@inproceedings{zhang-etal-2026-understanding-training,
title = "Understanding Training Data Attribution in {RLVR}: From the Perspective of Causal Inference",
author = "Zhang, Chi and
Zhao, Jian and
Wang, Wei and
Wang, Kai and
Zhao, Yao and
Li, Xuelong",
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.2023/",
pages = "43042--43056",
ISBN = "979-8-89176-462-0",
abstract = "Training data attribution is increasingly used to filter, reweight, and prioritize rollouts in reinforcement learning with verifiable rewards (RLVR). However, these interventions often assume that individual samples, transitions, or trajectories can be attributed independently of the other data used in the same update. We show that this assumption is not merely a property of the attribution estimator, but a structural property of the training objective. We recast training data attribution as individual treatment effect estimation and use the Stable Unit Treatment Value Assumption (SUTVA) to determine when a proposed attribution unit is well-defined. This yields a simple criterion that attribution is only valid at the level where the algorithm assigns a stable contribution to that unit. Applying this criterion to RLVR, we show that the PPO objective admits trajectory-level attribution, whereas naive trajectory-only attribution is not well-defined for GRPO unless peer context is modeled. Experimental results support this distinction and provide preliminary evidence for group-level attribution under GRPO."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-understanding-training">
<titleInfo>
<title>Understanding Training Data Attribution in RLVR: From the Perspective of Causal Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chi</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yao</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuelong</namePart>
<namePart type="family">Li</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>Training data attribution is increasingly used to filter, reweight, and prioritize rollouts in reinforcement learning with verifiable rewards (RLVR). However, these interventions often assume that individual samples, transitions, or trajectories can be attributed independently of the other data used in the same update. We show that this assumption is not merely a property of the attribution estimator, but a structural property of the training objective. We recast training data attribution as individual treatment effect estimation and use the Stable Unit Treatment Value Assumption (SUTVA) to determine when a proposed attribution unit is well-defined. This yields a simple criterion that attribution is only valid at the level where the algorithm assigns a stable contribution to that unit. Applying this criterion to RLVR, we show that the PPO objective admits trajectory-level attribution, whereas naive trajectory-only attribution is not well-defined for GRPO unless peer context is modeled. Experimental results support this distinction and provide preliminary evidence for group-level attribution under GRPO.</abstract>
<identifier type="citekey">zhang-etal-2026-understanding-training</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2023/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43042</start>
<end>43056</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Understanding Training Data Attribution in RLVR: From the Perspective of Causal Inference
%A Zhang, Chi
%A Zhao, Jian
%A Wang, Wei
%A Wang, Kai
%A Zhao, Yao
%A Li, Xuelong
%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 zhang-etal-2026-understanding-training
%X Training data attribution is increasingly used to filter, reweight, and prioritize rollouts in reinforcement learning with verifiable rewards (RLVR). However, these interventions often assume that individual samples, transitions, or trajectories can be attributed independently of the other data used in the same update. We show that this assumption is not merely a property of the attribution estimator, but a structural property of the training objective. We recast training data attribution as individual treatment effect estimation and use the Stable Unit Treatment Value Assumption (SUTVA) to determine when a proposed attribution unit is well-defined. This yields a simple criterion that attribution is only valid at the level where the algorithm assigns a stable contribution to that unit. Applying this criterion to RLVR, we show that the PPO objective admits trajectory-level attribution, whereas naive trajectory-only attribution is not well-defined for GRPO unless peer context is modeled. Experimental results support this distinction and provide preliminary evidence for group-level attribution under GRPO.
%U https://aclanthology.org/2026.findings-emnlp.2023/
%P 43042-43056
Markdown (Informal)
[Understanding Training Data Attribution in RLVR: From the Perspective of Causal Inference](https://aclanthology.org/2026.findings-emnlp.2023/) (Zhang et al., Findings 2026)
ACL