<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ozaki-etal-2026-first">
    <titleInfo>
        <title>First-Stage Groundwork Shapes Downstream RL Outcomes: Two-Stage Post-Training for LLM Personality Acquisition</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Taisei</namePart>
        <namePart type="family">Ozaki</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Tetsuya</namePart>
        <namePart type="family">Hashimoto</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Miki</namePart>
        <namePart type="family">Ota</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Keisuke</namePart>
        <namePart type="family">Hori</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhenyu</namePart>
        <namePart type="family">He</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Mitsuhiro</namePart>
        <namePart type="family">Mabuchi</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>Stable personality is a key requirement for personal AI assistants, yet how reasoning models acquire such traits remains underexplored. Existing work largely treats personality as role-play and inherits the standard reasoning-model recipe: SFT warm-up followed by RLVR. We ask whether this recipe is appropriate for natural personality acquisition, where a trait should shape both final answers and intermediate reasoning, persist under prompt variation, and preserve general capability. Holding Stage 2 fixed as rubric-based GRPO, we vary Stage 1 on Qwen3-8B across four Big Five trait directions and compare standard SFT with MCQ-GRPO, a simple on-policy warm-up over the model’s trait-relevant choices and reasoning traces. We evaluate each endpoint using an LLM-judged Rubric for stylistic expression and TRAIT for introspective disposition. We find that SFT-then-GRPO achieves high stylistic scores but is weaker on the introspective probe. In contrast, MCQ-GRPO-then-GRPO yields stronger introspective acquisition at a smaller mathematical-reasoning cost. Weight- and activation-space analyses further support this finding, suggesting that personality acquisition is not only a question of how much trait is displayed, but where it is carried.</abstract>
    <identifier type="citekey">ozaki-etal-2026-first</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1983/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>45104</start>
            <end>45131</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T First-Stage Groundwork Shapes Downstream RL Outcomes: Two-Stage Post-Training for LLM Personality Acquisition
%A Ozaki, Taisei
%A Hashimoto, Tetsuya
%A Ota, Miki
%A Hori, Keisuke
%A He, Zhenyu
%A Mabuchi, Mitsuhiro
%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 ozaki-etal-2026-first
%X Stable personality is a key requirement for personal AI assistants, yet how reasoning models acquire such traits remains underexplored. Existing work largely treats personality as role-play and inherits the standard reasoning-model recipe: SFT warm-up followed by RLVR. We ask whether this recipe is appropriate for natural personality acquisition, where a trait should shape both final answers and intermediate reasoning, persist under prompt variation, and preserve general capability. Holding Stage 2 fixed as rubric-based GRPO, we vary Stage 1 on Qwen3-8B across four Big Five trait directions and compare standard SFT with MCQ-GRPO, a simple on-policy warm-up over the model’s trait-relevant choices and reasoning traces. We evaluate each endpoint using an LLM-judged Rubric for stylistic expression and TRAIT for introspective disposition. We find that SFT-then-GRPO achieves high stylistic scores but is weaker on the introspective probe. In contrast, MCQ-GRPO-then-GRPO yields stronger introspective acquisition at a smaller mathematical-reasoning cost. Weight- and activation-space analyses further support this finding, suggesting that personality acquisition is not only a question of how much trait is displayed, but where it is carried.
%U https://aclanthology.org/2026.emnlp-main.1983/
%P 45104-45131

Markdown (Informal)

[First-Stage Groundwork Shapes Downstream RL Outcomes: Two-Stage Post-Training for LLM Personality Acquisition](https://aclanthology.org/2026.emnlp-main.1983/) (Ozaki et al., EMNLP 2026)

• First-Stage Groundwork Shapes Downstream RL Outcomes: Two-Stage Post-Training for LLM Personality Acquisition (Ozaki et al., EMNLP 2026)

ACL

• Taisei Ozaki, Tetsuya Hashimoto, Miki Ota, Keisuke Hori, Zhenyu He, and Mitsuhiro Mabuchi. 2026. First-Stage Groundwork Shapes Downstream RL Outcomes: Two-Stage Post-Training for LLM Personality Acquisition. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 45104–45131, Budapest, Hungary. Association for Computational Linguistics.