@inproceedings{yu-etal-2026-olivia,
title = "{OLIVIA}: Online Learning via Inference-time Action Adaptation",
author = "Yu, Sheldon and
Wu, Junda and
Li, Xintong and
Kuang, Nikki Lijing and
Zhou, Sizhe and
Yu, Tong and
Han, Jiawei and
Shang, Jingbo and
McAuley, Julian",
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.764/",
pages = "17719--17734",
ISBN = "979-8-89176-459-0",
abstract = "Large language models (LLMs) tool-use policies select actions from a finite candidate set at each reasoning step. We observe that this action-selection interface is already a linear contextual scorer: under a restricted output vocabulary, the LM head computes per-action scores from the last-layer hidden state, exactly the structure assumed by linear contextual bandits. Existing online adaptation methods for LLMs operate indirectly through prompts or retrieved memories, leaving this decision layer fixed and providing no explicit mechanism for action-level credit assignment or uncertainty-aware exploration. We propose OLIVIA, a lightweight online adaptation layer that replaces the frozen action-selection head with a per-action contextual linear bandit, selects actions via LinUCB exploration on the LLM{'}s hidden state and updates only its per-action sufficient statistics from step-level feedback via a rank-one Sherman{--}Morrison rule. Without backbone fine-tuning and additional LLM calls per step, OLIVIA adds negligible overhead relative to LLM inference. Across four tool-use benchmarks ToolBench, TaskBench, TaskBench-MM, BFCL and two backbones Qwen3-4B, Mistral-7B, OLIVIA outperforms prompt-based online adaptation and contextual-bandit variants."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yu-etal-2026-olivia">
<titleInfo>
<title>OLIVIA: Online Learning via Inference-time Action Adaptation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sheldon</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junda</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xintong</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikki</namePart>
<namePart type="given">Lijing</namePart>
<namePart type="family">Kuang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sizhe</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiawei</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingbo</namePart>
<namePart type="family">Shang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Julian</namePart>
<namePart type="family">McAuley</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>Large language models (LLMs) tool-use policies select actions from a finite candidate set at each reasoning step. We observe that this action-selection interface is already a linear contextual scorer: under a restricted output vocabulary, the LM head computes per-action scores from the last-layer hidden state, exactly the structure assumed by linear contextual bandits. Existing online adaptation methods for LLMs operate indirectly through prompts or retrieved memories, leaving this decision layer fixed and providing no explicit mechanism for action-level credit assignment or uncertainty-aware exploration. We propose OLIVIA, a lightweight online adaptation layer that replaces the frozen action-selection head with a per-action contextual linear bandit, selects actions via LinUCB exploration on the LLM’s hidden state and updates only its per-action sufficient statistics from step-level feedback via a rank-one Sherman–Morrison rule. Without backbone fine-tuning and additional LLM calls per step, OLIVIA adds negligible overhead relative to LLM inference. Across four tool-use benchmarks ToolBench, TaskBench, TaskBench-MM, BFCL and two backbones Qwen3-4B, Mistral-7B, OLIVIA outperforms prompt-based online adaptation and contextual-bandit variants.</abstract>
<identifier type="citekey">yu-etal-2026-olivia</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.764/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17719</start>
<end>17734</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T OLIVIA: Online Learning via Inference-time Action Adaptation
%A Yu, Sheldon
%A Wu, Junda
%A Li, Xintong
%A Kuang, Nikki Lijing
%A Zhou, Sizhe
%A Yu, Tong
%A Han, Jiawei
%A Shang, Jingbo
%A McAuley, Julian
%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 yu-etal-2026-olivia
%X Large language models (LLMs) tool-use policies select actions from a finite candidate set at each reasoning step. We observe that this action-selection interface is already a linear contextual scorer: under a restricted output vocabulary, the LM head computes per-action scores from the last-layer hidden state, exactly the structure assumed by linear contextual bandits. Existing online adaptation methods for LLMs operate indirectly through prompts or retrieved memories, leaving this decision layer fixed and providing no explicit mechanism for action-level credit assignment or uncertainty-aware exploration. We propose OLIVIA, a lightweight online adaptation layer that replaces the frozen action-selection head with a per-action contextual linear bandit, selects actions via LinUCB exploration on the LLM’s hidden state and updates only its per-action sufficient statistics from step-level feedback via a rank-one Sherman–Morrison rule. Without backbone fine-tuning and additional LLM calls per step, OLIVIA adds negligible overhead relative to LLM inference. Across four tool-use benchmarks ToolBench, TaskBench, TaskBench-MM, BFCL and two backbones Qwen3-4B, Mistral-7B, OLIVIA outperforms prompt-based online adaptation and contextual-bandit variants.
%U https://aclanthology.org/2026.emnlp-main.764/
%P 17719-17734
Markdown (Informal)
[OLIVIA: Online Learning via Inference-time Action Adaptation](https://aclanthology.org/2026.emnlp-main.764/) (Yu et al., EMNLP 2026)
ACL
- Sheldon Yu, Junda Wu, Xintong Li, Nikki Lijing Kuang, Sizhe Zhou, Tong Yu, Jiawei Han, Jingbo Shang, and Julian McAuley. 2026. OLIVIA: Online Learning via Inference-time Action Adaptation. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 17719–17734, Budapest, Hungary. Association for Computational Linguistics.