Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback
Abstract
At the start of every session, LLM agents load a fixed context file, such as AGENTS.md. Each loaded token in the file is charged again in every later round of the session, and these files can degrade performance as they grow in size. However, in practice, human or automated curators usually grow these files by appending. We formulate context curation as a capacitated assortment problem. Instructions consume tokens under a finite attention capacity; adding an instruction never raises the compliance of the others, while retained instructions incur a per-session setup cost. We prove an upper bound on the optimal file size, regardless of the number of available candidate instructions, and that appending every instruction with positive standalone value can be arbitrarily worse in net value than selecting an optimal subset. A token budget also limits the loss when the token price is underestimated. We then examine what can be learned from past sessions and how this information can guide decisions to add or remove instructions. Feedback is inherently censored: the benefits and harms of loaded instructions are observable, whereas missing instructions generate feedback only when their absence causes harm. In this setting, we show that deleting instructions ignored by agents can inevitably remove helpful ones. We characterize how much evidence should be collected before adding an instruction. Besides, we bound regret when human reviewers can inspect only a limited number of edits per period. Empirical experiments further show that irrelevant rules drawn from real context files reduce language-model compliance.
Related
- RELAGRAM: Recursive Latent Graph Model for the Capacitated Vehicle Routing Problem
- Discrete-Time MDP Modeling for Multi-Item Capacitated Lot Sizing with Stochastic Demand Timing
- A Unified Framework for Assortment Optimization and Matching under MNL Choice
- AI-Mediation in Retail Markets: Revenue Effects and Assortment Optimization
- Assortment optimization under the perception-adjusted Luce model
BibTeX
@misc{liu2026curating,
title = {{Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback}},
author = {Zexuan Liu and Yuning Yang and Tiancheng Zhao},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.11007},
eprint = {2610.11007},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.11007},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback
AU - Zexuan Liu
AU - Yuning Yang
AU - Tiancheng Zhao
PY - 2026
DA - 2026/10/07
PB - arXiv
DO - 10.48550/arxiv.2610.11007
UR - https://arxiv.org/abs/2610.11007
AB - At the start of every session, LLM agents load a fixed context file, such as AGENTS.md. Each loaded token in the file is charged again in every later round of the session, and these files can degrade performance as they grow in size. However, in practice, human or automated curators usually grow these files by appending. We formulate context curation as a capacitated assortment problem. Instructions consume tokens under a finite attention capacity; adding an instruction never raises the compliance of the others, while retained instructions incur a per-session setup cost. We prove an upper bound on the optimal file size, regardless of the number of available candidate instructions, and that appending every instruction with positive standalone value can be arbitrarily worse in net value than selecting an optimal subset. A token budget also limits the loss when the token price is underestimated. We then examine what can be learned from past sessions and how this information can guide decisions to add or remove instructions. Feedback is inherently censored: the benefits and harms of loaded instructions are observable, whereas missing instructions generate feedback only when their absence causes harm. In this setting, we show that deleting instructions ignored by agents can inevitably remove helpful ones. We characterize how much evidence should be collected before adding an instruction. Besides, we bound regret when human reviewers can inspect only a limited number of edits per period. Empirical experiments further show that irrelevant rules drawn from real context files reduce language-model compliance.
ER -