@inproceedings{liu-etal-2026-aggregate,
title = "Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon {LLM} Search Agents",
author = "Liu, Feixue and
Chen, Ruoyu and
Cui, Songyi and
Yan, Yimo and
Kuo, Yong-Hong",
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.503/",
pages = "11555--11570",
ISBN = "979-8-89176-459-0",
abstract = "Parallel test-time scaling treats each rollout as an independent candidate answer. For long-horizon search agents, however, these rollouts also produce rich traces of tool calls, retrieved evidence, and partial reasoning states. Intermediate failures such as reasoning drift are difficult to self-correct within a single rollout, yet useful corrective signals often already exist in other live rollouts before termination. This creates an opportunity to turn cross-trajectory information into online correction during execution, rather than harvesting it only after rollouts terminate. However, simple broadcast is unsafe because partial states can be redundant, irrelevant, or risky for the receiving trajectory. We propose OPTA, a live aggregation framework based on checkpoint-level synchronization: each rollout maintains a private report, while a controller distills a typed global shared state, and injects only relevant and useful feedback into each trajectory, preserving its exploration identity. On BrowseComp-Plus and DeepSearchQA across four LLM backbones, OPTA is consistently the strongest test-time scaling method, lifting Qwen3-30B on BrowseComp-Plus from 37.9{\%} to 59.6{\%}. Our results show that OPTA is a practical route toward timely and selective test-time scaling for long-horizon LLM search agents."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-aggregate">
<titleInfo>
<title>Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon LLM Search Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Feixue</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruoyu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Songyi</namePart>
<namePart type="family">Cui</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yimo</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yong-Hong</namePart>
<namePart type="family">Kuo</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>Parallel test-time scaling treats each rollout as an independent candidate answer. For long-horizon search agents, however, these rollouts also produce rich traces of tool calls, retrieved evidence, and partial reasoning states. Intermediate failures such as reasoning drift are difficult to self-correct within a single rollout, yet useful corrective signals often already exist in other live rollouts before termination. This creates an opportunity to turn cross-trajectory information into online correction during execution, rather than harvesting it only after rollouts terminate. However, simple broadcast is unsafe because partial states can be redundant, irrelevant, or risky for the receiving trajectory. We propose OPTA, a live aggregation framework based on checkpoint-level synchronization: each rollout maintains a private report, while a controller distills a typed global shared state, and injects only relevant and useful feedback into each trajectory, preserving its exploration identity. On BrowseComp-Plus and DeepSearchQA across four LLM backbones, OPTA is consistently the strongest test-time scaling method, lifting Qwen3-30B on BrowseComp-Plus from 37.9% to 59.6%. Our results show that OPTA is a practical route toward timely and selective test-time scaling for long-horizon LLM search agents.</abstract>
<identifier type="citekey">liu-etal-2026-aggregate</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.503/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>11555</start>
<end>11570</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon LLM Search Agents
%A Liu, Feixue
%A Chen, Ruoyu
%A Cui, Songyi
%A Yan, Yimo
%A Kuo, Yong-Hong
%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 liu-etal-2026-aggregate
%X Parallel test-time scaling treats each rollout as an independent candidate answer. For long-horizon search agents, however, these rollouts also produce rich traces of tool calls, retrieved evidence, and partial reasoning states. Intermediate failures such as reasoning drift are difficult to self-correct within a single rollout, yet useful corrective signals often already exist in other live rollouts before termination. This creates an opportunity to turn cross-trajectory information into online correction during execution, rather than harvesting it only after rollouts terminate. However, simple broadcast is unsafe because partial states can be redundant, irrelevant, or risky for the receiving trajectory. We propose OPTA, a live aggregation framework based on checkpoint-level synchronization: each rollout maintains a private report, while a controller distills a typed global shared state, and injects only relevant and useful feedback into each trajectory, preserving its exploration identity. On BrowseComp-Plus and DeepSearchQA across four LLM backbones, OPTA is consistently the strongest test-time scaling method, lifting Qwen3-30B on BrowseComp-Plus from 37.9% to 59.6%. Our results show that OPTA is a practical route toward timely and selective test-time scaling for long-horizon LLM search agents.
%U https://aclanthology.org/2026.emnlp-main.503/
%P 11555-11570
Markdown (Informal)
[Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon LLM Search Agents](https://aclanthology.org/2026.emnlp-main.503/) (Liu et al., EMNLP 2026)
ACL