<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kang-etal-2026-evidence">
    <titleInfo>
        <title>From Evidence to Exact Answers: Copy-Preserving Compression and Conversion-Aware Calibration for Frozen SLM RAG</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Dong</namePart>
        <namePart type="given">Ho</namePart>
        <namePart type="family">Kang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">HyeonJeong</namePart>
        <namePart type="family">Cha</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Daein</namePart>
        <namePart type="family">Weon</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>Frozen small language model (SLM) readers in retrieval-augmented generation (RAG) can fail even after retrieval has found answer-bearing evidence: on five LongBench tasks with LFM2.5-Instruct 1.2B, Raw RAG with its default prompt converts only 43% of answer-bearing predictions into exact matches. We treat frozen SLM RAG as a two-stage problem—copy-preserving evidence compression and conversion-aware deployment calibration—and instantiate each stage with one system. PEC-HOP is a non-parametric pointer-preserved evidence compressor; the Minimal Copy-preserving Evidence Compass (MCE-Compass) is a dev-time calibrator that selects an (interface, prompt) pair per (reader, task) cell, calibrated offline from ~50 labeled examples per cell for fixed deployments. In a matched-prompt head-to-head on three multi-hop LongBench tasks × three frozen readers, PEC-HOP is Pareto-best among compressors, with +4.28 pp macro EM over LLMLingua-2 and +1.61 pp over Provence at 2.8–4.8× faster median latency and no trained compressor on the critical path. Across five tasks, MCE-Compass improves macro EM over Raw RAG by +9.36, +15.39, and +2.39 pp on LFM2.5-Instruct 1.2B, Qwen3-4B, and Gemma-4-E4B at 2.02×, 5.69×, and 1.34× latency speedups, and beats the single globally dev-selected policy. A 20-partition dev-size sweep and a 66,581-article HotpotQA retrieval stress test indicate the gains are not single-split or bundled-context artifacts.</abstract>
    <identifier type="citekey">kang-etal-2026-evidence</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.849/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>19678</start>
            <end>19694</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T From Evidence to Exact Answers: Copy-Preserving Compression and Conversion-Aware Calibration for Frozen SLM RAG
%A Kang, Dong Ho
%A Cha, HyeonJeong
%A Weon, Daein
%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 kang-etal-2026-evidence
%X Frozen small language model (SLM) readers in retrieval-augmented generation (RAG) can fail even after retrieval has found answer-bearing evidence: on five LongBench tasks with LFM2.5-Instruct 1.2B, Raw RAG with its default prompt converts only 43% of answer-bearing predictions into exact matches. We treat frozen SLM RAG as a two-stage problem—copy-preserving evidence compression and conversion-aware deployment calibration—and instantiate each stage with one system. PEC-HOP is a non-parametric pointer-preserved evidence compressor; the Minimal Copy-preserving Evidence Compass (MCE-Compass) is a dev-time calibrator that selects an (interface, prompt) pair per (reader, task) cell, calibrated offline from ~50 labeled examples per cell for fixed deployments. In a matched-prompt head-to-head on three multi-hop LongBench tasks × three frozen readers, PEC-HOP is Pareto-best among compressors, with +4.28 pp macro EM over LLMLingua-2 and +1.61 pp over Provence at 2.8–4.8× faster median latency and no trained compressor on the critical path. Across five tasks, MCE-Compass improves macro EM over Raw RAG by +9.36, +15.39, and +2.39 pp on LFM2.5-Instruct 1.2B, Qwen3-4B, and Gemma-4-E4B at 2.02×, 5.69×, and 1.34× latency speedups, and beats the single globally dev-selected policy. A 20-partition dev-size sweep and a 66,581-article HotpotQA retrieval stress test indicate the gains are not single-split or bundled-context artifacts.
%U https://aclanthology.org/2026.emnlp-main.849/
%P 19678-19694

Markdown (Informal)

[From Evidence to Exact Answers: Copy-Preserving Compression and Conversion-Aware Calibration for Frozen SLM RAG](https://aclanthology.org/2026.emnlp-main.849/) (Kang et al., EMNLP 2026)

• From Evidence to Exact Answers: Copy-Preserving Compression and Conversion-Aware Calibration for Frozen SLM RAG (Kang et al., EMNLP 2026)

ACL

• Dong Ho Kang, HyeonJeong Cha, and Daein Weon. 2026. From Evidence to Exact Answers: Copy-Preserving Compression and Conversion-Aware Calibration for Frozen SLM RAG. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 19678–19694, Budapest, Hungary. Association for Computational Linguistics.