@inproceedings{liao-wang-2026-training,
title = "No Training, No Architecture Change: Per-Cell Warm-Up Calibration for {LLM} Agent Memory",
author = "Liao, Wenwen and
Wang, Yuansong",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1957/",
pages = "44556--44575",
ISBN = "979-8-89176-459-0",
abstract = "LLM agent memory modules universally adopt a single global retrieval depth k across all deployment scenarios; this global constraint produces an irreducible Oracle performance gap. We carry out a systematic evaluation on 19 heterogeneous deployment cells covering three LLM backbones, three multi-agent coordinators, three tasks, and two representative memory architectures (A-Mem and G-Memory). We confirm that the per-cell optimal k varies dramatically across cells, that low-capacity LLMs exhibit catastrophic accuracy collapse at larger k , and that no cross-cell feature or lookup-table shortcut reliably predicts the per-cell optimum. A theoretical analysis further shows that in the small-budget paired-binary-reward regime, every training-free bandit method we evaluate reduces to uniform-mean argmax. Motivated by this collapse, we propose per-cell warmup calibration: paired probe queries iteratively estimate the mean reward of each candidate k, the procedure commits to the empirical argmax, and a closed-form (2 {\ensuremath{\sigma}}) criterion provides earlystop. The recipe is training-free, requires no architectural modification, and introduces no hyperparameter beyond the probe budget. On all 19 deployment cells it attains 100{\%} of the per-cell Oracle, with an average absolute improvement of +3.33 percentage points and a peak gain of +12.00 percentage points."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liao-wang-2026-training">
<titleInfo>
<title>No Training, No Architecture Change: Per-Cell Warm-Up Calibration for LLM Agent Memory</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wenwen</namePart>
<namePart type="family">Liao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuansong</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>LLM agent memory modules universally adopt a single global retrieval depth k across all deployment scenarios; this global constraint produces an irreducible Oracle performance gap. We carry out a systematic evaluation on 19 heterogeneous deployment cells covering three LLM backbones, three multi-agent coordinators, three tasks, and two representative memory architectures (A-Mem and G-Memory). We confirm that the per-cell optimal k varies dramatically across cells, that low-capacity LLMs exhibit catastrophic accuracy collapse at larger k , and that no cross-cell feature or lookup-table shortcut reliably predicts the per-cell optimum. A theoretical analysis further shows that in the small-budget paired-binary-reward regime, every training-free bandit method we evaluate reduces to uniform-mean argmax. Motivated by this collapse, we propose per-cell warmup calibration: paired probe queries iteratively estimate the mean reward of each candidate k, the procedure commits to the empirical argmax, and a closed-form (2 \ensuremathσ) criterion provides earlystop. The recipe is training-free, requires no architectural modification, and introduces no hyperparameter beyond the probe budget. On all 19 deployment cells it attains 100% of the per-cell Oracle, with an average absolute improvement of +3.33 percentage points and a peak gain of +12.00 percentage points.</abstract>
<identifier type="citekey">liao-wang-2026-training</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1957/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>44556</start>
<end>44575</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T No Training, No Architecture Change: Per-Cell Warm-Up Calibration for LLM Agent Memory
%A Liao, Wenwen
%A Wang, Yuansong
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F liao-wang-2026-training
%X LLM agent memory modules universally adopt a single global retrieval depth k across all deployment scenarios; this global constraint produces an irreducible Oracle performance gap. We carry out a systematic evaluation on 19 heterogeneous deployment cells covering three LLM backbones, three multi-agent coordinators, three tasks, and two representative memory architectures (A-Mem and G-Memory). We confirm that the per-cell optimal k varies dramatically across cells, that low-capacity LLMs exhibit catastrophic accuracy collapse at larger k , and that no cross-cell feature or lookup-table shortcut reliably predicts the per-cell optimum. A theoretical analysis further shows that in the small-budget paired-binary-reward regime, every training-free bandit method we evaluate reduces to uniform-mean argmax. Motivated by this collapse, we propose per-cell warmup calibration: paired probe queries iteratively estimate the mean reward of each candidate k, the procedure commits to the empirical argmax, and a closed-form (2 \ensuremathσ) criterion provides earlystop. The recipe is training-free, requires no architectural modification, and introduces no hyperparameter beyond the probe budget. On all 19 deployment cells it attains 100% of the per-cell Oracle, with an average absolute improvement of +3.33 percentage points and a peak gain of +12.00 percentage points.
%U https://aclanthology.org/2026.emnlp-main.1957/
%P 44556-44575
Markdown (Informal)
[No Training, No Architecture Change: Per-Cell Warm-Up Calibration for LLM Agent Memory](https://aclanthology.org/2026.emnlp-main.1957/) (Liao & Wang, EMNLP 2026)
ACL