<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="qiu-etal-2026-bag">
    <titleInfo>
        <title>Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Chen-Meng</namePart>
        <namePart type="family">Qiu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ze-Hao</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Deng-Bao</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Linan</namePart>
        <namePart type="family">Yue</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Min-Ling</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>Token probabilities play a central role in large language models. However, these probabilities often fail to reflect either answer correctness or probability distributions explicitly specified in the context. We attribute this failure to the **Context-Logit Probability Gap**: contextual probabilities are not reliably reflected in the model’s next-token logits. We show that post-hoc temperature scaling has a limited effect on addressing this gap, because it rescales logits without correcting erroneous probability rankings. Motivated by this, we propose Bag-of-Tokens (BoT) tuning, a very simple fine-tuning approach that directly calibrates the logits with the probability distribution conveyed by synthetic contexts. More specifically, we construct BoT examples by forming a bag of candidate tokens with an explicitly assigned probability distribution. This distribution is expressed in the context and used as soft supervision to align the model’s logit probabilities via KL-divergence minimization. Experiments show that BoT tuning generalizes to LLM-based sampling tasks, such as random-number generation, and improves downstream QA accuracy and calibration. These results suggest that calibrating LLM logits with contextual probabilities provides a lightweight and effective way to enhance their uncertainty awareness.</abstract>
    <identifier type="citekey">qiu-etal-2026-bag</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1592/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>36198</start>
            <end>36215</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities
%A Qiu, Chen-Meng
%A Wang, Ze-Hao
%A Wang, Deng-Bao
%A Yue, Linan
%A Zhang, Min-Ling
%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 qiu-etal-2026-bag
%X Token probabilities play a central role in large language models. However, these probabilities often fail to reflect either answer correctness or probability distributions explicitly specified in the context. We attribute this failure to the **Context-Logit Probability Gap**: contextual probabilities are not reliably reflected in the model’s next-token logits. We show that post-hoc temperature scaling has a limited effect on addressing this gap, because it rescales logits without correcting erroneous probability rankings. Motivated by this, we propose Bag-of-Tokens (BoT) tuning, a very simple fine-tuning approach that directly calibrates the logits with the probability distribution conveyed by synthetic contexts. More specifically, we construct BoT examples by forming a bag of candidate tokens with an explicitly assigned probability distribution. This distribution is expressed in the context and used as soft supervision to align the model’s logit probabilities via KL-divergence minimization. Experiments show that BoT tuning generalizes to LLM-based sampling tasks, such as random-number generation, and improves downstream QA accuracy and calibration. These results suggest that calibrating LLM logits with contextual probabilities provides a lightweight and effective way to enhance their uncertainty awareness.
%U https://aclanthology.org/2026.emnlp-main.1592/
%P 36198-36215

Markdown (Informal)

[Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities](https://aclanthology.org/2026.emnlp-main.1592/) (Qiu et al., EMNLP 2026)

• Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities (Qiu et al., EMNLP 2026)

ACL

• Chen-Meng Qiu, Ze-Hao Wang, Deng-Bao Wang, Linan Yue, and Min-Ling Zhang. 2026. Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 36198–36215, Budapest, Hungary. Association for Computational Linguistics.