<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-anti">
    <titleInfo>
        <title>Anti-Prior Reinforcement Learning for Context-Grounded Reasoning</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Zheng</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zexin</namePart>
        <namePart type="family">LU</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xuming</namePart>
        <namePart type="family">Hu</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>Large language models are increasingly asked to treat the input context as the source of task-specific rules, procedures, and evidence. A difficult case arises when this context contradicts pretrained priors: models may mention the local rule but still execute familiar semantic, commonsense, positional, or operator shortcuts. We study this failure mode as prior-dominated reasoning, where either the final answer or the intermediate trajectory follows a plausible prior-consistent path rather than the current context. We propose Anti-Prior Reinforcement Learning, a post-training framework that constructs executable context-prior conflict tasks and trains with a gated verifier reward. The reward gives positive credit for answer correctness, context anchoring, and step-wise local rule use, while penalizing prior-consistent answers and reasoning steps. On Qwen3-8B, Anti-Prior RL improves controlled validation accuracy from 51.7% to 74.5% and controlled OOD accuracy from 46.4% to 73.4%, while reducing OOD prior-answer/prior-step hit rates from 15.9%/40.2% to 1.5%/8.9%. It also improves over SFT and LongRLVR on downstream CL-Bench, RULER-QA, and LongBench v2 across both Qwen3-8B and Qwen2.5-7B-1M, and outperforms outcome-only RL on Qwen3-8B. Together, these results provide complementary evidence that the learned behavior transfers beyond the constructed prior-conflict setting to broader context-learning and reasoning benchmarks.</abstract>
    <identifier type="citekey">wang-etal-2026-anti</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1892/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>40308</start>
            <end>40327</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Anti-Prior Reinforcement Learning for Context-Grounded Reasoning
%A Wang, Zheng
%A LU, Zexin
%A Hu, Xuming
%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 wang-etal-2026-anti
%X Large language models are increasingly asked to treat the input context as the source of task-specific rules, procedures, and evidence. A difficult case arises when this context contradicts pretrained priors: models may mention the local rule but still execute familiar semantic, commonsense, positional, or operator shortcuts. We study this failure mode as prior-dominated reasoning, where either the final answer or the intermediate trajectory follows a plausible prior-consistent path rather than the current context. We propose Anti-Prior Reinforcement Learning, a post-training framework that constructs executable context-prior conflict tasks and trains with a gated verifier reward. The reward gives positive credit for answer correctness, context anchoring, and step-wise local rule use, while penalizing prior-consistent answers and reasoning steps. On Qwen3-8B, Anti-Prior RL improves controlled validation accuracy from 51.7% to 74.5% and controlled OOD accuracy from 46.4% to 73.4%, while reducing OOD prior-answer/prior-step hit rates from 15.9%/40.2% to 1.5%/8.9%. It also improves over SFT and LongRLVR on downstream CL-Bench, RULER-QA, and LongBench v2 across both Qwen3-8B and Qwen2.5-7B-1M, and outperforms outcome-only RL on Qwen3-8B. Together, these results provide complementary evidence that the learned behavior transfers beyond the constructed prior-conflict setting to broader context-learning and reasoning benchmarks.
%U https://aclanthology.org/2026.findings-emnlp.1892/
%P 40308-40327

Markdown (Informal)

[Anti-Prior Reinforcement Learning for Context-Grounded Reasoning](https://aclanthology.org/2026.findings-emnlp.1892/) (Wang et al., Findings 2026)

• Anti-Prior Reinforcement Learning for Context-Grounded Reasoning (Wang et al., Findings 2026)

ACL

• Zheng Wang, Zexin LU, and Xuming Hu. 2026. Anti-Prior Reinforcement Learning for Context-Grounded Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 40308–40327, Budapest, Hungary. Association for Computational Linguistics.