@inproceedings{perez-beltrachini-lapata-2026-action,
title = "Action-Level Credit Assignment for Retrieval-Augmented Reasoning",
author = "Perez-Beltrachini, Laura and
Lapata, Mirella",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1203/",
doi = "10.18653/2026.findings-emnlp.1203",
pages = "25414--25434",
ISBN = "979-8-89176-462-0",
abstract = "Complex question answering over large document collections requires models to identify missing information, issue sub-queries, and iteratively gather evidence before producing an answer. Reinforcement learning offers a natural framework for training such behaviour, but standard GRPO-style training provides only a sparse final-answer reward and assigns the same advantage to every token in a rollout. We introduce PACE-GRPO (Potential-based Action Credit Estimation GRPO), a variant of GRPO that combines potential-based reward shaping with action-level credit assignment. PACE-GRPO computes a state potential from inexpensive retrieval-derived signals, including retrieval utility, coverage, and passage diversity, and uses potential differences to provide step-level feedback. Unlike prior intermediate-reward methods, it requires no gold evidence annotations, teacher traces, reward model, or additional forward passes. Across six multi-hop QA benchmarks, PACE-GRPO learns a tool-use policy that selectively invokes the more expensive retrieval operation only when it is beneficial. This adaptive tool use yields consistent improvements over standard GRPO and strong search-augmented baselines. Our code and data are in https://github.com/lauhaide/pace."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="perez-beltrachini-lapata-2026-action">
<titleInfo>
<title>Action-Level Credit Assignment for Retrieval-Augmented Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Laura</namePart>
<namePart type="family">Perez-Beltrachini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mirella</namePart>
<namePart type="family">Lapata</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Complex question answering over large document collections requires models to identify missing information, issue sub-queries, and iteratively gather evidence before producing an answer. Reinforcement learning offers a natural framework for training such behaviour, but standard GRPO-style training provides only a sparse final-answer reward and assigns the same advantage to every token in a rollout. We introduce PACE-GRPO (Potential-based Action Credit Estimation GRPO), a variant of GRPO that combines potential-based reward shaping with action-level credit assignment. PACE-GRPO computes a state potential from inexpensive retrieval-derived signals, including retrieval utility, coverage, and passage diversity, and uses potential differences to provide step-level feedback. Unlike prior intermediate-reward methods, it requires no gold evidence annotations, teacher traces, reward model, or additional forward passes. Across six multi-hop QA benchmarks, PACE-GRPO learns a tool-use policy that selectively invokes the more expensive retrieval operation only when it is beneficial. This adaptive tool use yields consistent improvements over standard GRPO and strong search-augmented baselines. Our code and data are in https://github.com/lauhaide/pace.</abstract>
<identifier type="citekey">perez-beltrachini-lapata-2026-action</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1203</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1203/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>25414</start>
<end>25434</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Action-Level Credit Assignment for Retrieval-Augmented Reasoning
%A Perez-Beltrachini, Laura
%A Lapata, Mirella
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F perez-beltrachini-lapata-2026-action
%X Complex question answering over large document collections requires models to identify missing information, issue sub-queries, and iteratively gather evidence before producing an answer. Reinforcement learning offers a natural framework for training such behaviour, but standard GRPO-style training provides only a sparse final-answer reward and assigns the same advantage to every token in a rollout. We introduce PACE-GRPO (Potential-based Action Credit Estimation GRPO), a variant of GRPO that combines potential-based reward shaping with action-level credit assignment. PACE-GRPO computes a state potential from inexpensive retrieval-derived signals, including retrieval utility, coverage, and passage diversity, and uses potential differences to provide step-level feedback. Unlike prior intermediate-reward methods, it requires no gold evidence annotations, teacher traces, reward model, or additional forward passes. Across six multi-hop QA benchmarks, PACE-GRPO learns a tool-use policy that selectively invokes the more expensive retrieval operation only when it is beneficial. This adaptive tool use yields consistent improvements over standard GRPO and strong search-augmented baselines. Our code and data are in https://github.com/lauhaide/pace.
%R 10.18653/2026.findings-emnlp.1203
%U https://aclanthology.org/2026.findings-emnlp.1203/
%U https://doi.org/10.18653/2026.findings-emnlp.1203
%P 25414-25434
Markdown (Informal)
[Action-Level Credit Assignment for Retrieval-Augmented Reasoning](https://aclanthology.org/2026.findings-emnlp.1203/) (Perez-Beltrachini & Lapata, Findings 2026)
ACL