a fundamental probabilistic principle. Second, we employ a Yes-No Game to show that as the number of questions increases, LLMs suffer from "concept drift," leading to inevitable self-contradictions due to the lack of LSP. Finally, inspired by Mathematical Mentalism, we task models with tracking transformations on hidden variables, revealing a failure in variable binding and state evolution when the initial state is not explicitly present in the context. Collectively, these findings suggest that LLMs function as reactive post-hoc solvers rather than proactive planners with LSP. Our work provides a framework for evaluating the fidelity of internal representations and highlights a fundamental architectural divergence between autoregressive transformers and human-like cognition.