Searcharxiv⌕ Search

arXiv · 2611.13760

Recursive memory for prediction and generation

Abstract

Recurrent and state-space language models compress the past into a state of fixed size. The state must answer questions about the future now, and it must remain useful as new tokens arrive, including tokens the model samples itself. We show that the second requirement can cost far more memory than the first. For a known hidden Markov source with positive transition probabilities, a state of $O(\log(1/\varepsilon))$ bits answers every question about the future to accuracy $\varepsilon$, however far ahead it looks, because the Bayesian filter forgets its rounding errors. Generating the future is different. When a model feeds each sampled token back into its state, the small errors in its next-token distributions add up, and $T$ tokens can require a state of order $\log(T/\varepsilon)$ bits. The lower bound comes from the KL chain rule: at every step, a model with $K$ states pays for approximating the true next-token distribution by one of only $K$ distributions. The argument applies to any source, including a language model continuing a prompt, and a small-ball estimate makes it tight for two-state sources. The extra $\log T$ disappears if the model keeps private randomness, or if error is budgeted per token, as in perplexity. Statistics of the recent past show the same gap between describing and maintaining. The number of ones among the last $n$ bits takes $\log n$ bits to write down. But if each bit is a noisy copy of the one $n$ steps earlier, tracking that count to a mean-square error $E^2$ well below its variance costs order $n/(1+E^2)+\log n$ bits. Information about each bit must be kept until the bit leaves the window, and a queue holding the counts of recent blocks keeps just enough of it.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Pengbin Feng, Chunlei Meng. 2026-11-10. Recursive memory for prediction and generation. https://searcharxiv.com/abs/2611.13760

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

A Bayesian Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance

In pragmatic cluster randomized controlled trials (PCRCTs), healthcare providers are randomized, but intervention delivery and patient receipt may vary after assignment. In many PCRCTs, cluster-level implementation is measured using continuous implementation metrics, while individual compliance is recorded as a binary indicator. Standard complier average causal effect (CACE) estimands focus on individual-level compliance and do not account for heterogeneity in implementation across clusters. When intervention uptake is shaped by both cluster- and individual-level processes, it is of scientific interest to characterize how effects vary across these sources of compliance. We propose a Bayesian framework for PCRCTs with one-sided binary noncompliance at the individual level and one-sided partial compliance at the cluster level. The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on baseline characteristics and treatment-arm implementation measures, and links these latent implementation strata to individual compliance and outcomes through a joint model. Because individual compliance and treatment-defined implementation measures are observed only under treatment assignment, the model imputes these quantities for control clusters and individuals assigned to control. The framework enables estimation of finite- and super-population intent-to-treat (ITT) and CACE estimands, both marginally and within latent implementation strata. We apply the method to the METRIcAL trial, a pragmatic cluster randomized study evaluating a personalized music intervention for nursing home residents with dementia. The analysis illustrates how accounting for implementation heterogeneity and individual compliance can provide insights beyond standard ITT analyses.

stat.AP↗

Copula-Based Synthetic Data Generation: A Use-Inspired Perspective on Trade-offs

The generation of synthetic data offers a promising way to address challenges related to data scarcity and privacy concerns. While many approaches have been developed, tabular synthetic data generation methods often require complex models and extensive training data, while lacking interpretability. In this work, we investigate versatile probabilistic approaches using copulas, and leverage in particular the Gaussian Mixture Copula Model, to model the joint distribution of multivariate data, with a main focus on moderate-dimensional data sets. Taking a use-inspired perspective, we examine key trade-offs inherent to synthetic data evaluation, spanning fidelity, utility, and privacy, and demonstrate that copula-based methods achieve a sound balance across all three dimensions while remaining interpretable and providing an explicit probabilistic model. We illustrate and highlight the potential benefits of the method on a test case focusing on systematic missing values in a cardiovascular risk calculator and develop specific evaluation metrics for this application.

stat.AP↗

JALE: A new Python toolbox for coordinate-based neuroimaging meta-analysis using activation likelihood estimation

Coordinate-based neuroimaging meta-analysis using activation likelihood estimation (ALE) is commonly used to synthesize reported results of functional or structural magnetic resonance imaging studies. We here present JALE, a new, open-source Python package for conducting ALE meta-analyses. Besides computing standard ALE main effects and contrast analyses, this toolbox implements several novel analytical approaches including new options for thresholding (threshold-free cluster enhancement statistics, significance threshold prediction) as well as sensitivity and cluster analyses that were only available as in-house scripts in the past. The purpose of this paper is to provide an integrated conceptual foundation of recent developments in the ALE meta-analysis framework, with a particular focus on novel methodological implementations that can be executed in a straightforward and standardized manner using the new JALE toolbox.

stat.AP↗