<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhan-zhang-2026-spire">
    <titleInfo>
        <title>SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Chenlu</namePart>
        <namePart type="family">Zhan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yufei</namePart>
        <namePart type="family">Zhang</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>Production RL pipelines for large language models routinely violate the on-policy assumption along three independent axes. Partial rollouts resume long chain-of-thought trajectories after a parameter update. Rollout replay re-injects past successful trajectories. Asynchronous training decouples sampling from optimization, so trajectories arrive several updates stale. Today these biases are absorbed by a single, theoretically unjustified instrument: a token-level PPO clip whose magnitude is invariant to the actual policy divergence. We show that all three regimes are instances of one mathematical object: a trajectory composed of segments produced by distinct behavior policies. Building on this view, we introduce **SPIRE** (**Segment-wise Prefix Importance with Replay Equilibrium**), a drop-in replacement for the policy-gradient estimator in GRPO-style optimizers. **SPIRE** is: 1. **Unbiased** under a prefix-measurable boundary condition; 2. **Variance-controlled**, admitting a tight per-segment second-moment bound via Renyi-2 divergence with a closed-form MSE-optimal clip; 3. **Replay-equilibrium aware**, admitting a closed-form Pareto-optimal replay/fresh mixing ratio from a KKT analysis of the constrained MSE/throughput problem. On Qwen2.5-Math-1.5B, Qwen2.5-7B, and Llama-3.1-8B trained with verifiable rewards on MATH-500, AIME-2024, OlympiadBench, LiveCodeBench-v5, and GPQA-D, **SPIRE** delivers **+3.6 to +6.8 absolute points on AIME-2024** and **+2.1 to +4.4 on OlympiadBench** at iso-FLOPs, while remaining stable up to **4x trainer/sampler staleness**, where AReaL+PPO collapses. Baselines marked with a double-dagger symbol are reproduced verbatim from the corresponding paper on the matching model-benchmark pair; **SPIRE** numbers reflect the model’s own runs released with this submission.</abstract>
    <identifier type="citekey">zhan-zhang-2026-spire</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1613/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>36688</start>
            <end>36701</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling
%A Zhan, Chenlu
%A Zhang, Yufei
%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 zhan-zhang-2026-spire
%X Production RL pipelines for large language models routinely violate the on-policy assumption along three independent axes. Partial rollouts resume long chain-of-thought trajectories after a parameter update. Rollout replay re-injects past successful trajectories. Asynchronous training decouples sampling from optimization, so trajectories arrive several updates stale. Today these biases are absorbed by a single, theoretically unjustified instrument: a token-level PPO clip whose magnitude is invariant to the actual policy divergence. We show that all three regimes are instances of one mathematical object: a trajectory composed of segments produced by distinct behavior policies. Building on this view, we introduce **SPIRE** (**Segment-wise Prefix Importance with Replay Equilibrium**), a drop-in replacement for the policy-gradient estimator in GRPO-style optimizers. **SPIRE** is: 1. **Unbiased** under a prefix-measurable boundary condition; 2. **Variance-controlled**, admitting a tight per-segment second-moment bound via Renyi-2 divergence with a closed-form MSE-optimal clip; 3. **Replay-equilibrium aware**, admitting a closed-form Pareto-optimal replay/fresh mixing ratio from a KKT analysis of the constrained MSE/throughput problem. On Qwen2.5-Math-1.5B, Qwen2.5-7B, and Llama-3.1-8B trained with verifiable rewards on MATH-500, AIME-2024, OlympiadBench, LiveCodeBench-v5, and GPQA-D, **SPIRE** delivers **+3.6 to +6.8 absolute points on AIME-2024** and **+2.1 to +4.4 on OlympiadBench** at iso-FLOPs, while remaining stable up to **4x trainer/sampler staleness**, where AReaL+PPO collapses. Baselines marked with a double-dagger symbol are reproduced verbatim from the corresponding paper on the matching model-benchmark pair; **SPIRE** numbers reflect the model’s own runs released with this submission.
%U https://aclanthology.org/2026.emnlp-main.1613/
%P 36688-36701

Markdown (Informal)

[SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling](https://aclanthology.org/2026.emnlp-main.1613/) (Zhan & Zhang, EMNLP 2026)

• SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling (Zhan & Zhang, EMNLP 2026)

ACL

• Chenlu Zhan and Yufei Zhang. 2026. SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 36688–36701, Budapest, Hungary. Association for Computational Linguistics.