@inproceedings{teng-etal-2026-skip,
title = "Skip-Connected Rollouts for Mid-Trajectory Rewards in {GRPO}",
author = "Teng, Fengwei and
Bai, Jinyi and
Yao, Xinhao and
Wang, Demi and
Zhao, Jiahao and
Guo, Zhijiang",
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.934/",
pages = "19710--19728",
ISBN = "979-8-89176-462-0",
abstract = "Group Relative Policy Optimization (GRPO) and its variants have proven effective for reinforcement learning with verifiable rewards through outcome-based feedback. Incorporating intermediate credit, however, remains challenging. We show that, under practical sampling budgets, token-level Monte Carlo estimates at early reasoning positions have high variance, which can cause frequent advantage-sign errors after group-relative normalization and yield wrong-direction token-level gradient contributions. At later positions, these estimates increasingly resemble the terminal outcome reward and provide limited additional information.\par We propose SCRoll, a rollout-side module that augments group-relative policy optimizers with randomized single-boundary segment-level rewards while leaving the downstream optimizer interface unchanged. SCRoll uses a skip-connected two-stream rollout. An upstream stream generates an early-stopped reasoning segment $s$, while a downstream stream samples $G$ continuations conditioned on $[s,q]$, which places the segment before the original problem. This construction retains the segment as preceding reasoning context while mitigating the reward homogenization associated with pure-prefix continuation. The same downstream group supplies terminal outcomes for the standard group-relative objective and a Monte Carlo aggregate used as the upstream segment reward, avoiding a separate objective-specific rollout group. The upstream segment is optimized with a sample-efficient single-stream objective.\par Across Qwen2.5-Math-7B and Llama-3.2-3B-Instruct, pairing SCRoll with four representative group-relative optimizers improves performance in most evaluated settings. The clearest gains occur with CISPO, with smaller positive average gains for GRPO and DAPO and marginal gains for GSPO. A decoupled two-process baseline can achieve comparable performance with substantially more generated tokens, while SCRoll performs better under a matched generated-token budget, supporting the sample-efficiency benefit of rollout reuse."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="teng-etal-2026-skip">
<titleInfo>
<title>Skip-Connected Rollouts for Mid-Trajectory Rewards in GRPO</title>
</titleInfo>
<name type="personal">
<namePart type="given">Fengwei</namePart>
<namePart type="family">Teng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinyi</namePart>
<namePart type="family">Bai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinhao</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Demi</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiahao</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhijiang</namePart>
<namePart type="family">Guo</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>Group Relative Policy Optimization (GRPO) and its variants have proven effective for reinforcement learning with verifiable rewards through outcome-based feedback. Incorporating intermediate credit, however, remains challenging. We show that, under practical sampling budgets, token-level Monte Carlo estimates at early reasoning positions have high variance, which can cause frequent advantage-sign errors after group-relative normalization and yield wrong-direction token-level gradient contributions. At later positions, these estimates increasingly resemble the terminal outcome reward and provide limited additional information.\par We propose SCRoll, a rollout-side module that augments group-relative policy optimizers with randomized single-boundary segment-level rewards while leaving the downstream optimizer interface unchanged. SCRoll uses a skip-connected two-stream rollout. An upstream stream generates an early-stopped reasoning segment s, while a downstream stream samples G continuations conditioned on [s,q], which places the segment before the original problem. This construction retains the segment as preceding reasoning context while mitigating the reward homogenization associated with pure-prefix continuation. The same downstream group supplies terminal outcomes for the standard group-relative objective and a Monte Carlo aggregate used as the upstream segment reward, avoiding a separate objective-specific rollout group. The upstream segment is optimized with a sample-efficient single-stream objective.\par Across Qwen2.5-Math-7B and Llama-3.2-3B-Instruct, pairing SCRoll with four representative group-relative optimizers improves performance in most evaluated settings. The clearest gains occur with CISPO, with smaller positive average gains for GRPO and DAPO and marginal gains for GSPO. A decoupled two-process baseline can achieve comparable performance with substantially more generated tokens, while SCRoll performs better under a matched generated-token budget, supporting the sample-efficiency benefit of rollout reuse.</abstract>
<identifier type="citekey">teng-etal-2026-skip</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.934/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19710</start>
<end>19728</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Skip-Connected Rollouts for Mid-Trajectory Rewards in GRPO
%A Teng, Fengwei
%A Bai, Jinyi
%A Yao, Xinhao
%A Wang, Demi
%A Zhao, Jiahao
%A Guo, Zhijiang
%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 teng-etal-2026-skip
%X Group Relative Policy Optimization (GRPO) and its variants have proven effective for reinforcement learning with verifiable rewards through outcome-based feedback. Incorporating intermediate credit, however, remains challenging. We show that, under practical sampling budgets, token-level Monte Carlo estimates at early reasoning positions have high variance, which can cause frequent advantage-sign errors after group-relative normalization and yield wrong-direction token-level gradient contributions. At later positions, these estimates increasingly resemble the terminal outcome reward and provide limited additional information.\par We propose SCRoll, a rollout-side module that augments group-relative policy optimizers with randomized single-boundary segment-level rewards while leaving the downstream optimizer interface unchanged. SCRoll uses a skip-connected two-stream rollout. An upstream stream generates an early-stopped reasoning segment s, while a downstream stream samples G continuations conditioned on [s,q], which places the segment before the original problem. This construction retains the segment as preceding reasoning context while mitigating the reward homogenization associated with pure-prefix continuation. The same downstream group supplies terminal outcomes for the standard group-relative objective and a Monte Carlo aggregate used as the upstream segment reward, avoiding a separate objective-specific rollout group. The upstream segment is optimized with a sample-efficient single-stream objective.\par Across Qwen2.5-Math-7B and Llama-3.2-3B-Instruct, pairing SCRoll with four representative group-relative optimizers improves performance in most evaluated settings. The clearest gains occur with CISPO, with smaller positive average gains for GRPO and DAPO and marginal gains for GSPO. A decoupled two-process baseline can achieve comparable performance with substantially more generated tokens, while SCRoll performs better under a matched generated-token budget, supporting the sample-efficiency benefit of rollout reuse.
%U https://aclanthology.org/2026.findings-emnlp.934/
%P 19710-19728
Markdown (Informal)
[Skip-Connected Rollouts for Mid-Trajectory Rewards in GRPO](https://aclanthology.org/2026.findings-emnlp.934/) (Teng et al., Findings 2026)
ACL
- Fengwei Teng, Jinyi Bai, Xinhao Yao, Demi Wang, Jiahao Zhao, and Zhijiang Guo. 2026. Skip-Connected Rollouts for Mid-Trajectory Rewards in GRPO. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 19710–19728, Budapest, Hungary. Association for Computational Linguistics.