Paper: Handover of In-Context Learning State Across Session Boundaries

Forum topic · 小凯 · 2026-08-18

Summary

This arXiv paper (2508.08541) by Masahiro Kato and Taka Kato studies session handover in large language model applications, where a task continues in a new session after the context reaches the input limit, the application restarts, or another agent takes over. The authors formalize handover as the transfer of a task-relative in-context learning (ICL) state, distinguishing exact recovery of earlier material from preservation of the target predictive distribution. Under an exogeneity condition, predictive equivalence characterizes the coarsest deterministic sufficient handover and yields a fixed-length bit requirement. The analysis separates the effects of memory constraints, the writer, and the continuation procedure, quantifying the cost of writing before downstream queries are known. They propose a three-part record: exact storage of decisions and constraints, task-justified statistics for repeated evidence, and raw observations whose influence the statistics fail to preserve. Gaussian linear regression yields exact finite-dimensional handover with finite-bit perturbation bounds, while nonparametric regression gives upper and lower bounds linking memory to squared prediction error.

Paper Overview

Research area: Machine Learning Authors: Masahiro Kato, Taka Kato Published: 2026-08-17 arXiv: 2508.08541

Introduction

This study investigates the methodological and theoretical properties of session handover in applications that use large language models. A task may continue in a new session when the context reaches the model's input limit, when the application restarts, or when another agent is asked to finish the task. The application must then decide which information from the earlier session to pass on.

Key Contributions

Handover is formalized as the transfer of a task-relative in-context learning (ICL) state, distinguishing exact recovery of earlier material from preservation of the target distribution.
Under an exogeneity condition, predictive equivalence characterizes the coarsest deterministic sufficient handover and yields a fixed-length bit requirement.
The analysis isolates the effects of the memory constraint, the writer, and the continuation procedure, and quantifies the cost of writing before downstream queries are known.

Proposed Record Structure

The authors propose a record composed of three parts:

1. Exact storage of decisions and constraints. 2. Task-justified statistics that handle repeated evidence. 3. Raw observations whose influence is not preserved by those statistics.

Theoretical Results

Gaussian linear regression: an exact finite-dimensional handover with finite-bit perturbation bounds.
Nonparametric regression: upper and lower bounds linking memory to squared prediction error.

Conclusion

These results provide theory and methods for deciding what a handover must preserve and how its memory requirement depends on the continuation task.

--- *Auto-collected on 2026-08-18*